diff --git a/CMakeLists.txt b/CMakeLists.txt
index c89d2d979e30142e96da464c9d9d46ec9aec9f2e..29a2c9066886554fc561f0a645f15459fbea99d9 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -130,7 +130,7 @@ install(TARGETS arbor-config-defs EXPORT arbor-targets)
# for the arbor library.
add_library(arbor-private-deps INTERFACE)
-target_link_libraries(arbor-private-deps INTERFACE arbor-config-defs)
+target_link_libraries(arbor-private-deps INTERFACE arbor-config-defs ext-random123)
install(TARGETS arbor-private-deps EXPORT arbor-targets)
# Interface library `arborenv-private-deps` collects dependencies, options etc.
diff --git a/arbor/include/arbor/morph/embed_pwlin.hpp b/arbor/include/arbor/morph/embed_pwlin.hpp
index 82d83a570bc6f95d163acd8d838869f1a47a9591..d6947ac87ed594ea52918aaf5e59296120b7c9fc 100644
--- a/arbor/include/arbor/morph/embed_pwlin.hpp
+++ b/arbor/include/arbor/morph/embed_pwlin.hpp
@@ -28,6 +28,10 @@ struct embed_pwlin {
// Interpolated radius at location.
double radius(mlocation) const;
+ mcable_list radius_cmp(msize_t bid, double rad_lim, comp_op op) const;
+
+ double directed_projection(mlocation) const;
+ mcable_list projection_cmp(msize_t bid, double proj_lim, comp_op op) const;
// Computed length of mcable.
double integrate_length(mcable c) const;
diff --git a/arbor/include/arbor/morph/locset.hpp b/arbor/include/arbor/morph/locset.hpp
index d4785d89a8e9bd53260ff89cadaed3f56083d505..3429717526f9a75a4a1c10de62bb2342bf8b1f43 100644
--- a/arbor/include/arbor/morph/locset.hpp
+++ b/arbor/include/arbor/morph/locset.hpp
@@ -118,8 +118,9 @@ private:
};
};
-namespace ls {
+class region;
+namespace ls {
// Explicit location on morphology.
locset location(msize_t branch, double pos);
@@ -138,6 +139,16 @@ locset named(std::string);
// The null (empty) set.
locset nil();
+// Most distal points of a region
+locset most_distal(region reg);
+
+// Most proximal point of a region
+locset most_proximal(region reg);
+
+// A range `left` to `right` of randomly selected locations with a
+// uniform distribution from region `reg` generated using `seed`
+locset uniform(region reg, unsigned left, unsigned right, uint64_t seed);
+
// Proportional location on every branch.
locset on_branches(double pos);
diff --git a/arbor/include/arbor/morph/primitives.hpp b/arbor/include/arbor/morph/primitives.hpp
index 4b9e223b993db2659e94c00d78796c9b5f53f76c..fa6cb3eab733136757acfc7a715a0b090a746b0a 100644
--- a/arbor/include/arbor/morph/primitives.hpp
+++ b/arbor/include/arbor/morph/primitives.hpp
@@ -28,6 +28,14 @@ mpoint lerp(const mpoint& a, const mpoint& b, double u);
bool is_collocated(const mpoint& a, const mpoint& b);
double distance(const mpoint& a, const mpoint& b);
+// Indicate allowed comparison operations for classifying regions
+enum class comp_op {
+ lt,
+ le,
+ gt,
+ ge
+};
+
// A morphology sample consists of a location and an integer tag.
// When loaded from an SWC file, the tag will correspond to the SWC label,
// which are standardised as follows:
diff --git a/arbor/include/arbor/morph/region.hpp b/arbor/include/arbor/morph/region.hpp
index bd049fd68225f314ca2d92e1cf1e0bce33a774d1..629c7345b05e040f14baccecd2a08e65ca8190e6 100644
--- a/arbor/include/arbor/morph/region.hpp
+++ b/arbor/include/arbor/morph/region.hpp
@@ -116,6 +116,8 @@ private:
};
};
+class locset;
+
namespace reg {
// An empty region.
@@ -130,6 +132,28 @@ region branch(msize_t);
// Region with all segments with segment tag id.
region tagged(int id);
+// Region with all segments distal from another region
+region distal_interval(locset start, double distance);
+
+// Region with all segments proximal from another region
+region proximal_interval(locset end, double distance);
+
+// Region with all segments with radius less than/less than or equal to r
+region radius_lt(region reg, double r);
+region radius_le(region reg, double r);
+
+// Region with all segments with radius greater than/greater than or equal to r
+region radius_gt(region reg, double r);
+region radius_ge(region reg, double r);
+
+// Region with all segments with projection less than/less than or equal to r
+region z_dist_from_soma_lt(double r);
+region z_dist_from_soma_le(double r);
+
+// Region with all segments with projection greater than/greater than or equal to r
+region z_dist_from_soma_gt(double r);
+region z_dist_from_soma_ge(double r);
+
// Region with all segments in a cell.
region all();
diff --git a/arbor/morph/embed_pwlin.cpp b/arbor/morph/embed_pwlin.cpp
index e04436b4598224ab3a975ac60f5c527e40fd11b2..d81e9e4c6145c40c4223ed7149b7200abc3c33fc 100644
--- a/arbor/morph/embed_pwlin.cpp
+++ b/arbor/morph/embed_pwlin.cpp
@@ -53,14 +53,48 @@ double integrate(const branch_pw_ratpoly<p, q>& f, unsigned bid, const pw_consta
return accum;
}
+template <typename operation>
+mcable_list data_cmp(const branch_pw_ratpoly<1, 0>& f, unsigned bid, double val, operation op) {
+ mcable_list L;
+ const auto& pw = f.at(bid);
+ for (const auto& piece: pw) {
+ auto extents = piece.first;
+ auto left_val = piece.second(0);
+ auto right_val = piece.second(1);
+
+ if (!op(left_val, val) && !op(right_val, val)) {
+ continue;
+ }
+ if (op(left_val, val) && op(right_val, val)) {
+ L.push_back({bid, extents.first, extents.second});
+ continue;
+ }
+
+ auto cable_loc = (val - left_val)/(right_val - left_val);
+ auto edge = math::lerp(extents.first, extents.second, cable_loc);
+
+ if (op(left_val, val)) {
+ L.push_back({bid, extents.first, edge});
+ continue;
+ }
+ if (!op(left_val, val)) {
+ L.push_back({bid, edge, extents.second});
+ continue;
+ }
+ }
+ return L;
+}
+
struct embed_pwlin_data {
branch_pw_ratpoly<1, 0> length;
+ branch_pw_ratpoly<1, 0> directed_projection;
branch_pw_ratpoly<1, 0> radius;
branch_pw_ratpoly<2, 0> area;
branch_pw_ratpoly<1, 1> ixa;
explicit embed_pwlin_data(msize_t n_branch):
length(n_branch),
+ directed_projection(n_branch),
radius(n_branch),
area(n_branch),
ixa(n_branch)
@@ -71,6 +105,10 @@ double embed_pwlin::radius(mlocation loc) const {
return interpolate(data_->radius, loc.branch, loc.pos);
}
+double embed_pwlin::directed_projection(arb::mlocation loc) const {
+ return interpolate(data_->directed_projection, loc.branch, loc.pos);
+}
+
double embed_pwlin::integrate_length(msize_t bid, const pw_constant_fn& g) const {
return integrate(data_->length, bid, g);
}
@@ -97,6 +135,26 @@ double embed_pwlin::integrate_ixa(mcable c) const {
return integrate_ixa(c.branch, pw_constant_fn{{c.prox_pos, c.dist_pos}, {1.}});
}
+mcable_list embed_pwlin::radius_cmp(msize_t bid, double val, comp_op op) const {
+ switch (op) {
+ case comp_op::lt: return data_cmp(data_->radius, bid, val, [](auto l, auto r){return l < r;});
+ case comp_op::le: return data_cmp(data_->radius, bid, val, [](auto l, auto r){return l <= r;});
+ case comp_op::gt: return data_cmp(data_->radius, bid, val, [](auto l, auto r){return l > r;});
+ case comp_op::ge: return data_cmp(data_->radius, bid, val, [](auto l, auto r){return l >= r;});
+ default: return {};
+ }
+}
+
+mcable_list embed_pwlin::projection_cmp(msize_t bid, double val, comp_op op) const {
+ switch (op) {
+ case comp_op::lt: return data_cmp(data_->directed_projection, bid, val, [](auto l, auto r){return l < r;});
+ case comp_op::le: return data_cmp(data_->directed_projection, bid, val, [](auto l, auto r){return l <= r;});
+ case comp_op::gt: return data_cmp(data_->directed_projection, bid, val, [](auto l, auto r){return l > r;});
+ case comp_op::ge: return data_cmp(data_->directed_projection, bid, val, [](auto l, auto r){return l >= r;});
+ default: return {};
+ }
+}
+
// Initialization, creation of geometric data.
embed_pwlin::embed_pwlin(const arb::morphology& m) {
@@ -109,16 +167,18 @@ embed_pwlin::embed_pwlin(const arb::morphology& m) {
const auto& samples = m.samples();
sample_locations_.resize(m.num_samples());
+ double proj_shift = samples.front().loc.z;
+
for (msize_t bid = 0; bid<n_branch; ++bid) {
unsigned parent = m.branch_parent(bid);
auto sample_indices = util::make_range(m.branch_indexes(bid));
-
if (bid==0 && m.spherical_root()) {
arb_assert(sample_indices.size()==1);
// Treat spherical root as area-equivalent cylinder.
double r = samples[0].loc.radius;
+ data_->directed_projection[bid].push_back(0., 1., rat_element<1, 0>(-r, r));
data_->length[bid].push_back(0., 1., rat_element<1, 0>(0, r*2));
data_->radius[bid].push_back(0., 1., rat_element<1, 0>(r, r));
@@ -155,13 +215,15 @@ embed_pwlin::embed_pwlin(const arb::morphology& m) {
double length_0 = parent==mnpos? 0: data_->length[parent].back().second[1];
data_->length[bid].push_back(0., 1, rat_element<1, 0>(length_0, length_0+branch_length));
- double area_0 = parent=mnpos? 0: data_->area[parent].back().second[1];
- double ixa_0 = parent=mnpos? 0: data_->ixa[parent].back().second[1];
+ double area_0 = parent==mnpos? 0: data_->area[parent].back().second[1];
+ double ixa_0 = parent==mnpos? 0: data_->ixa[parent].back().second[1];
if (length_scale==0) {
// Zero-length branch? Weird, but make best show of it.
double r = samples[sample_indices[0]].loc.radius;
+ double z = samples[sample_indices[0]].loc.z;
data_->radius[bid].push_back(0., 1., rat_element<1, 0>(r, r));
+ data_->directed_projection[bid].push_back(0., 1., rat_element<1, 0>(z-proj_shift, z-proj_shift));
data_->area[bid].push_back(0., 1., rat_element<2, 0>(area_0, area_0, area_0));
data_->ixa[bid].push_back(0., 1., rat_element<1, 1>(ixa_0, ixa_0, ixa_0));
}
@@ -173,6 +235,10 @@ embed_pwlin::embed_pwlin(const arb::morphology& m) {
double p1 = sample_locations_[sample_indices[i]].pos;
if (p0==p1) continue;
+ double z0 = samples[sample_indices[i-1]].loc.z - proj_shift;
+ double z1 = samples[sample_indices[i]].loc.z - proj_shift;
+ data_->directed_projection[bid].push_back(p0, p1, rat_element<1, 0>(z0, z1));
+
double r0 = samples[sample_indices[i-1]].loc.radius;
double r1 = samples[sample_indices[i]].loc.radius;
data_->radius[bid].push_back(p0, p1, rat_element<1, 0>(r0, r1));
diff --git a/arbor/morph/locset.cpp b/arbor/morph/locset.cpp
index 1f7eafaa123195bda775cca63d6a4c638029b1ff..19b50d84e20745ee7ff2acedba3d7ddf2c7bb54a 100644
--- a/arbor/morph/locset.cpp
+++ b/arbor/morph/locset.cpp
@@ -2,13 +2,19 @@
#include <iostream>
#include <numeric>
+#include <arbor/math.hpp>
#include <arbor/morph/locset.hpp>
#include <arbor/morph/morphexcept.hpp>
#include <arbor/morph/morphology.hpp>
#include <arbor/morph/mprovider.hpp>
#include <arbor/morph/primitives.hpp>
+#include <arbor/morph/region.hpp>
+#include "util/cbrng.hpp"
+#include "util/partition.hpp"
#include "util/rangeutil.hpp"
+#include "util/transform.hpp"
+#include "util/span.hpp"
#include "util/strprintf.hpp"
namespace arb {
@@ -160,6 +166,125 @@ std::ostream& operator<<(std::ostream& o, const named_& x) {
return o << "(locset \"" << x.name << "\")";
}
+// Most distal points of a region
+
+struct most_distal_: locset_tag {
+ explicit most_distal_(region reg): reg(std::move(reg)) {}
+ region reg;
+};
+
+locset most_distal(region reg) {
+ return locset(most_distal_{std::move(reg)});
+}
+
+mlocation_list thingify_(const most_distal_& n, const mprovider& p) {
+ mlocation_list L;
+
+ auto cables = thingify(n.reg, p);
+ util::sort(cables, [](const auto& l, const auto& r){return (l.branch < r.branch) && (l.dist_pos < r.dist_pos);});
+
+ std::unordered_set<msize_t> branches_visited;
+ for (auto it= cables.rbegin(); it!= cables.rend(); it++) {
+ auto bid = (*it).branch;
+ auto pos = (*it).dist_pos;
+
+ // Check if any other points on the branch or any of its children has been added as a distal point
+ if (branches_visited.count(bid)) continue;
+ L.push_back({bid, pos});
+ while (bid != mnpos) {
+ branches_visited.insert(bid);
+ bid = p.morphology().branch_parent(bid);
+ }
+ }
+
+ util::sort(L);
+ return L;
+}
+
+std::ostream& operator<<(std::ostream& o, const most_distal_& x) {
+ return o << "(locset \"" << x.reg << "\")";
+}
+
+// Most distal points of a region
+
+struct most_proximal_: locset_tag {
+ explicit most_proximal_(region reg): reg(std::move(reg)) {}
+ region reg;
+};
+
+locset most_proximal(region reg) {
+ return locset(most_proximal_{std::move(reg)});
+}
+
+mlocation_list thingify_(const most_proximal_& n, const mprovider& p) {
+ auto cables = thingify(n.reg, p);
+ arb_assert(test_invariants(cables));
+
+ auto most_prox = cables.front();
+ return {{most_prox.branch, most_prox.prox_pos}};
+}
+
+std::ostream& operator<<(std::ostream& o, const most_proximal_& x) {
+ return o << "(locset \"" << x.reg << "\")";
+}
+
+
+// Uniform locset.
+
+struct uniform_ {
+ region reg;
+ unsigned left;
+ unsigned right;
+ uint64_t seed;
+};
+
+locset uniform(arb::region reg, unsigned left, unsigned right, uint64_t seed) {
+ return locset(uniform_{reg, left, right, seed});
+}
+
+mlocation_list thingify_(const uniform_& u, const mprovider& p) {
+ mlocation_list L;
+ auto morpho = p.morphology();
+ auto embed = p.embedding();
+
+ // Thingify the region and store relevant data
+ auto reg_cables = thingify(u.reg, p);
+
+ std::vector<double> lengths_bounds;
+ auto lengths_part = util::make_partition(lengths_bounds,
+ util::transform_view(reg_cables, [&embed](const auto& c) {
+ return embed.integrate_length(c);
+ }));
+
+ auto region_length = lengths_part.bounds().second;
+
+ // Generate uniform random positions along the extent of the full region
+ auto random_pos = util::uniform(u.seed, u.left, u.right);
+ std::transform(random_pos.begin(), random_pos.end(), random_pos.begin(),
+ [®ion_length](auto& c){return c*region_length;});
+ util::sort(random_pos);
+
+ // Match random_extents to cables and find position on the associated branch
+ unsigned cable_idx = 0;
+ auto range = lengths_part[cable_idx];
+
+ for (auto e: random_pos) {
+ while (e > range.second) {
+ range = lengths_part[++cable_idx];
+ }
+ auto cable = reg_cables[cable_idx];
+ auto pos_on_cable = (e - range.first)/(range.second - range.first);
+ auto pos_on_branch = math::lerp(cable.prox_pos, cable.dist_pos, pos_on_cable);
+ L.push_back({cable.branch, pos_on_branch});
+ }
+
+ return L;
+}
+
+std::ostream& operator<<(std::ostream& o, const uniform_& u) {
+ return o << "(uniform from region: \"" << u.reg << "\"; using seed: " << u.seed
+ << "; range: {" << u.left << ", " << u.right << "})";
+}
// Intersection of two point sets.
diff --git a/arbor/morph/region.cpp b/arbor/morph/region.cpp
index 3db12f50fb317606054a817f66befb99ca12446f..9aafa123c7beacdcf28f3b542a272093427e658b 100644
--- a/arbor/morph/region.cpp
+++ b/arbor/morph/region.cpp
@@ -1,6 +1,7 @@
#include <set>
#include <string>
#include <vector>
+#include <stack>
#include <arbor/morph/locset.hpp>
#include <arbor/morph/primitives.hpp>
@@ -163,6 +164,7 @@ mcable_list remove_covered_points(mcable_list cables, const morphology& m) {
}
}
}
+
util::sort(erase_indices);
for (auto it = erase_indices.rbegin(); it != erase_indices.rend(); it++) {
cables.erase(cables.begin() + *it);
@@ -283,7 +285,6 @@ std::ostream& operator<<(std::ostream& o, const tagged_& t) {
return o << "(tag " << t.tag << ")";
}
-
// Region comprising whole morphology.
struct all_: region_tag {};
@@ -306,9 +307,323 @@ std::ostream& operator<<(std::ostream& o, const all_& t) {
return o << "(all)";
}
+// Region with all segments distal from another region
-// Named region.
+struct distal_interval_ {
+ locset start;
+ double distance; //um
+};
+
+region distal_interval(locset start, double distance) {
+ return region(distal_interval_{start, distance});
+}
+
+mcable_list thingify_(const distal_interval_& reg, const mprovider& p) {
+ const auto& m = p.morphology();
+ const auto& e = p.embedding();
+
+ std::vector<mcable> L;
+
+ auto start = thingify(reg.start, p);
+ auto distance = reg.distance;
+
+ struct branch_interval {
+ msize_t bid;
+ double distance;
+ };
+
+ for (auto c: start) {
+ std::stack<branch_interval> branches_reached;
+ bool first_branch = true;
+
+ // if we're starting at the end of a branch, start traversal with its children
+ if (c.pos < 1) {
+ branches_reached.push({c.branch, distance});
+ } else {
+ first_branch = false;
+ L.push_back({c.branch,1,1});
+ for (auto child: m.branch_children(c.branch)) {
+ branches_reached.push({child, distance});
+ }
+ }
+
+ while (!branches_reached.empty()) {
+ auto bi = branches_reached.top();
+ branches_reached.pop();
+
+ auto branch = bi.bid;
+ auto rem_dist = bi.distance;
+
+ auto branch_length = e.branch_length(branch);
+ auto prox_pos = first_branch*c.pos;
+ auto dist_pos = rem_dist / branch_length + prox_pos;
+
+ if (dist_pos <= 1) {
+ L.push_back({branch, prox_pos, dist_pos});
+ } else {
+ L.push_back({branch, prox_pos, 1});
+ rem_dist = rem_dist - (1 - prox_pos)*branch_length;
+ for (auto child: m.branch_children(branch)) {
+ branches_reached.push({child, rem_dist});
+ }
+ }
+ first_branch = false;
+ }
+ }
+ return remove_covered_points(remove_cover(L, m), m);
+}
+
+std::ostream& operator<<(std::ostream& o, const distal_interval_& d) {
+ return o << "(distal_interval: " << d.start << ", " << d.distance << ")";
+}
+
+// Region with all segments proximal from another region
+
+struct proximal_interval_ {
+ locset end;
+ double distance; //um
+};
+
+region proximal_interval(locset end, double distance) {
+ return region(proximal_interval_{end, distance});
+}
+
+mcable_list thingify_(const proximal_interval_& reg, const mprovider& p) {
+ const auto& m = p.morphology();
+ const auto& e = p.embedding();
+
+ std::vector<mcable> L;
+
+ auto start = thingify(reg.end, p);
+ auto distance = reg.distance;
+
+ for (auto c: start) {
+ auto branch = c.branch;
+ auto branch_length = e.branch_length(branch);
+ auto rem_dist = distance;
+
+ auto dist_pos = c.pos;
+ auto prox_pos = dist_pos - distance / branch_length;
+
+ while (prox_pos < 0) {
+ L.push_back({branch, 0, dist_pos});
+
+ rem_dist = rem_dist - dist_pos*branch_length;
+
+ branch = m.branch_parent(branch);
+ if (branch == mnpos) {
+ break;
+ }
+
+ dist_pos = 1;
+ prox_pos = dist_pos - rem_dist / e.branch_length(branch);
+ }
+ if (branch != mnpos) {
+ L.push_back({branch, prox_pos, dist_pos});
+ }
+ }
+ return remove_cover(L, m);
+}
+
+std::ostream& operator<<(std::ostream& o, const proximal_interval_& d) {
+ return o << "(distal_interval: " << d.end << ", " << d.distance << ")";
+}
+
+mcable_list radius_cmp(const mprovider& p, region r, double v, comp_op op) {
+ const auto& e = p.embedding();
+
+ std::vector<mcable> L;
+ auto reg = thingify(r, p);
+ auto val = v;
+ for (auto c: reg) {
+ for (auto r: e.radius_cmp(c.branch, val, op)) {
+ if (is_disjoint(c, r)) continue;
+ L.push_back(make_intersection(c, r));
+ }
+ }
+ return remove_cover(L, p.morphology());
+}
+
+// Region with all segments with radius less than r
+struct radius_lt_ {
+ region reg;
+ double val; //um
+};
+
+region radius_lt(region reg, double val) {
+ return region(radius_lt_{reg, val});
+}
+
+mcable_list thingify_(const radius_lt_& r, const mprovider& p) {
+ return radius_cmp(p, r.reg, r.val, comp_op::lt);
+}
+
+std::ostream& operator<<(std::ostream& o, const radius_lt_& r) {
+ return o << "(radius_lt: " << r.reg << ", " << r.val << ")";
+}
+
+// Region with all segments with radius less than r
+struct radius_le_ {
+ region reg;
+ double val; //um
+};
+
+region radius_le(region reg, double val) {
+ return region(radius_le_{reg, val});
+}
+mcable_list thingify_(const radius_le_& r, const mprovider& p) {
+ return radius_cmp(p, r.reg, r.val, comp_op::le);
+}
+
+std::ostream& operator<<(std::ostream& o, const radius_le_& r) {
+ return o << "(radius_le: " << r.reg << ", " << r.val << ")";
+}
+
+// Region with all segments with radius greater than r
+struct radius_gt_ {
+ region reg;
+ double val; //um
+};
+
+region radius_gt(region reg, double val) {
+ return region(radius_gt_{reg, val});
+}
+
+mcable_list thingify_(const radius_gt_& r, const mprovider& p) {
+ return radius_cmp(p, r.reg, r.val, comp_op::gt);
+}
+
+std::ostream& operator<<(std::ostream& o, const radius_gt_& r) {
+ return o << "(radius_gt: " << r.reg << ", " << r.val << ")";
+}
+
+// Region with all segments with radius greater than or equal to r
+struct radius_ge_ {
+ region reg;
+ double val; //um
+};
+
+region radius_ge(region reg, double val) {
+ return region(radius_gt_{reg, val});
+}
+
+mcable_list thingify_(const radius_ge_& r, const mprovider& p) {
+ return radius_cmp(p, r.reg, r.val, comp_op::ge);
+}
+
+std::ostream& operator<<(std::ostream& o, const radius_ge_& r) {
+ return o << "(radius_ge: " << r.reg << ", " << r.val << ")";
+}
+
+mcable_list projection_cmp(const mprovider& p, double v, comp_op op) {
+ const auto& m = p.morphology();
+ const auto& e = p.embedding();
+
+ std::vector<mcable> L;
+ auto val = v;
+ for (auto i: util::make_span(m.num_branches())) {
+ util::append(L, e.projection_cmp(i, val, op));
+ }
+ return remove_cover(L, p.morphology());
+}
+
+// Region with all segments with projection less than val
+struct projection_lt_{
+ double val; //um
+};
+
+region projection_lt(double val) {
+ return region(projection_lt_{val});
+}
+
+mcable_list thingify_(const projection_lt_& r, const mprovider& p) {
+ return projection_cmp(p, r.val, comp_op::lt);
+}
+
+std::ostream& operator<<(std::ostream& o, const projection_lt_& r) {
+ return o << "(projection_lt: " << r.val << ")";
+}
+
+// Region with all segments with projection less than or equal to val
+struct projection_le_{
+ double val; //um
+};
+
+region projection_le(double val) {
+ return region(projection_le_{val});
+}
+
+mcable_list thingify_(const projection_le_& r, const mprovider& p) {
+ return projection_cmp(p, r.val, comp_op::le);
+}
+
+std::ostream& operator<<(std::ostream& o, const projection_le_& r) {
+ return o << "(projection_le: " << r.val << ")";
+}
+
+// Region with all segments with projection greater than val
+struct projection_gt_ {
+ double val; //um
+};
+
+region projection_gt(double val) {
+ return region(projection_gt_{val});
+}
+
+mcable_list thingify_(const projection_gt_& r, const mprovider& p) {
+ return projection_cmp(p, r.val, comp_op::gt);
+}
+
+std::ostream& operator<<(std::ostream& o, const projection_gt_& r) {
+ return o << "(projection_gt: " << r.val << ")";
+}
+
+// Region with all segments with projection greater than val
+struct projection_ge_ {
+ double val; //um
+};
+
+region projection_ge(double val) {
+ return region(projection_ge_{val});
+}
+
+mcable_list thingify_(const projection_ge_& r, const mprovider& p) {
+ return projection_cmp(p, r.val, comp_op::ge);
+}
+
+std::ostream& operator<<(std::ostream& o, const projection_ge_& r) {
+ return o << "(projection_ge: " << r.val << ")";
+}
+
+region z_dist_from_soma_lt(double r0) {
+ if (r0 == 0) {
+ return {};
+ }
+ region lt = reg::projection_lt(r0);
+ region gt = reg::projection_gt(-r0);
+ return intersect(std::move(lt), std::move(gt));
+}
+
+region z_dist_from_soma_le(double r0) {
+ region le = reg::projection_le(r0);
+ region ge = reg::projection_ge(-r0);
+ return intersect(std::move(le), std::move(ge));
+}
+
+region z_dist_from_soma_gt(double r0) {
+ region lt = reg::projection_lt(-r0);
+ region gt = reg::projection_gt(r0);
+ return region{join(std::move(lt), std::move(gt))};
+}
+
+region z_dist_from_soma_ge(double r0) {
+ region lt = reg::projection_le(-r0);
+ region gt = reg::projection_ge(r0);
+ return region{join(std::move(lt), std::move(gt))};
+}
+
+// Named region.
struct named_: region_tag {
explicit named_(std::string name): name(std::move(name)) {}
std::string name;
diff --git a/arbor/util/cbrng.hpp b/arbor/util/cbrng.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..6f78e44942c6f8ba882265f2b9d669c6541a3191
--- /dev/null
+++ b/arbor/util/cbrng.hpp
@@ -0,0 +1,41 @@
+#pragma once
+#include <vector>
+
+#include <Random123/threefry.h>
+#include <Random123/uniform.hpp>
+
+namespace arb {
+namespace util {
+
+std::vector<double> uniform(uint64_t seed, unsigned left, unsigned right) {
+ typedef r123::Threefry2x64 cbrng;
+ std::vector<double> r;
+
+ cbrng::key_type key = {{seed}};
+ cbrng::ctr_type ctr = {{0,0}};
+ cbrng g;
+
+ unsigned i = left;
+ if (i%2 && i<=right) {
+ ctr[0] = i/2;
+ cbrng::ctr_type rand = g(ctr, key);
+ r.push_back(r123::u01<double>(rand[1]));;
+ ++i;
+ }
+ while (i < 2*((right+1)/2)) {
+ ctr[0] = i/2;
+ cbrng::ctr_type rand = g(ctr, key);
+ r.push_back(r123::u01<double>(rand[0]));
+ r.push_back(r123::u01<double>(rand[1]));
+ i += 2;
+ }
+ if (i<=right) {
+ ctr[0] = i/2;
+ cbrng::ctr_type rand = g(ctr, key);
+ r.push_back(r123::u01<double>(rand[0]));
+ }
+ return r;
+}
+
+}
+}
\ No newline at end of file
diff --git a/ext/CMakeLists.txt b/ext/CMakeLists.txt
index 49aa903e132f58a154f43f7f64c3b22320ef7171..634a28898410c4bbd6d3217dc9d7a04655dd82f1 100644
--- a/ext/CMakeLists.txt
+++ b/ext/CMakeLists.txt
@@ -8,6 +8,12 @@ target_include_directories(ext-json INTERFACE json/single_include)
add_library(ext-tclap INTERFACE)
target_include_directories(ext-tclap INTERFACE tclap/include)
+# Random123 (DE Shaw Research) counter-based random number generators (header-only)
+
+add_library(ext-random123 INTERFACE)
+target_include_directories(ext-random123 INTERFACE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>/random123/include)
+install(TARGETS ext-random123 EXPORT arbor-targets)
+
# Google benchmark for microbenchmarks:
check_git_submodule(gbench google-benchmark)
diff --git a/ext/random123/LICENSE b/ext/random123/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..c6094acaf20c26df516f01549101dabc7fde5c93
--- /dev/null
+++ b/ext/random123/LICENSE
@@ -0,0 +1,31 @@
+/** @page LICENSE
+Copyright 2010-2012, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
diff --git a/ext/random123/include/Random123/MicroURNG.hpp b/ext/random123/include/Random123/MicroURNG.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..9ea77325514c07ab3e4393467dba018e42c4f412
--- /dev/null
+++ b/ext/random123/include/Random123/MicroURNG.hpp
@@ -0,0 +1,146 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __MicroURNG_dot_hpp__
+#define __MicroURNG_dot_hpp__
+
+#include <stdexcept>
+#include <limits>
+
+namespace r123{
+/**
+ Given a CBRNG whose ctr_type has an unsigned integral value_type,
+ MicroURNG<CBRNG>(c, k) is a type that satisfies the
+ requirements of a C++11 Uniform Random Number Generator.
+
+ The intended purpose is for a MicroURNG to be passed
+ as an argument to a C++11 Distribution, e.g.,
+ std::normal_distribution. See examples/MicroURNG.cpp.
+
+ The MicroURNG functor has a period of "only"
+
+ ctr_type.size()*2^32,
+
+ after which it will silently repeat.
+
+ The high 32 bits of the highest word in the counter c, passed to
+ the constructor must be zero. MicroURNG uses these bits to
+ "count".
+
+ Older versions of the library permitted a second template
+ parameter by which the caller could control the number of
+ bits devoted to the URNG's internal counter. This flexibility
+ has been disabled because URNGs created with different
+ numbers of counter bits could, conceivably "collide".
+
+\code
+ typedef ?someCBRNG? RNG;
+ RNG::ctr_type c = ...; // under application control
+ RNG::key_type k = ...; //
+ std::normal_distribution<float> nd;
+ MicroURNG<RNG> urng(c, k);
+ for(???){
+ ...
+ nd(urng); // may be called several hundred times with BITS=10
+ ...
+ }
+\endcode
+*/
+
+template<typename CBRNG>
+class MicroURNG{
+ // According to C++11, a URNG requires only a result_type,
+ // operator()(), min() and max() methods. Everything else
+ // (ctr_type, key_type, reset() method, etc.) is "value added"
+ // for the benefit of users that "know" that they're dealing with
+ // a MicroURNG.
+public:
+ typedef CBRNG cbrng_type;
+ static const int BITS = 32;
+ typedef typename cbrng_type::ctr_type ctr_type;
+ typedef typename cbrng_type::key_type key_type;
+ typedef typename cbrng_type::ukey_type ukey_type;
+ typedef typename ctr_type::value_type result_type;
+
+ R123_STATIC_ASSERT( std::numeric_limits<result_type>::digits >= BITS, "The result_type must have at least 32 bits" );
+
+ result_type operator()(){
+ if(last_elem == 0){
+ // jam n into the high bits of c
+ const size_t W = std::numeric_limits<result_type>::digits;
+ ctr_type c = c0;
+ c[c0.size()-1] |= n<<(W-BITS);
+ rdata = b(c,k);
+ n++;
+ last_elem = rdata.size();
+ }
+ return rdata[--last_elem];
+ }
+ MicroURNG(cbrng_type _b, ctr_type _c0, ukey_type _uk) : b(_b), c0(_c0), k(_uk), n(0), last_elem(0) {
+ chkhighbits();
+ }
+ MicroURNG(ctr_type _c0, ukey_type _uk) : b(), c0(_c0), k(_uk), n(0), last_elem(0) {
+ chkhighbits();
+ }
+
+ // _Min and _Max work around a bug in the library shipped with MacOS Xcode 4.5.2.
+ // See the commment in conventional/Engine.hpp.
+ const static result_type _Min = 0;
+ const static result_type _Max = ~((result_type)0);
+
+ static R123_CONSTEXPR result_type min R123_NO_MACRO_SUBST () { return _Min; }
+ static R123_CONSTEXPR result_type max R123_NO_MACRO_SUBST () { return _Max; }
+ // extra methods:
+ const ctr_type& counter() const{ return c0; }
+ void reset(ctr_type _c0, ukey_type _uk){
+ c0 = _c0;
+ chkhighbits();
+ k = _uk;
+ n = 0;
+ last_elem = 0;
+ }
+
+private:
+ cbrng_type b;
+ ctr_type c0;
+ key_type k;
+ R123_ULONG_LONG n;
+ size_t last_elem;
+ ctr_type rdata;
+ void chkhighbits(){
+ result_type r = c0[c0.size()-1];
+ result_type mask = ((uint64_t)std::numeric_limits<result_type>::max R123_NO_MACRO_SUBST ())>>BITS;
+ if((r&mask) != r)
+ throw std::runtime_error("MicroURNG: c0, does not have high bits clear");
+ }
+};
+} // namespace r123
+#endif
diff --git a/ext/random123/include/Random123/ReinterpretCtr.hpp b/ext/random123/include/Random123/ReinterpretCtr.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..164a38b0a56bc09232d9a7caa8601c29ebc2b0a9
--- /dev/null
+++ b/ext/random123/include/Random123/ReinterpretCtr.hpp
@@ -0,0 +1,88 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __ReinterpretCtr_dot_hpp__
+#define __ReinterpretCtr_dot_hpp__
+
+#include "features/compilerfeatures.h"
+#include <cstring>
+
+namespace r123{
+/*!
+ ReinterpretCtr uses memcpy to map back and forth
+ between a CBRNG's ctr_type and the specified ToType. For example,
+ after:
+
+ typedef ReinterpretCtr<r123array4x32, Philox2x64> G;
+
+ G is a bona fide CBRNG with ctr_type r123array4x32.
+
+ WARNING: ReinterpretCtr is endian dependent. The
+ values returned by G, declared as above,
+ will depend on the endianness of the machine on which it runs.
+ */
+
+template <typename ToType, typename CBRNG>
+struct ReinterpretCtr{
+ typedef ToType ctr_type;
+ typedef typename CBRNG::key_type key_type;
+ typedef typename CBRNG::ctr_type bctype;
+ typedef typename CBRNG::ukey_type ukey_type;
+ R123_STATIC_ASSERT(sizeof(ToType) == sizeof(bctype) && sizeof(typename bctype::value_type) != 16,
+ "ReinterpretCtr: sizeof(ToType) is not the same as sizeof(CBRNG::ctr_type) or CBRNG::ctr_type::value_type looks like it might be __m128i");
+ // It's amazingly difficult to safely do conversions with __m128i.
+ // If we use the operator() implementation below with a CBRNG
+ // whose ctr_type is r123array1xm128i, gcc4.6 optimizes away the
+ // memcpys, inlines the operator()(c,k), and produces assembly
+ // language that ends with an aesenclast instruction with a
+ // destination operand pointing to an unaligned memory address ...
+ // Segfault! See: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50444
+ // MSVC also produces code that crashes. We suspect a
+ // similar mechanism but haven't done the debugging necessary to
+ // be sure. We were able to 'fix' gcc4.6 by making bc a mutable
+ // data member rather than declaring it in the scope of
+ // operator(). That didn't fix the MSVC problems, though.
+ //
+ // Conclusion - don't touch __m128i, at least for now. The
+ // easiest (but highly imprecise) way to do that is the static
+ // assertion above that rejects bctype::value_types of size 16. -
+ // Sep 2011.
+ ctr_type operator()(ctr_type c, key_type k){
+ bctype bc;
+ std::memcpy(&bc, &c, sizeof(c));
+ CBRNG b;
+ bc = b(bc, k);
+ std::memcpy(&c, &bc, sizeof(bc));
+ return c;
+ }
+};
+} // namespace r123
+#endif
diff --git a/ext/random123/include/Random123/aes.h b/ext/random123/include/Random123/aes.h
new file mode 100644
index 0000000000000000000000000000000000000000..3095fac37de5d051b602fb53eb887ebf1ac59863
--- /dev/null
+++ b/ext/random123/include/Random123/aes.h
@@ -0,0 +1,398 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __Random123_aes_dot_hpp__
+#define __Random123_aes_dot_hpp__
+
+#include "features/compilerfeatures.h"
+#include "array.h"
+
+/* Implement a bona fide AES block cipher. It's minimally
+// checked against the test vector in FIPS-197 in ut_aes.cpp. */
+#if R123_USE_AES_NI
+
+/** @ingroup AESNI */
+typedef struct r123array1xm128i aesni1xm128i_ctr_t;
+/** @ingroup AESNI */
+typedef struct r123array1xm128i aesni1xm128i_ukey_t;
+/** @ingroup AESNI */
+typedef struct r123array4x32 aesni4x32_ukey_t;
+/** @ingroup AESNI */
+enum r123_enum_aesni1xm128i { aesni1xm128i_rounds = 10 };
+
+/** \cond HIDDEN_FROM_DOXYGEN */
+R123_STATIC_INLINE __m128i AES_128_ASSIST (__m128i temp1, __m128i temp2) {
+ __m128i temp3;
+ temp2 = _mm_shuffle_epi32 (temp2 ,0xff);
+ temp3 = _mm_slli_si128 (temp1, 0x4);
+ temp1 = _mm_xor_si128 (temp1, temp3);
+ temp3 = _mm_slli_si128 (temp3, 0x4);
+ temp1 = _mm_xor_si128 (temp1, temp3);
+ temp3 = _mm_slli_si128 (temp3, 0x4);
+ temp1 = _mm_xor_si128 (temp1, temp3);
+ temp1 = _mm_xor_si128 (temp1, temp2);
+ return temp1;
+}
+
+R123_STATIC_INLINE void aesni1xm128iexpand(aesni1xm128i_ukey_t uk, __m128i ret[11])
+{
+ __m128i rkey = uk.v[0].m;
+ __m128i tmp2;
+
+ ret[0] = rkey;
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x1);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[1] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x2);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[2] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x4);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[3] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x8);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[4] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x10);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[5] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x20);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[6] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x40);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[7] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x80);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[8] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x1b);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[9] = rkey;
+
+ tmp2 = _mm_aeskeygenassist_si128(rkey, 0x36);
+ rkey = AES_128_ASSIST(rkey, tmp2);
+ ret[10] = rkey;
+}
+/** \endcond */
+
+#ifdef __cplusplus
+/** @ingroup AESNI */
+struct aesni1xm128i_key_t{
+ __m128i k[11];
+ aesni1xm128i_key_t(){
+ aesni1xm128i_ukey_t uk;
+ uk.v[0].m = _mm_setzero_si128();
+ aesni1xm128iexpand(uk, k);
+ }
+ aesni1xm128i_key_t(const aesni1xm128i_ukey_t& uk){
+ aesni1xm128iexpand(uk, k);
+ }
+ aesni1xm128i_key_t(const aesni4x32_ukey_t& uk){
+ aesni1xm128i_ukey_t uk128;
+ uk128.v[0].m = _mm_set_epi32(uk.v[3], uk.v[2], uk.v[1], uk.v[0]);
+ aesni1xm128iexpand(uk128, k);
+ }
+ aesni1xm128i_key_t& operator=(const aesni1xm128i_ukey_t& uk){
+ aesni1xm128iexpand(uk, k);
+ return *this;
+ }
+ aesni1xm128i_key_t& operator=(const aesni4x32_ukey_t& uk){
+ aesni1xm128i_ukey_t uk128;
+ uk128.v[0].m = _mm_set_epi32(uk.v[3], uk.v[2], uk.v[1], uk.v[0]);
+ aesni1xm128iexpand(uk128, k);
+ return *this;
+ }
+ bool operator==(const aesni1xm128i_key_t& rhs) const{
+ for(int i=0; i<11; ++i){
+ // Sigh... No r123m128i(__m128i) constructor!
+ r123m128i li; li.m = k[i];
+ r123m128i ri; ri.m = rhs.k[i];
+ if( li != ri ) return false;
+ }
+ return true;
+ }
+ bool operator!=(const aesni1xm128i_key_t& rhs) const{
+ return !(*this == rhs);
+ }
+ friend std::ostream& operator<<(std::ostream& os, const aesni1xm128i_key_t& v){
+ r123m128i ki;
+ for(int i=0; i<10; ++i){
+ ki.m = v.k[i];
+ os << ki << " ";
+ }
+ ki.m = v.k[10];
+ return os << ki;
+ }
+ friend std::istream& operator>>(std::istream& is, aesni1xm128i_key_t& v){
+ r123m128i ki;
+ for(int i=0; i<11; ++i){
+ is >> ki;
+ v.k[i] = ki;
+ }
+ return is;
+ }
+};
+#else
+typedef struct {
+ __m128i k[11];
+}aesni1xm128i_key_t;
+
+/** @ingroup AESNI */
+R123_STATIC_INLINE aesni1xm128i_key_t aesni1xm128ikeyinit(aesni1xm128i_ukey_t uk){
+ aesni1xm128i_key_t ret;
+ aesni1xm128iexpand(uk, ret.k);
+ return ret;
+}
+#endif
+
+/** @ingroup AESNI */
+R123_STATIC_INLINE aesni1xm128i_ctr_t aesni1xm128i(aesni1xm128i_ctr_t in, aesni1xm128i_key_t k) {
+ __m128i x = _mm_xor_si128(k.k[0], in.v[0].m);
+ x = _mm_aesenc_si128(x, k.k[1]);
+ x = _mm_aesenc_si128(x, k.k[2]);
+ x = _mm_aesenc_si128(x, k.k[3]);
+ x = _mm_aesenc_si128(x, k.k[4]);
+ x = _mm_aesenc_si128(x, k.k[5]);
+ x = _mm_aesenc_si128(x, k.k[6]);
+ x = _mm_aesenc_si128(x, k.k[7]);
+ x = _mm_aesenc_si128(x, k.k[8]);
+ x = _mm_aesenc_si128(x, k.k[9]);
+ x = _mm_aesenclast_si128(x, k.k[10]);
+ {
+ aesni1xm128i_ctr_t ret;
+ ret.v[0].m = x;
+ return ret;
+ }
+}
+
+/** @ingroup AESNI */
+R123_STATIC_INLINE aesni1xm128i_ctr_t aesni1xm128i_R(unsigned R, aesni1xm128i_ctr_t in, aesni1xm128i_key_t k){
+ R123_ASSERT(R==10);
+ return aesni1xm128i(in, k);
+}
+
+
+/** @ingroup AESNI */
+typedef struct r123array4x32 aesni4x32_ctr_t;
+/** @ingroup AESNI */
+typedef aesni1xm128i_key_t aesni4x32_key_t;
+/** @ingroup AESNI */
+enum r123_enum_aesni4x32 { aesni4x32_rounds = 10 };
+/** @ingroup AESNI */
+R123_STATIC_INLINE aesni4x32_key_t aesni4x32keyinit(aesni4x32_ukey_t uk){
+ aesni1xm128i_ukey_t uk128;
+ aesni4x32_key_t ret;
+ uk128.v[0].m = _mm_set_epi32(uk.v[3], uk.v[2], uk.v[1], uk.v[0]);
+ aesni1xm128iexpand(uk128, ret.k);
+ return ret;
+}
+
+/** @ingroup AESNI */
+/** The aesni4x32_R function provides a C API to the @ref AESNI "AESNI" CBRNG, allowing the number of rounds to be specified explicitly **/
+R123_STATIC_INLINE aesni4x32_ctr_t aesni4x32_R(unsigned int Nrounds, aesni4x32_ctr_t c, aesni4x32_key_t k){
+ aesni1xm128i_ctr_t c128;
+ c128.v[0].m = _mm_set_epi32(c.v[3], c.v[2], c.v[1], c.v[0]);
+ c128 = aesni1xm128i_R(Nrounds, c128, k);
+ _mm_storeu_si128((__m128i*)&c.v[0], c128.v[0].m);
+ return c;
+}
+
+#define aesni4x32_rounds aesni1xm128i_rounds
+
+/** The aesni4x32 macro provides a C API to the @ref AESNI "AESNI" CBRNG, uses the default number of rounds i.e. \c aesni4x32_rounds **/
+/** @ingroup AESNI */
+#define aesni4x32(c,k) aesni4x32_R(aesni4x32_rounds, c, k)
+
+#ifdef __cplusplus
+namespace r123{
+/**
+@defgroup AESNI ARS and AESNI Classes and Typedefs
+
+The ARS4x32, ARS1xm128i, AESNI4x32 and AESNI1xm128i classes export the member functions, typedefs and
+operator overloads required by a @ref CBRNG "CBRNG" class.
+
+ARS1xm128i and AESNI1xm128i are based on the AES block cipher and rely on the AES-NI hardware instructions
+available on some some new (2011) CPUs.
+
+The ARS1xm128i CBRNG and the use of AES for random number generation are described in
+<a href="http://dl.acm.org/citation.cfm?doid=2063405"><i>Parallel Random Numbers: As Easy as 1, 2, 3</i> </a>.
+Although it uses some cryptographic primitives, ARS1xm128i uses a cryptographically weak key schedule and is \b not suitable for cryptographic use.
+
+@class AESNI1xm128i
+@ingroup AESNI
+AESNI exports the member functions, typedefs and operator overloads required by a @ref CBRNG class.
+
+AESNI1xm128i uses the crypotgraphic AES round function, including the cryptographic key schedule.
+
+In contrast to the other CBRNGs in the Random123 library, the AESNI1xm128i_R::key_type is opaque
+and is \b not identical to the AESNI1xm128i_R::ukey_type. Creating a key_type, using either the constructor
+or assignment operator, is significantly more time-consuming than running the bijection (hundreds
+of clock cycles vs. tens of clock cycles).
+
+AESNI1xm128i is only available when the feature-test macro R123_USE_AES_NI is true, which
+should occur only when the compiler is configured to generate AES-NI instructions (or
+when defaults are overridden by compile-time, compiler-command-line options).
+
+As of September 2011, the authors know of no statistical flaws with AESNI1xm128i. It
+would be an event of major cryptographic note if any such flaws were ever found.
+*/
+struct AESNI1xm128i{
+ typedef aesni1xm128i_ctr_t ctr_type;
+ typedef aesni1xm128i_ukey_t ukey_type;
+ typedef aesni1xm128i_key_t key_type;
+ static const unsigned int rounds=10;
+ ctr_type operator()(ctr_type ctr, key_type key) const{
+ return aesni1xm128i(ctr, key);
+ }
+};
+
+/* @class AESNI4x32 */
+struct AESNI4x32{
+ typedef aesni4x32_ctr_t ctr_type;
+ typedef aesni4x32_ukey_t ukey_type;
+ typedef aesni4x32_key_t key_type;
+ static const unsigned int rounds=10;
+ ctr_type operator()(ctr_type ctr, key_type key) const{
+ return aesni4x32(ctr, key);
+ }
+};
+
+/** @ingroup AESNI
+ @class AESNI1xm128i_R
+
+AESNI1xm128i_R is provided for completeness, but is only instantiable with ROUNDS=10, in
+which case it is identical to AESNI1xm128i */
+template <unsigned ROUNDS=10>
+struct AESNI1xm128i_R : public AESNI1xm128i{
+ R123_STATIC_ASSERT(ROUNDS==10, "AESNI1xm128i_R<R> is only valid with R=10");
+};
+
+/** @class AESNI4x32_R **/
+template <unsigned ROUNDS=10>
+struct AESNI4x32_R : public AESNI4x32{
+ R123_STATIC_ASSERT(ROUNDS==10, "AESNI4x32_R<R> is only valid with R=10");
+};
+} // namespace r123
+#endif /* __cplusplus */
+
+#endif /* R123_USE_AES_NI */
+
+#if R123_USE_AES_OPENSSL
+#include "string.h"
+#include <openssl/aes.h>
+typedef struct r123array16x8 aesopenssl16x8_ctr_t;
+typedef struct r123array16x8 aesopenssl16x8_ukey_t;
+#ifdef __cplusplus
+struct aesopenssl16x8_key_t{
+ AES_KEY k;
+ aesopenssl16x8_key_t(){
+ aesopenssl16x8_ukey_t ukey={{}};
+ AES_set_encrypt_key((const unsigned char *)&ukey.v[0], 128, &k);
+ }
+ aesopenssl16x8_key_t(const aesopenssl16x8_ukey_t& ukey){
+ AES_set_encrypt_key((const unsigned char *)&ukey.v[0], 128, &k);
+ }
+ aesopenssl16x8_key_t& operator=(const aesopenssl16x8_ukey_t& ukey){
+ AES_set_encrypt_key((const unsigned char *)&ukey.v[0], 128, &k);
+ return *this;
+ }
+ bool operator==(const aesopenssl16x8_key_t& rhs) const{
+ return (k.rounds == rhs.k.rounds) && 0==::memcmp(&k.rd_key[0], &rhs.k.rd_key[0], (k.rounds+1) * 4 * sizeof(uint32_t));
+ }
+ bool operator!=(const aesopenssl16x8_key_t& rhs) const{
+ return !(*this == rhs);
+ }
+ friend std::ostream& operator<<(std::ostream& os, const aesopenssl16x8_key_t& v){
+ os << v.k.rounds;
+ const unsigned int *p = &v.k.rd_key[0];
+ for(int i=0; i<(v.k.rounds+1); ++i){
+ os << " " << p[0] << " " << p[1] << " " << p[2] << " " << p[3];
+ p += 4;
+ }
+ return os;
+ }
+ friend std::istream& operator>>(std::istream& is, aesopenssl16x8_key_t& v){
+ is >> v.k.rounds;
+ unsigned int *p = &v.k.rd_key[0];
+ for(int i=0; i<(v.k.rounds+1); ++i){
+ is >> p[0] >> p[1] >> p[2] >> p[3];
+ p += 4;
+ }
+ return is;
+ }
+};
+#else
+typedef struct aesopenssl16x8_key_t{
+ AES_KEY k;
+}aesopenssl16x8_key_t;
+R123_STATIC_INLINE struct aesopenssl16x8_key_t aesopenssl16x8keyinit(aesopenssl16x8_ukey_t uk){
+ aesopenssl16x8_key_t ret;
+ AES_set_encrypt_key((const unsigned char *)&uk.v[0], 128, &ret.k);
+ return ret;
+}
+#endif
+
+R123_STATIC_INLINE R123_FORCE_INLINE(aesopenssl16x8_ctr_t aesopenssl16x8_R(aesopenssl16x8_ctr_t ctr, aesopenssl16x8_key_t key));
+R123_STATIC_INLINE
+aesopenssl16x8_ctr_t aesopenssl16x8_R(aesopenssl16x8_ctr_t ctr, aesopenssl16x8_key_t key){
+ aesopenssl16x8_ctr_t ret;
+ AES_encrypt((const unsigned char*)&ctr.v[0], (unsigned char *)&ret.v[0], &key.k);
+ return ret;
+}
+
+#define aesopenssl16x8_rounds aesni4x32_rounds
+#define aesopenssl16x8(c,k) aesopenssl16x8_R(aesopenssl16x8_rounds)
+
+#ifdef __cplusplus
+namespace r123{
+struct AESOpenSSL16x8{
+ typedef aesopenssl16x8_ctr_t ctr_type;
+ typedef aesopenssl16x8_key_t key_type;
+ typedef aesopenssl16x8_ukey_t ukey_type;
+ static const unsigned int rounds=10;
+ ctr_type operator()(const ctr_type& in, const key_type& k){
+ ctr_type out;
+ AES_encrypt((const unsigned char *)&in[0], (unsigned char *)&out[0], &k.k);
+ return out;
+ }
+};
+} // namespace r123
+#endif /* __cplusplus */
+#endif /* R123_USE_AES_OPENSSL */
+
+#endif
diff --git a/ext/random123/include/Random123/array.h b/ext/random123/include/Random123/array.h
new file mode 100644
index 0000000000000000000000000000000000000000..c560c3fee1a860f5c8f1785c1b181961be285226
--- /dev/null
+++ b/ext/random123/include/Random123/array.h
@@ -0,0 +1,348 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef _r123array_dot_h__
+#define _r123array_dot_h__
+#include "features/compilerfeatures.h"
+#include "features/sse.h"
+
+#if !defined(__cplusplus) || defined(__METAL_MACOS__)
+#define CXXMETHODS(_N, W, T)
+#define CXXOVERLOADS(_N, W, T)
+#define CXXMETHODS_REQUIRING_STL
+#else
+
+#include <stddef.h>
+#include <algorithm>
+#include <stdexcept>
+#include <iterator>
+#include <limits>
+#include <iostream>
+
+/** @defgroup arrayNxW The r123arrayNxW classes
+
+ Each of the r123arrayNxW is a fixed size array of N W-bit unsigned integers.
+ It is functionally equivalent to the C++11 std::array<N, uintW_t>,
+ but does not require C++11 features or libraries.
+
+ In addition to meeting most of the requirements of a Container,
+ it also has a member function, incr(), which increments the zero-th
+ element and carrys overflows into higher indexed elements. Thus,
+ by using incr(), sequences of up to 2^(N*W) distinct values
+ can be produced.
+
+ If SSE is supported by the compiler, then the class
+ r123array1xm128i is also defined, in which the data member is an
+ array of one r123m128i object.
+
+ When compiling with __CUDA_ARCH__ defined, the reverse iterator
+ methods (rbegin, rend, crbegin, crend) are not defined because
+ CUDA does not support std::reverse_iterator.
+
+*/
+
+/** @cond HIDDEN_FROM_DOXYGEN */
+
+template <typename value_type>
+inline R123_CUDA_DEVICE value_type assemble_from_u32(uint32_t *p32){
+ value_type v=0;
+ for(size_t i=0; i<(3+sizeof(value_type))/4; ++i)
+ v |= ((value_type)(*p32++)) << (32*i);
+ return v;
+}
+
+/** @endcond */
+
+#ifdef __CUDA_ARCH__
+/* CUDA can't handle std::reverse_iterator. We *could* implement it
+ ourselves, but let's not bother until somebody really feels a need
+ to reverse-iterate through an r123array */
+#define CXXMETHODS_REQUIRING_STL
+#else
+#define CXXMETHODS_REQUIRING_STL \
+ public: \
+ typedef std::reverse_iterator<iterator> reverse_iterator; \
+ typedef std::reverse_iterator<const_iterator> const_reverse_iterator; \
+ R123_CUDA_DEVICE reverse_iterator rbegin(){ return reverse_iterator(end()); } \
+ R123_CUDA_DEVICE const_reverse_iterator rbegin() const{ return const_reverse_iterator(end()); } \
+ R123_CUDA_DEVICE reverse_iterator rend(){ return reverse_iterator(begin()); } \
+ R123_CUDA_DEVICE const_reverse_iterator rend() const{ return const_reverse_iterator(begin()); } \
+ R123_CUDA_DEVICE const_reverse_iterator crbegin() const{ return const_reverse_iterator(cend()); } \
+ R123_CUDA_DEVICE const_reverse_iterator crend() const{ return const_reverse_iterator(cbegin()); }
+#endif
+
+// Work-alike methods and typedefs modeled on std::array:
+#define CXXMETHODS(_N, W, T) \
+ typedef T value_type; \
+ typedef T* iterator; \
+ typedef const T* const_iterator; \
+ typedef value_type& reference; \
+ typedef const value_type& const_reference; \
+ typedef size_t size_type; \
+ typedef ptrdiff_t difference_type; \
+ typedef T* pointer; \
+ typedef const T* const_pointer; \
+ /* Boost.array has static_size. C++11 specializes tuple_size */ \
+ enum {static_size = _N}; \
+ R123_CUDA_DEVICE reference operator[](size_type i){return v[i];} \
+ R123_CUDA_DEVICE const_reference operator[](size_type i) const {return v[i];} \
+ R123_CUDA_DEVICE reference at(size_type i){ if(i >= _N) R123_THROW(std::out_of_range("array index out of range")); return (*this)[i]; } \
+ R123_CUDA_DEVICE const_reference at(size_type i) const { if(i >= _N) R123_THROW(std::out_of_range("array index out of range")); return (*this)[i]; } \
+ R123_CUDA_DEVICE size_type size() const { return _N; } \
+ R123_CUDA_DEVICE size_type max_size() const { return _N; } \
+ R123_CUDA_DEVICE bool empty() const { return _N==0; }; \
+ R123_CUDA_DEVICE iterator begin() { return &v[0]; } \
+ R123_CUDA_DEVICE iterator end() { return &v[_N]; } \
+ R123_CUDA_DEVICE const_iterator begin() const { return &v[0]; } \
+ R123_CUDA_DEVICE const_iterator end() const { return &v[_N]; } \
+ R123_CUDA_DEVICE const_iterator cbegin() const { return &v[0]; } \
+ R123_CUDA_DEVICE const_iterator cend() const { return &v[_N]; } \
+ R123_CUDA_DEVICE pointer data(){ return &v[0]; } \
+ R123_CUDA_DEVICE const_pointer data() const{ return &v[0]; } \
+ R123_CUDA_DEVICE reference front(){ return v[0]; } \
+ R123_CUDA_DEVICE const_reference front() const{ return v[0]; } \
+ R123_CUDA_DEVICE reference back(){ return v[_N-1]; } \
+ R123_CUDA_DEVICE const_reference back() const{ return v[_N-1]; } \
+ R123_CUDA_DEVICE bool operator==(const r123array##_N##x##W& rhs) const{ \
+ /* CUDA3 does not have std::equal */ \
+ for (size_t i = 0; i < _N; ++i) \
+ if (v[i] != rhs.v[i]) return false; \
+ return true; \
+ } \
+ R123_CUDA_DEVICE bool operator!=(const r123array##_N##x##W& rhs) const{ return !(*this == rhs); } \
+ /* CUDA3 does not have std::fill_n */ \
+ R123_CUDA_DEVICE void fill(const value_type& val){ for (size_t i = 0; i < _N; ++i) v[i] = val; } \
+ R123_CUDA_DEVICE void swap(r123array##_N##x##W& rhs){ \
+ /* CUDA3 does not have std::swap_ranges */ \
+ for (size_t i = 0; i < _N; ++i) { \
+ T tmp = v[i]; \
+ v[i] = rhs.v[i]; \
+ rhs.v[i] = tmp; \
+ } \
+ } \
+ R123_CUDA_DEVICE r123array##_N##x##W& incr(R123_ULONG_LONG n=1){ \
+ /* This test is tricky because we're trying to avoid spurious \
+ complaints about illegal shifts, yet still be compile-time \
+ evaulated. */ \
+ if(sizeof(T)<sizeof(n) && n>>((sizeof(T)<sizeof(n))?8*sizeof(T):0) ) \
+ return incr_carefully(n); \
+ if(n==1){ \
+ ++v[0]; \
+ if(_N==1 || R123_BUILTIN_EXPECT(!!v[0], 1)) return *this; \
+ }else{ \
+ v[0] += n; \
+ if(_N==1 || R123_BUILTIN_EXPECT(n<=v[0], 1)) return *this; \
+ } \
+ /* We expect that the N==?? tests will be \
+ constant-folded/optimized away by the compiler, so only the \
+ overflow tests (!!v[i]) remain to be done at runtime. For \
+ small values of N, it would be better to do this as an \
+ uncondtional sequence of adc. An experiment/optimization \
+ for another day... \
+ N.B. The weird subscripting: v[_N>3?3:0] is to silence \
+ a spurious error from icpc \
+ */ \
+ ++v[_N>1?1:0]; \
+ if(_N==2 || R123_BUILTIN_EXPECT(!!v[_N>1?1:0], 1)) return *this; \
+ ++v[_N>2?2:0]; \
+ if(_N==3 || R123_BUILTIN_EXPECT(!!v[_N>2?2:0], 1)) return *this; \
+ ++v[_N>3?3:0]; \
+ for(size_t i=4; i<_N; ++i){ \
+ if( R123_BUILTIN_EXPECT(!!v[i-1], 1) ) return *this; \
+ ++v[i]; \
+ } \
+ return *this; \
+ } \
+ /* seed(SeedSeq) would be a constructor if having a constructor */ \
+ /* didn't cause headaches with defaults */ \
+ template <typename SeedSeq> \
+ R123_CUDA_DEVICE static r123array##_N##x##W seed(SeedSeq &ss){ \
+ r123array##_N##x##W ret; \
+ const size_t Ngen = _N*((3+sizeof(value_type))/4); \
+ uint32_t u32[Ngen]; \
+ uint32_t *p32 = &u32[0]; \
+ ss.generate(&u32[0], &u32[Ngen]); \
+ for(size_t i=0; i<_N; ++i){ \
+ ret.v[i] = assemble_from_u32<value_type>(p32); \
+ p32 += (3+sizeof(value_type))/4; \
+ } \
+ return ret; \
+ } \
+protected: \
+ R123_CUDA_DEVICE r123array##_N##x##W& incr_carefully(R123_ULONG_LONG n){ \
+ /* n may be greater than the maximum value of a single value_type */ \
+ value_type vtn; \
+ vtn = n; \
+ v[0] += n; \
+ const unsigned rshift = 8* ((sizeof(n)>sizeof(value_type))? sizeof(value_type) : 0); \
+ for(size_t i=1; i<_N; ++i){ \
+ if(rshift){ \
+ n >>= rshift; \
+ }else{ \
+ n=0; \
+ } \
+ if( v[i-1] < vtn ) \
+ ++n; \
+ if( n==0 ) break; \
+ vtn = n; \
+ v[i] += n; \
+ } \
+ return *this; \
+ } \
+
+/** @cond HIDDEN_FROM_DOXYGEN */
+
+// There are several tricky considerations for the insertion and extraction
+// operators:
+// - we would like to be able to print r123array16x8 as a sequence of 16 integers,
+// not as 16 bytes.
+// - we would like to be able to print r123array1xm128i.
+// - we do not want an int conversion operator in r123m128i because it causes
+// lots of ambiguity problems with automatic promotions.
+// Solution: r123arrayinsertable and r123arrayextractable
+
+template<typename T>
+struct r123arrayinsertable{
+ const T& v;
+ r123arrayinsertable(const T& t_) : v(t_) {}
+ friend std::ostream& operator<<(std::ostream& os, const r123arrayinsertable<T>& t){
+ return os << t.v;
+ }
+};
+
+template<>
+struct r123arrayinsertable<uint8_t>{
+ const uint8_t& v;
+ r123arrayinsertable(const uint8_t& t_) : v(t_) {}
+ friend std::ostream& operator<<(std::ostream& os, const r123arrayinsertable<uint8_t>& t){
+ return os << (int)t.v;
+ }
+};
+
+template<typename T>
+struct r123arrayextractable{
+ T& v;
+ r123arrayextractable(T& t_) : v(t_) {}
+ friend std::istream& operator>>(std::istream& is, r123arrayextractable<T>& t){
+ return is >> t.v;
+ }
+};
+
+template<>
+struct r123arrayextractable<uint8_t>{
+ uint8_t& v;
+ r123arrayextractable(uint8_t& t_) : v(t_) {}
+ friend std::istream& operator>>(std::istream& is, r123arrayextractable<uint8_t>& t){
+ int i;
+ is >> i;
+ t.v = i;
+ return is;
+ }
+};
+/** @endcond */
+
+#define CXXOVERLOADS(_N, W, T) \
+ \
+inline std::ostream& operator<<(std::ostream& os, const r123array##_N##x##W& a){ \
+ os << r123arrayinsertable<T>(a.v[0]); \
+ for(size_t i=1; i<_N; ++i) \
+ os << " " << r123arrayinsertable<T>(a.v[i]); \
+ return os; \
+} \
+ \
+inline std::istream& operator>>(std::istream& is, r123array##_N##x##W& a){ \
+ for(size_t i=0; i<_N; ++i){ \
+ r123arrayextractable<T> x(a.v[i]); \
+ is >> x; \
+ } \
+ return is; \
+} \
+ \
+namespace r123{ \
+ typedef r123array##_N##x##W Array##_N##x##W; \
+}
+
+#endif /* __cplusplus */
+
+/* _r123array_tpl expands to a declaration of struct r123arrayNxW.
+
+ In C, it's nothing more than a struct containing an array of N
+ objects of type T.
+
+ In C++ it's the same, but endowed with an assortment of member
+ functions, typedefs and friends. In C++, r123arrayNxW looks a lot
+ like std::array<T,N>, has most of the capabilities of a container,
+ and satisfies the requirements outlined in compat/Engine.hpp for
+ counter and key types. ArrayNxW, in the r123 namespace is
+ a typedef equivalent to r123arrayNxW.
+*/
+
+#define _r123array_tpl(_N, W, T) \
+ /** @ingroup arrayNxW */ \
+ /** @see arrayNxW */ \
+struct r123array##_N##x##W{ \
+ T v[_N]; \
+ CXXMETHODS(_N, W, T) \
+ CXXMETHODS_REQUIRING_STL \
+}; \
+ \
+CXXOVERLOADS(_N, W, T)
+
+
+_r123array_tpl(1, 32, uint32_t) /* r123array1x32 */
+_r123array_tpl(2, 32, uint32_t) /* r123array2x32 */
+_r123array_tpl(4, 32, uint32_t) /* r123array4x32 */
+_r123array_tpl(8, 32, uint32_t) /* r123array8x32 */
+
+#if R123_USE_64BIT
+_r123array_tpl(1, 64, uint64_t) /* r123array1x64 */
+_r123array_tpl(2, 64, uint64_t) /* r123array2x64 */
+_r123array_tpl(4, 64, uint64_t) /* r123array4x64 */
+#endif
+
+_r123array_tpl(16, 8, uint8_t) /* r123array16x8 for ARSsw, AESsw */
+
+#if R123_USE_SSE
+_r123array_tpl(1, m128i, r123m128i) /* r123array1x128i for ARSni, AESni */
+#endif
+
+/* In C++, it's natural to use sizeof(a::value_type), but in C it's
+ pretty convoluted to figure out the width of the value_type of an
+ r123arrayNxW:
+*/
+#define R123_W(a) (8*sizeof(((a *)0)->v[0]))
+
+/** @namespace r123
+ Most of the Random123 C++ API is contained in the r123 namespace.
+*/
+
+#endif
+
diff --git a/ext/random123/include/Random123/ars.h b/ext/random123/include/Random123/ars.h
new file mode 100644
index 0000000000000000000000000000000000000000..a027b6fe043e55da528569cf4b22aa10296fd6a8
--- /dev/null
+++ b/ext/random123/include/Random123/ars.h
@@ -0,0 +1,204 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __Random123_ars_dot_hpp__
+#define __Random123_ars_dot_hpp__
+
+#include "features/compilerfeatures.h"
+#include "array.h"
+
+#if R123_USE_AES_NI
+
+#ifndef ARS1xm128i_DEFAULT_ROUNDS
+#define ARS1xm128i_DEFAULT_ROUNDS 7
+#endif
+
+/** @ingroup AESNI */
+enum r123_enum_ars1xm128i {ars1xm128i_rounds = ARS1xm128i_DEFAULT_ROUNDS};
+
+/* ARS1xm128i with Weyl keys. Fast, and Crush-resistant, but NOT CRYPTO. */
+/** @ingroup AESNI */
+typedef struct r123array1xm128i ars1xm128i_ctr_t;
+/** @ingroup AESNI */
+typedef struct r123array1xm128i ars1xm128i_key_t;
+/** @ingroup AESNI */
+typedef struct r123array1xm128i ars1xm128i_ukey_t;
+/** @ingroup AESNI */
+R123_STATIC_INLINE ars1xm128i_key_t ars1xm128ikeyinit(ars1xm128i_ukey_t uk) { return uk; }
+/** @ingroup AESNI */
+R123_STATIC_INLINE ars1xm128i_ctr_t ars1xm128i_R(unsigned int Nrounds, ars1xm128i_ctr_t in, ars1xm128i_key_t k){
+ __m128i kweyl = _mm_set_epi64x(R123_64BIT(0xBB67AE8584CAA73B), /* sqrt(3) - 1.0 */
+ R123_64BIT(0x9E3779B97F4A7C15)); /* golden ratio */
+ /* N.B. the aesenc instructions do the xor *after*
+ // so if we want to follow the AES pattern, we
+ // have to do the initial xor explicitly */
+ __m128i kk = k.v[0].m;
+ __m128i v = _mm_xor_si128(in.v[0].m, kk);
+ ars1xm128i_ctr_t ret;
+ R123_ASSERT(Nrounds<=10);
+ if( Nrounds>1 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>2 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>3 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>4 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>5 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>6 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>7 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>8 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ if( Nrounds>9 ){
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenc_si128(v, kk);
+ }
+ kk = _mm_add_epi64(kk, kweyl);
+ v = _mm_aesenclast_si128(v, kk);
+ ret.v[0].m = v;
+ return ret;
+}
+
+/** @def ars1xm128i
+@ingroup AESNI
+The ars1mx128i macro provides a C API interface to the @ref AESNI "ARS" CBRNG with the default number of rounds i.e. \c ars1xm128i_rounds **/
+#define ars1xm128i(c,k) ars1xm128i_R(ars1xm128i_rounds, c, k)
+
+/** @ingroup AESNI */
+typedef struct r123array4x32 ars4x32_ctr_t;
+/** @ingroup AESNI */
+typedef struct r123array4x32 ars4x32_key_t;
+/** @ingroup AESNI */
+typedef struct r123array4x32 ars4x32_ukey_t;
+/** @ingroup AESNI */
+enum r123_enum_ars4x32 {ars4x32_rounds = ARS1xm128i_DEFAULT_ROUNDS};
+/** @ingroup AESNI */
+R123_STATIC_INLINE ars4x32_key_t ars4x32keyinit(ars4x32_ukey_t uk) { return uk; }
+/** @ingroup AESNI */
+R123_STATIC_INLINE ars4x32_ctr_t ars4x32_R(unsigned int Nrounds, ars4x32_ctr_t c, ars4x32_key_t k){
+ ars1xm128i_ctr_t c128;
+ ars1xm128i_key_t k128;
+ c128.v[0].m = _mm_set_epi32(c.v[3], c.v[2], c.v[1], c.v[0]);
+ k128.v[0].m = _mm_set_epi32(k.v[3], k.v[2], k.v[1], k.v[0]);
+ c128 = ars1xm128i_R(Nrounds, c128, k128);
+ _mm_storeu_si128((__m128i*)&c.v[0], c128.v[0].m);
+ return c;
+}
+
+/** @def ars4x32
+@ingroup AESNI
+The ars4x32 macro provides a C API interface to the @ref AESNI "ARS" CBRNG with the default number of rounds i.e. \c ars4x32_rounds **/
+#define ars4x32(c,k) ars4x32_R(ars4x32_rounds, c, k)
+
+#ifdef __cplusplus
+namespace r123{
+/**
+@ingroup AESNI
+
+ARS1xm128i_R exports the member functions, typedefs and operator overloads required by a @ref CBRNG class.
+
+ARS1xm128i uses the crypotgraphic AES round function, but a @b non-cryptographc key schedule
+to save time and space.
+
+ARS1xm128i is only available when the feature-test macro R123_USE_AES_NI is true, which
+should occur only when the compiler is configured to generate AES-NI instructions (or
+when defaults are overridden by compile-time, compiler-command-line options).
+
+The template argument, ROUNDS, is the number of times the ARS round
+functions will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=5 or more.
+
+@class ARS1xm128i_R
+
+*/
+template<unsigned int ROUNDS>
+struct ARS1xm128i_R{
+ typedef ars1xm128i_ctr_t ctr_type;
+ typedef ars1xm128i_key_t key_type;
+ typedef ars1xm128i_key_t ukey_type;
+ static const unsigned int rounds=ROUNDS;
+ R123_FORCE_INLINE(ctr_type operator()(ctr_type ctr, key_type key) const){
+ return ars1xm128i_R(ROUNDS, ctr, key);
+ }
+};
+
+/** @class ARS4x32_R
+ @ingroup AESNI
+*/
+
+template<unsigned int ROUNDS>
+struct ARS4x32_R{
+ typedef ars4x32_ctr_t ctr_type;
+ typedef ars4x32_key_t key_type;
+ typedef ars4x32_key_t ukey_type;
+ static const unsigned int rounds=ROUNDS;
+ R123_FORCE_INLINE(ctr_type operator()(ctr_type ctr, key_type key) const){
+ return ars4x32_R(ROUNDS, ctr, key);
+ }
+};
+/**
+@ingroup AESNI
+
+@class ARS1xm128i_R
+ ARS1xm128i is equivalent to ARS1xm128i_R<7>. With 7 rounds,
+ the ARS1xm128i CBRNG has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance. */
+typedef ARS1xm128i_R<ars1xm128i_rounds> ARS1xm128i;
+typedef ARS4x32_R<ars4x32_rounds> ARS4x32;
+} // namespace r123
+
+#endif /* __cplusplus */
+
+#endif /* R123_USE_AES_NI */
+
+#endif
diff --git a/ext/random123/include/Random123/boxmuller.hpp b/ext/random123/include/Random123/boxmuller.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..9c91cf879109133a80844c7c69f26d8e448578fa
--- /dev/null
+++ b/ext/random123/include/Random123/boxmuller.hpp
@@ -0,0 +1,139 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+// This file implements the Box-Muller method for generating gaussian
+// random variables (GRVs). Box-Muller has the advantage of
+// deterministically requiring exactly two uniform random variables as
+// input and producing exactly two GRVs as output, which makes it
+// especially well-suited to the counter-based generators in
+// Random123. Other methods (e.g., Ziggurat, polar) require an
+// indeterminate number of inputs for each output and so require a
+// 'MicroURNG' to be used with Random123. The down side of Box-Muller
+// is that it calls sincos, log and sqrt, which may be slow. However,
+// on GPUs, these functions are remarkably fast, which makes
+// Box-Muller the fastest GRV generator we know of on GPUs.
+//
+// This file exports two structs and one overloaded function,
+// all in the r123 namespace:
+// struct r123::float2{ float x,y; }
+// struct r123::double2{ double x,y; }
+//
+// r123::float2 r123::boxmuller(uint32_t u0, uint32_t u1);
+// r123::double2 r123::boxmuller(uint64_t u0, uint64_t u1);
+//
+// float2 and double2 are identical to their synonymous global-
+// namespace structures in CUDA.
+//
+// This file may not be as portable, and has not been tested as
+// rigorously as other files in the library, e.g., the generators.
+// Nevertheless, we hope it is useful and we encourage developers to
+// copy it and modify it for their own use. We invite comments and
+// improvements.
+
+#ifndef _r123_BOXMULLER_HPP__
+#define _r123_BOXMULLER_HPP__
+
+#include <Random123/features/compilerfeatures.h>
+#include <Random123/uniform.hpp>
+#include <math.h>
+
+namespace r123{
+
+#if !defined(__CUDACC__)
+typedef struct { float x, y; } float2;
+typedef struct { double x, y; } double2;
+#else
+typedef ::float2 float2;
+typedef ::double2 double2;
+#endif
+
+#if !defined(R123_NO_SINCOS) && defined(__APPLE__)
+/* MacOS X 10.10.5 (2015) doesn't have sincosf */
+#define R123_NO_SINCOS 1
+#endif
+
+#if R123_NO_SINCOS /* enable this if sincos and sincosf are not in the math library */
+R123_CUDA_DEVICE R123_STATIC_INLINE void sincosf(float x, float *s, float *c) {
+ *s = sinf(x);
+ *c = cosf(x);
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE void sincos(double x, double *s, double *c) {
+ *s = sin(x);
+ *c = cos(x);
+}
+#endif /* sincos is not in the math library */
+
+#if !defined(CUDART_VERSION) || CUDART_VERSION < 5000 /* enabled if sincospi and sincospif are not in math lib */
+
+R123_CUDA_DEVICE R123_STATIC_INLINE void sincospif(float x, float *s, float *c){
+ const float PIf = 3.1415926535897932f;
+ sincosf(PIf*x, s, c);
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE void sincospi(double x, double *s, double *c) {
+ const double PI = 3.1415926535897932;
+ sincos(PI*x, s, c);
+}
+#endif /* sincospi is not in math lib */
+
+/*
+ * take two 32bit unsigned random values and return a float2 with
+ * two random floats in a normal distribution via a Box-Muller transform
+ */
+R123_CUDA_DEVICE R123_STATIC_INLINE float2 boxmuller(uint32_t u0, uint32_t u1) {
+ float r;
+ float2 f;
+ sincospif(uneg11<float>(u0), &f.x, &f.y);
+ r = sqrtf(-2.f * logf(u01<float>(u1))); // u01 is guaranteed to avoid 0.
+ f.x *= r;
+ f.y *= r;
+ return f;
+}
+
+/*
+ * take two 64bit unsigned random values and return a double2 with
+ * two random doubles in a normal distribution via a Box-Muller transform
+ */
+R123_CUDA_DEVICE R123_STATIC_INLINE double2 boxmuller(uint64_t u0, uint64_t u1) {
+ double r;
+ double2 f;
+
+ sincospi(uneg11<double>(u0), &f.x, &f.y);
+ r = sqrt(-2. * log(u01<double>(u1))); // u01 is guaranteed to avoid 0.
+ f.x *= r;
+ f.y *= r;
+ return f;
+}
+} // namespace r123
+
+#endif /* BOXMULLER_H__ */
diff --git a/ext/random123/include/Random123/conventional/Engine.hpp b/ext/random123/include/Random123/conventional/Engine.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..bd2da2e1911b789adbba32c51e9835128ed96842
--- /dev/null
+++ b/ext/random123/include/Random123/conventional/Engine.hpp
@@ -0,0 +1,276 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __Engine_dot_hpp_
+#define __Engine_dot_hpp_
+
+#include "../features/compilerfeatures.h"
+#include "../array.h"
+#include <limits>
+#include <stdexcept>
+#include <sstream>
+#include <algorithm>
+#include <vector>
+#if R123_USE_CXX11_TYPE_TRAITS
+#include <type_traits>
+#endif
+
+namespace r123{
+/**
+ If G satisfies the requirements of a CBRNG, and has a ctr_type whose
+ value_type is an unsigned integral type, then Engine<G> satisfies
+ the requirements of a C++11 "Uniform Random Number Engine" and can
+ be used in any context where such an object is expected.
+
+ Note that wrapping a counter based RNG with a traditional API in
+ this way obscures much of the power of counter based PRNGs.
+ Nevertheless, it may be of value in applications that are already
+ coded to work with the C++11 random number engines.
+
+ The MicroURNG template in MicroURNG.hpp
+ provides the more limited functionality of a C++11 "Uniform
+ Random Number Generator", but leaves the application in control
+ of counters and keys and hence may be preferable to the Engine template.
+ For example, a MicroURNG allows one to use C++11 "Random Number
+ Distributions" without giving up control over the counters
+ and keys.
+*/
+
+template<typename CBRNG>
+struct Engine {
+ typedef CBRNG cbrng_type;
+ typedef typename CBRNG::ctr_type ctr_type;
+ typedef typename CBRNG::key_type key_type;
+ typedef typename CBRNG::ukey_type ukey_type;
+ typedef typename ctr_type::value_type result_type;
+
+protected:
+ cbrng_type b;
+ key_type key;
+ ctr_type c;
+ ctr_type v;
+
+ void fix_invariant(){
+ if( v.back() != 0 ) {
+ result_type vv = v.back();
+ v = b(c, key);
+ v.back() = vv;
+ }
+ }
+public:
+ explicit Engine() : b(), c() {
+ ukey_type x = {{}};
+ v.back() = 0;
+ key = x;
+ }
+ explicit Engine(result_type r) : b(), c() {
+ ukey_type x = {{typename ukey_type::value_type(r)}};
+ v.back() = 0;
+ key = x;
+ }
+ // 26.5.3 says that the SeedSeq templates shouldn't particpate in
+ // overload resolution unless the type qualifies as a SeedSeq.
+ // How that is determined is unspecified, except that "as a
+ // minimum a type shall not qualify as a SeedSeq if it is
+ // implicitly convertible to a result_type."
+ //
+ // First, we make sure that even the non-const copy constructor
+ // works as expected. In addition, if we've got C++11
+ // type_traits, we use enable_if and is_convertible to implement
+ // the convertible-to-result_type restriction. Otherwise, the
+ // template is unconditional and will match in some surpirsing
+ // and undesirable situations.
+ Engine(Engine& e) : b(e.b), key(e.key), c(e.c){
+ v.back() = e.v.back();
+ fix_invariant();
+ }
+ Engine(const Engine& e) : b(e.b), key(e.key), c(e.c){
+ v.back() = e.v.back();
+ fix_invariant();
+ }
+
+ template <typename SeedSeq>
+ explicit Engine(SeedSeq &s
+#if R123_USE_CXX11_TYPE_TRAITS
+ , typename std::enable_if<!std::is_convertible<SeedSeq, result_type>::value>::type* =0
+#endif
+ )
+ : b(), c() {
+ ukey_type ukey = ukey_type::seed(s);
+ key = ukey;
+ v.back() = 0;
+ }
+ void seed(result_type r){
+ *this = Engine(r);
+ }
+ template <typename SeedSeq>
+ void seed(SeedSeq &s
+#if R123_USE_CXX11_TYPE_TRAITS
+ , typename std::enable_if<!std::is_convertible<SeedSeq, result_type>::value>::type* =0
+#endif
+ ){
+ *this = Engine(s);
+ }
+ void seed(){
+ *this = Engine();
+ }
+ friend bool operator==(const Engine& lhs, const Engine& rhs){
+ return lhs.c==rhs.c && lhs.v.back() == rhs.v.back() && lhs.key == rhs.key;
+ }
+ friend bool operator!=(const Engine& lhs, const Engine& rhs){
+ return lhs.c!=rhs.c || lhs.v.back()!=rhs.v.back() || lhs.key!=rhs.key;
+ }
+
+ friend std::ostream& operator<<(std::ostream& os, const Engine& be){
+ return os << be.c << " " << be.key << " " << be.v.back();
+ }
+
+ friend std::istream& operator>>(std::istream& is, Engine& be){
+ is >> be.c >> be.key >> be.v.back();
+ be.fix_invariant();
+ return is;
+ }
+
+ // The <random> shipped with MacOS Xcode 4.5.2 imposes a
+ // non-standard requirement that URNGs also have static data
+ // members: _Min and _Max. Later versions of libc++ impose the
+ // requirement only when constexpr isn't supported. Although the
+ // Xcode 4.5.2 requirement is clearly non-standard, it is unlikely
+ // to be fixed and it is very easy work around. We certainly
+ // don't want to go to great lengths to accommodate every buggy
+ // library we come across, but in this particular case, the effort
+ // is low and the benefit is high, so it's worth doing. Thanks to
+ // Yan Zhou for pointing this out to us. See similar code in
+ // ../MicroURNG.hpp
+ const static result_type _Min = 0;
+ const static result_type _Max = ~((result_type)0);
+
+ static R123_CONSTEXPR result_type min R123_NO_MACRO_SUBST () { return _Min; }
+ static R123_CONSTEXPR result_type max R123_NO_MACRO_SUBST () { return _Max; }
+
+ result_type operator()(){
+ if( c.size() == 1 ) // short-circuit the scalar case. Compilers aren't mind-readers.
+ return b(c.incr(), key)[0];
+ result_type& elem = v.back();
+ if( elem == 0 ){
+ v = b(c.incr(), key);
+ result_type ret = v.back();
+ elem = c.size()-1;
+ return ret;
+ }
+ return v[--elem];
+ }
+
+ void discard(R123_ULONG_LONG skip){
+ // don't forget: elem counts down
+ size_t nelem = c.size();
+ size_t sub = skip % nelem;
+ result_type& elem = v.back();
+ skip /= nelem;
+ if (elem < sub) {
+ elem += nelem;
+ skip++;
+ }
+ elem -= sub;
+ c.incr(skip);
+ fix_invariant();
+ }
+
+ //--------------------------
+ // Some bonus methods, not required for a Random Number
+ // Engine
+
+ // Constructors and seed() method for ukey_type seem useful
+ // We need const and non-const to supersede the SeedSeq template.
+ explicit Engine(const ukey_type &uk) : key(uk), c(){ v.back() = 0; }
+ explicit Engine(ukey_type &uk) : key(uk), c(){ v.back() = 0; }
+ void seed(const ukey_type& uk){
+ *this = Engine(uk);
+ }
+ void seed(ukey_type& uk){
+ *this = Engine(uk);
+ }
+
+#if R123_USE_CXX11_TYPE_TRAITS
+ template <typename DUMMY=void>
+ explicit Engine(const key_type& k,
+ typename std::enable_if<!std::is_same<ukey_type, key_type>::value, DUMMY>::type* = 0)
+ : key(k), c(){ v.back() = 0; }
+
+ template <typename DUMMY=void>
+ void seed(const key_type& k,
+ typename std::enable_if<!std::is_same<ukey_type, key_type>::value, DUMMY>::type* = 0){
+ *this = Engine(k);
+ }
+#endif
+
+ // Forward the e(counter) to the CBRNG we are templated
+ // on, using the current value of the key.
+ ctr_type operator()(const ctr_type& c) const{
+ return b(c, key);
+ }
+
+ key_type getkey() const{
+ return key;
+ }
+
+ // N.B. setkey(k) is different from seed(k) because seed(k) zeros
+ // the counter (per the C++11 requirements for an Engine), whereas
+ // setkey does not.
+ void setkey(const key_type& k){
+ key = k;
+ fix_invariant();
+ }
+
+ // Maybe the caller want's to know the details of
+ // the internal state, e.g., so it can call a different
+ // bijection with the same counter.
+ std::pair<ctr_type, result_type> getcounter() const {
+ return std::make_pair(c, v.back());
+ }
+
+ // And the inverse.
+ void setcounter(const ctr_type& _c, result_type _elem){
+ static const size_t nelem = c.size();
+ if( _elem >= nelem )
+ throw std::range_error("Engine::setcounter called with elem out of range");
+ c = _c;
+ v.back() = _elem;
+ fix_invariant();
+ }
+
+ void setcounter(const std::pair<ctr_type, result_type>& ce){
+ setcounter(ce.first, ce.second);
+ }
+};
+} // namespace r123
+
+#endif
diff --git a/ext/random123/include/Random123/conventional/gsl_cbrng.h b/ext/random123/include/Random123/conventional/gsl_cbrng.h
new file mode 100644
index 0000000000000000000000000000000000000000..44457d002b0d40ec5319236f67bcd3ec2a3ce629
--- /dev/null
+++ b/ext/random123/include/Random123/conventional/gsl_cbrng.h
@@ -0,0 +1,128 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __r123_compat_gslrng_dot_h__
+#define __r123_compat_gslrng_dot_h__
+
+#include <gsl/gsl_rng.h>
+#include <string.h>
+
+/**
+ The macro: GSL_CBRNG(NAME, CBRNGNAME)
+ declares the necessary structs and constants that define a
+ gsl_rng_NAME type based on the counter-based RNG CBRNGNAME. For example:
+
+ Usage:
+
+ @code
+ #include <Random123/threefry.h>
+ #include <Random123/conventional/gsl_cbrng.h> // this file
+ GSL_CBRNG(cbrng, threefry4x32); // creates gsl_rng_cbrng
+
+ int main(int argc, char **argv){
+ gsl_rng *r = gsl_rng_alloc(gsl_rng_cbrng);
+ ... use r as you would use any other gsl_rng ...
+ }
+ @endcode
+
+ It requires that NAME be the name of a CBRNG that follows the
+ naming and stylistic conventions of the Random123 library.
+
+ Note that wrapping a \ref CBRNG "counter-based PRNG" with a traditional API in
+ this way obscures much of the power of the CBRNG API.
+ Nevertheless, it may be of value to applications that are already
+ coded to work with GSL random number generators, and that wish
+ to use the RNGs in the Random123 library.
+
+ */
+
+#define GSL_CBRNG(NAME, CBRNGNAME) \
+const gsl_rng_type *gsl_rng_##NAME; \
+ \
+typedef struct{ \
+ CBRNGNAME##_ctr_t ctr; \
+ CBRNGNAME##_ctr_t r; \
+ CBRNGNAME##_key_t key; \
+ int elem; \
+} NAME##_state; \
+ \
+static unsigned long int NAME##_get(void *vstate){ \
+ NAME##_state *st = (NAME##_state *)vstate; \
+ const int N=sizeof(st->ctr.v)/sizeof(st->ctr.v[0]); \
+ if( st->elem == 0 ){ \
+ ++st->ctr.v[0]; \
+ if( N>1 && st->ctr.v[0] == 0 ) ++st->ctr.v[1]; \
+ if( N>2 && st->ctr.v[1] == 0 ) ++st->ctr.v[2]; \
+ if( N>3 && st->ctr.v[2] == 0 ) ++st->ctr.v[3]; \
+ st->r = CBRNGNAME(st->ctr, st->key); \
+ st->elem = N; \
+ } \
+ return 0xffffffffUL & st->r.v[--st->elem]; \
+} \
+ \
+static double \
+NAME##_get_double (void * vstate) \
+{ \
+ return NAME##_get (vstate)/4294967296.0; \
+} \
+ \
+static void NAME##_set(void *vstate, unsigned long int s){ \
+ NAME##_state *st = (NAME##_state *)vstate; \
+ st->elem = 0; \
+ /* Assume that key and ctr have an array member, v, \
+ as if they are r123arrayNxW. If not, this will fail \
+ to compile. In particular, this macro fails to compile \
+ when the underlying CBRNG requires use of keyinit */ \
+ memset(&st->ctr.v[0], 0, sizeof(st->ctr.v)); \
+ memset(&st->key.v[0], 0, sizeof(st->key.v)); \
+ /* GSL 1.15 documentation says this about gsl_rng_set: \
+ Note that the most generators only accept 32-bit seeds, with higher \
+ values being reduced modulo 2^32. For generators with smaller \
+ ranges the maximum seed value will typically be lower. \
+ so we won't jump through any hoops here to deal with \
+ high bits if sizeof(unsigned long) > sizeof(uint32_t). */ \
+ st->key.v[0] = s; \
+} \
+ \
+static const gsl_rng_type NAME##_type = { \
+ #NAME, \
+ 0xffffffffUL, \
+ 0, \
+ sizeof(NAME##_state), \
+ &NAME##_set, \
+ &NAME##_get, \
+ &NAME##_get_double \
+}; \
+ \
+const gsl_rng_type *gsl_rng_##NAME = &NAME##_type
+
+#endif
+
diff --git a/ext/random123/include/Random123/features/clangfeatures.h b/ext/random123/include/Random123/features/clangfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..1e3c8cfdb9ccaf5e5c0491af373922e9855f1620
--- /dev/null
+++ b/ext/random123/include/Random123/features/clangfeatures.h
@@ -0,0 +1,93 @@
+/*
+Copyright 2010-2016, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __clangfeatures_dot_hpp
+#define __clangfeatures_dot_hpp
+
+#ifndef R123_USE_X86INTRIN_H
+#if (defined(__x86_64__)||defined(__i386__))
+#define R123_USE_X86INTRIN_H 1
+#else
+#define R123_USE_X86INTRIN_H 0
+#endif
+#endif
+
+#ifndef R123_USE_CXX11_UNRESTRICTED_UNIONS
+#define R123_USE_CXX11_UNRESTRICTED_UNIONS __has_feature(cxx_unrestricted_unions)
+#endif
+
+#ifndef R123_USE_CXX11_STATIC_ASSERT
+#define R123_USE_CXX11_STATIC_ASSERT __has_feature(cxx_static_assert)
+#endif
+
+// With clang-3.6, -Wall warns about unused-local-typedefs.
+// The "obvious" thing to do is to ignore -Wunused-local-typedefs,
+// but that doesn't work because earlier versions of clang blow
+// up on an 'unknown warning group'. So we briefly ignore -Wall...
+// It's tempting to just give up on static assertions in pre-c++11 code.
+#if !R123_USE_CXX11_STATIC_ASSERT && !defined(R123_STATIC_ASSERT)
+#define R123_STATIC_ASSERT(expr, msg) \
+_Pragma("clang diagnostic push") \
+_Pragma("clang diagnostic ignored \"-Wall\"") \
+typedef char static_assertion[(!!(expr))*2-1] \
+_Pragma("clang diagnostic pop")
+#endif
+
+#ifndef R123_USE_CXX11_CONSTEXPR
+#define R123_USE_CXX11_CONSTEXPR __has_feature(cxx_constexpr)
+#endif
+
+#ifndef R123_USE_CXX11_EXPLICIT_CONVERSIONS
+#define R123_USE_CXX11_EXPLICIT_CONVERSIONS __has_feature(cxx_explicit_conversions)
+#endif
+
+// With clang-3.0, the apparently simpler:
+// #define R123_USE_CXX11_RANDOM __has_include(<random>)
+// dumps core.
+#ifndef R123_USE_CXX11_RANDOM
+#if __cplusplus>=201103L && __has_include(<random>)
+#define R123_USE_CXX11_RANDOM 1
+#else
+#define R123_USE_CXX11_RANDOM 0
+#endif
+#endif
+
+#ifndef R123_USE_CXX11_TYPE_TRAITS
+#if __cplusplus>=201103L && __has_include(<type_traits>)
+#define R123_USE_CXX11_TYPE_TRAITS 1
+#else
+#define R123_USE_CXX11_TYPE_TRAITS 0
+#endif
+#endif
+
+#include "gccfeatures.h"
+
+#endif
diff --git a/ext/random123/include/Random123/features/compilerfeatures.h b/ext/random123/include/Random123/features/compilerfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..2341a7a01ef53fd6add4a381588fa6a4e84029b5
--- /dev/null
+++ b/ext/random123/include/Random123/features/compilerfeatures.h
@@ -0,0 +1,343 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+/**
+
+@page porting Preprocessor symbols for porting Random123 to different platforms.
+
+The Random123 library is portable across C, C++, CUDA, OpenCL environments,
+and multiple operating systems (Linux, Windows 7, Mac OS X, FreeBSD, Solaris).
+This level of portability requires the abstraction of some features
+and idioms that are either not standardized (e.g., asm statments), or for which
+different vendors have their own standards (e.g., SSE intrinsics) or for
+which vendors simply refuse to conform to well-established standards (e.g., <inttypes.h>).
+
+Random123/features/compilerfeatures.h
+conditionally includes a compiler-or-OS-specific Random123/featires/XXXfeatures.h file which
+defines appropriate values for the preprocessor symbols which can be used with
+a specific compiler or OS. Those symbols will then
+be used by other header files and source files in the Random123
+library (and may be used by applications) to control what actually
+gets presented to the compiler.
+
+Most of the symbols are boolean valued. In general, they will
+\b always be defined with value either 1 or 0, so do
+\b NOT use \#ifdef. Use \#if R123_USE_SOMETHING instead.
+
+Library users can override any value by defining the pp-symbol with a compiler option,
+e.g.,
+
+ cc -DR123_USE_MULHILO64_C99
+
+will use a strictly c99 version of the full-width 64x64->128-bit multiplication
+function, even if it would be disabled by default.
+
+All boolean-valued pre-processor symbols in Random123/features/compilerfeatures.h start with the prefix R123_USE_
+@verbatim
+ AES_NI
+ AES_OPENSSL
+ SSE4_2
+ SSE4_1
+ SSE
+
+ STD_RANDOM
+
+ GNU_UINT128
+ ASM_GNU
+ ASM_MSASM
+
+ CPUID_MSVC
+
+ CXX11_RANDOM
+ CXX11_TYPE_TRAITS
+ CXX11_STATIC_ASSERT
+ CXX11_CONSTEXPR
+ CXX11_UNRESTRICTED_UNIONS
+ CXX11_EXPLICIT_CONVERSIONS
+ CXX11_LONG_LONG
+ CXX11_STD_ARRAY
+ CXX11
+
+ X86INTRIN_H
+ IA32INTRIN_H
+ XMMINTRIN_H
+ EMMINTRIN_H
+ SMMINTRIN_H
+ WMMINTRIN_H
+ INTRIN_H
+
+ MULHILO32_ASM
+ MULHILO64_ASM
+ MULHILO64_MSVC_INTRIN
+ MULHILO64_CUDA_INTRIN
+ MULHILO64_OPENCL_INTRIN
+ MULHILO64_C99
+
+ U01_DOUBLE
+
+@endverbatim
+Most have obvious meanings. Some non-obvious ones:
+
+AES_NI and AES_OPENSSL are not mutually exclusive. You can have one,
+both or neither.
+
+GNU_UINT128 says that it's safe to use __uint128_t, but it
+does not require its use. In particular, it should be
+used in mulhilo<uint64_t> only if MULHILO64_ASM is unset.
+
+If the XXXINTRIN_H macros are true, then one should
+@code
+#include <xxxintrin.h>
+@endcode
+to gain accesss to compiler intrinsics.
+
+The CXX11_SOME_FEATURE macros allow the code to use specific
+features of the C++11 language and library. The catchall
+In the absence of a specific CXX11_SOME_FEATURE, the feature
+is controlled by the catch-all R123_USE_CXX11 macro.
+
+U01_DOUBLE defaults on, and can be turned off (set to 0)
+if one does not want the utility functions that convert to double
+(i.e. u01_*_53()), e.g. on OpenCL without the cl_khr_fp64 extension.
+
+There are a number of invariants that are always true. Application code may
+choose to rely on these:
+
+<ul>
+<li>ASM_GNU and ASM_MASM are mutually exclusive
+<li>The "higher" SSE values imply the lower ones.
+</ul>
+
+There are also non-boolean valued symbols:
+
+<ul>
+<li>R123_STATIC_INLINE -
+ According to both C99 and GNU99, the 'static inline' declaration allows
+ the compiler to not emit code if the function is not used.
+ Note that the semantics of 'inline', 'static' and 'extern' in
+ gcc have changed over time and are subject to modification by
+ command line options, e.g., -std=gnu89, -fgnu-inline.
+ Nevertheless, it appears that the meaning of 'static inline'
+ has not changed over time and (with a little luck) the use of 'static inline'
+ here will be portable between versions of gcc and to other C99
+ compilers.
+ See: http://gcc.gnu.org/onlinedocs/gcc/Inline.html
+ http://www.greenend.org.uk/rjk/2003/03/inline.html
+
+<li>R123_FORCE_INLINE(decl) -
+ which expands to 'decl', adorned with the compiler-specific
+ embellishments to strongly encourage that the declared function be
+ inlined. If there is no such compiler-specific magic, it should
+ expand to decl, unadorned.
+
+<li>R123_CUDA_DEVICE - which expands to __device__ (or something else with
+ sufficiently similar semantics) when CUDA is in use, and expands
+ to nothing in other cases.
+
+<li>R123_METAL_THREAD_ADDRESS_SPACE - which expands to 'thread' (or
+ something else with sufficiently similar semantics) when compiling a
+ Metal kernel, and expands to nothing in other cases.
+
+<li>R123_ASSERT(x) - which expands to assert(x), or maybe to nothing at
+ all if we're in an environment so feature-poor that you can't even
+ call assert (I'm looking at you, CUDA and OpenCL), or even include
+ assert.h safely (OpenCL).
+
+<li>R123_STATIC_ASSERT(expr,msg) - which expands to
+ static_assert(expr,msg), or to an expression that
+ will raise a compile-time exception if expr is not true.
+
+<li>R123_ULONG_LONG - which expands to a declaration of the longest available
+ unsigned integer.
+
+<li>R123_64BIT(x) - expands to something equivalent to
+ UINT64_C(x) from <stdint.h>, even in environments where <stdint.h>
+ is not available, e.g., MSVC and OpenCL.
+
+<li>R123_BUILTIN_EXPECT(expr,likely_value) - expands to something with
+ the semantics of gcc's __builtin_expect(expr,likely_value). If
+ the environment has nothing like __builtin_expect, it should expand
+ to just expr.
+</ul>
+
+
+\cond HIDDEN_FROM_DOXYGEN
+*/
+
+/*
+N.B. When something is added to the list of features, it should be
+added to each of the *features.h files, AND to examples/ut_features.cpp.
+*/
+
+/* N.B. most other compilers (icc, nvcc, open64, llvm) will also define __GNUC__, so order matters. */
+#if defined(__METAL_MACOS__)
+#include "metalfeatures.h"
+#elif defined(__OPENCL_VERSION__) && __OPENCL_VERSION__ > 0
+#include "openclfeatures.h"
+#elif defined(__CUDACC__)
+#include "nvccfeatures.h"
+#elif defined(__ICC)
+#include "iccfeatures.h"
+#elif defined(__xlC__)
+#include "xlcfeatures.h"
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+#include "sunprofeatures.h"
+#elif defined(__OPEN64__)
+#include "open64features.h"
+#elif defined(__clang__)
+#include "clangfeatures.h"
+#elif defined(__GNUC__)
+#include "gccfeatures.h"
+#elif defined(__PGI)
+#include "pgccfeatures.h"
+#elif defined(_MSC_FULL_VER)
+#include "msvcfeatures.h"
+#else
+#error "Can't identify compiler. You'll need to add a new xxfeatures.hpp"
+{ /* maybe an unbalanced brace will terminate the compilation */
+#endif
+
+#ifndef R123_USE_CXX11
+#define R123_USE_CXX11 (__cplusplus >= 201103L)
+#endif
+
+#ifndef R123_USE_CXX11_UNRESTRICTED_UNIONS
+#define R123_USE_CXX11_UNRESTRICTED_UNIONS R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_STATIC_ASSERT
+#define R123_USE_CXX11_STATIC_ASSERT R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_CONSTEXPR
+#define R123_USE_CXX11_CONSTEXPR R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_EXPLICIT_CONVERSIONS
+#define R123_USE_CXX11_EXPLICIT_CONVERSIONS R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_RANDOM
+#define R123_USE_CXX11_RANDOM R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_TYPE_TRAITS
+#define R123_USE_CXX11_TYPE_TRAITS R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_LONG_LONG
+#define R123_USE_CXX11_LONG_LONG R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_CXX11_STD_ARRAY
+#define R123_USE_CXX11_STD_ARRAY R123_USE_CXX11
+#endif
+
+#ifndef R123_USE_MULHILO64_C99
+#define R123_USE_MULHILO64_C99 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MULHI_INTRIN
+#define R123_USE_MULHILO64_MULHI_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO32_MULHI_INTRIN
+#define R123_USE_MULHILO32_MULHI_INTRIN 0
+#endif
+
+#ifndef R123_STATIC_ASSERT
+#if R123_USE_CXX11_STATIC_ASSERT
+#define R123_STATIC_ASSERT(expr, msg) static_assert(expr, msg)
+#else
+ /* if msg always_looked_like_this, we could paste it into the name. Worth it? */
+#define R123_STATIC_ASSERT(expr, msg) typedef char static_assertion[(!!(expr))*2-1]
+#endif
+#endif
+
+#ifndef R123_CONSTEXPR
+#if R123_USE_CXX11_CONSTEXPR
+#define R123_CONSTEXPR constexpr
+#else
+#define R123_CONSTEXPR
+#endif
+#endif
+
+#ifndef R123_USE_64BIT
+#define R123_USE_64BIT 1
+#endif
+
+#ifndef R123_USE_PHILOX_64BIT
+#define R123_USE_PHILOX_64BIT (R123_USE_64BIT && (R123_USE_MULHILO64_ASM || R123_USE_MULHILO64_MSVC_INTRIN || R123_USE_MULHILO64_CUDA_INTRIN || R123_USE_GNU_UINT128 || R123_USE_MULHILO64_C99 || R123_USE_MULHILO64_OPENCL_INTRIN || R123_USE_MULHILO64_MULHI_INTRIN))
+#endif
+
+#ifndef R123_ULONG_LONG
+#if defined(__cplusplus) && !R123_USE_CXX11_LONG_LONG
+/* C++98 doesn't have long long. It doesn't have uint64_t either, but
+ we will have typedef'ed uint64_t to something in the xxxfeatures.h.
+ With luck, it won't elicit complaints from -pedantic. Cross your
+ fingers... */
+#define R123_ULONG_LONG uint64_t
+#else
+#define R123_ULONG_LONG unsigned long long
+#endif
+#endif
+
+/* UINT64_C should have been #defined by XXXfeatures.h, either by
+ #include <stdint.h> or through compiler-dependent hacks */
+#ifndef R123_64BIT
+#define R123_64BIT(x) UINT64_C(x)
+#endif
+
+#ifndef R123_THROW
+#define R123_THROW(x) throw (x)
+#endif
+
+#ifndef R123_METAL_THREAD_ADDRESS_SPACE
+#define R123_METAL_THREAD_ADDRESS_SPACE
+#endif
+
+#ifndef R123_METAL_CONSTANT_ADDRESS_SPACE
+#define R123_METAL_CONSTANT_ADDRESS_SPACE
+#endif
+
+/*
+ * Windows.h (and perhaps other "well-meaning" code define min and
+ * max, so there's a high chance that our definition of min, max
+ * methods or use of std::numeric_limits min and max will cause
+ * complaints in any program that happened to include Windows.h or
+ * suchlike first. We use the null macro below in our own header
+ * files definition or use of min, max to defensively preclude
+ * this problem. It may not be enough; one might need to #define
+ * NOMINMAX before including Windows.h or compile with -DNOMINMAX.
+ */
+#define R123_NO_MACRO_SUBST
+
+/** \endcond */
diff --git a/ext/random123/include/Random123/features/gccfeatures.h b/ext/random123/include/Random123/features/gccfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..701f3c667801421d7638d6cf7474f5bc461f8411
--- /dev/null
+++ b/ext/random123/include/Random123/features/gccfeatures.h
@@ -0,0 +1,263 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __gccfeatures_dot_hpp
+#define __gccfeatures_dot_hpp
+
+#define R123_GNUC_VERSION (__GNUC__*10000 + __GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
+
+#if !defined(__x86_64__) && !defined(__i386__) && !defined(__powerpc__) && !defined(__arm__) && !defined(__aarch64__)
+# error "This code has only been tested on x86, powerpc and a few arm platforms."
+#include <including_a_nonexistent_file_will_stop_some_compilers_from_continuing_with_a_hopeless_task>
+{ /* maybe an unbalanced brace will terminate the compilation */
+ /* Feel free to try the Random123 library on other architectures by changing
+ the conditions that reach this error, but you should consider it a
+ porting exercise and expect to encounter bugs and deficiencies.
+ Please let the authors know of any successes (or failures). */
+#endif
+
+#ifdef __powerpc__
+#include <ppu_intrinsics.h>
+#endif
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static __inline__
+#endif
+
+#ifndef R123_FORCE_INLINE
+#if R123_GNUC_VERSION >= 40000
+#define R123_FORCE_INLINE(decl) decl __attribute__((always_inline))
+#else
+#define R123_FORCE_INLINE(decl) decl
+#endif
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) __builtin_expect(expr,likely)
+#endif
+
+/* According to the C++0x standard, we should be able to test the numeric
+ value of __cplusplus == 199701L for C++98, __cplusplus == 201103L for C++11
+ But gcc has had an open bug http://gcc.gnu.org/bugzilla/show_bug.cgi?id=1773
+ since early 2001, which was finally fixed in 4.7 (early 2012). For
+ earlier versions, the only way to detect whether --std=c++0x was requested
+ on the command line is to look at the __GCC_EXPERIMENTAL_CXX0X__ pp-symbol.
+*/
+#if defined(__GCC_EXPERIMENTAL_CXX0X__)
+#define GNU_CXX11 (__cplusplus>=201103L || (R123_GNUC_VERSION<40700 && 1/* defined(__GCC_EXPERIMENTAL_CXX0X__) */))
+#else
+#define GNU_CXX11 (__cplusplus>=201103L || (R123_GNUC_VERSION<40700 && 0/* defined(__GCC_EXPERIMENTAL_CXX0X__) */))
+#endif
+
+#ifndef R123_USE_CXX11_UNRESTRICTED_UNIONS
+#define R123_USE_CXX11_UNRESTRICTED_UNIONS ((R123_GNUC_VERSION >= 40600) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_CXX11_STATIC_ASSERT
+#define R123_USE_CXX11_STATIC_ASSERT ((R123_GNUC_VERSION >= 40300) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_CXX11_CONSTEXPR
+#define R123_USE_CXX11_CONSTEXPR ((R123_GNUC_VERSION >= 40600) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_CXX11_EXPLICIT_CONVERSIONS
+#define R123_USE_CXX11_EXPLICIT_CONVERSIONS ((R123_GNUC_VERSION >= 40500) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_CXX11_RANDOM
+#define R123_USE_CXX11_RANDOM ((R123_GNUC_VERSION>=40500) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_CXX11_TYPE_TRAITS
+#define R123_USE_CXX11_TYPE_TRAITS ((R123_GNUC_VERSION>=40400) && GNU_CXX11)
+#endif
+
+#ifndef R123_USE_AES_NI
+#ifdef __AES__
+#define R123_USE_AES_NI 1
+#else
+#define R123_USE_AES_NI 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE4_2
+#ifdef __SSE4_2__
+#define R123_USE_SSE4_2 1
+#else
+#define R123_USE_SSE4_2 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE4_1
+#ifdef __SSE4_1__
+#define R123_USE_SSE4_1 1
+#else
+#define R123_USE_SSE4_1 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE
+/* There's no point in trying to compile SSE code in Random123
+ unless SSE2 is available. */
+#ifdef __SSE2__
+#define R123_USE_SSE 1
+#else
+#define R123_USE_SSE 0
+#endif
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+/* There isn't really a good way to tell at compile time whether
+ openssl is available. Without a pre-compilation configure-like
+ tool, it's less error-prone to guess that it isn't available. Add
+ -DR123_USE_AES_OPENSSL=1 and any necessary LDFLAGS or LDLIBS to
+ play with openssl */
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#if defined(__x86_64__) || defined(__aarch64__)
+#define R123_USE_GNU_UINT128 1
+#else
+#define R123_USE_GNU_UINT128 0
+#endif
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#if (defined(__x86_64__)||defined(__i386__))
+#define R123_USE_ASM_GNU 1
+#else
+#define R123_USE_ASM_GNU 1
+#endif
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 0
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#if (defined(__x86_64__)||defined(__i386__))
+#define R123_USE_X86INTRIN_H (1/* (defined(__x86_64__)||defined(__i386__)) */ && R123_GNUC_VERSION >= 40402)
+#else
+#define R123_USE_X86INTRIN_H (0/* (defined(__x86_64__)||defined(__i386__)) */ && R123_GNUC_VERSION >= 40402)
+#endif
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 0
+#endif
+
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+/* gcc -m64 on Solaris 10 defines __SSE2__ but doesn't have
+ emmintrin.h in the include search path. This is
+ so broken that I refuse to try to work around it. If this
+ affects you, figure out where your emmintrin.h lives and
+ add an appropriate -I to your CPPFLAGS. Or add -DR123_USE_SSE=0. */
+#define R123_USE_EMMINTRIN_H (R123_USE_SSE && (R123_GNUC_VERSION < 40402))
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H ((R123_USE_SSE4_1 || R123_USE_SSE4_2) && (R123_GNUC_VERSION < 40402))
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#define R123_USE_INTRIN_H 0
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MULHI_INTRIN
+#if (defined(__powerpc64__))
+#define R123_USE_MULHILO64_MULHI_INTRIN 1
+#else
+#define R123_USE_MULHILO64_MULHI_INTRIN 0
+#endif
+#endif
+
+#ifndef R123_MULHILO64_MULHI_INTRIN
+#define R123_MULHILO64_MULHI_INTRIN __mulhdu
+#endif
+
+#ifndef R123_USE_MULHILO32_MULHI_INTRIN
+#define R123_USE_MULHILO32_MULHI_INTRIN 0
+#endif
+
+#ifndef R123_MULHILO32_MULHI_INTRIN
+#define R123_MULHILO32_MULHI_INTRIN __mulhwu
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+/* If you add something, it must go in all the other XXfeatures.hpp
+ and in ../ut_features.cpp */
+#endif
diff --git a/ext/random123/include/Random123/features/iccfeatures.h b/ext/random123/include/Random123/features/iccfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..7e72dec1d31027eac66e00fe92b4cd926b235d23
--- /dev/null
+++ b/ext/random123/include/Random123/features/iccfeatures.h
@@ -0,0 +1,212 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __icpcfeatures_dot_hpp
+#define __icpcfeatures_dot_hpp
+
+// icc relies on gcc libraries and other toolchain components.
+#define R123_GNUC_VERSION (__GNUC__*10000 + __GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
+
+#if !defined(__x86_64__) && !defined(__i386__)
+# error "This code has only been tested on x86 platforms."
+{ // maybe an unbalanced brace will terminate the compilation
+// You are invited to try Easy123 on other architectures, by changing
+// the conditions that reach this error, but you should consider it a
+// porting exercise and expect to encounter bugs and deficiencies.
+// Please let the authors know of any successes (or failures).
+#endif
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl __attribute__((always_inline))
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) __builtin_expect(expr,likely)
+#endif
+
+// The basic idiom is:
+// #ifndef R123_SOMETHING
+// #if some condition
+// #define R123_SOMETHING 1
+// #else
+// #define R123_SOMETHING 0
+// #endif
+// #endif
+// This idiom allows an external user to override any decision
+// in this file with a command-line -DR123_SOMETHING=1 or -DR123_SOMETHINE=0
+
+// An alternative idiom is:
+// #ifndef R123_SOMETHING
+// #define R123_SOMETHING (some boolean expression)
+// #endif
+// where the boolean expression might contain previously-defined R123_SOMETHING_ELSE
+// pp-symbols.
+
+#ifndef R123_USE_SSE4_2
+#ifdef __SSE4_2__
+#define R123_USE_SSE4_2 1
+#else
+#define R123_USE_SSE4_2 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE4_1
+#ifdef __SSE4_1__
+#define R123_USE_SSE4_1 1
+#else
+#define R123_USE_SSE4_1 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE
+#ifdef __SSE2__
+#define R123_USE_SSE 1
+#else
+#define R123_USE_SSE 0
+#endif
+#endif
+
+#ifndef R123_USE_AES_NI
+// Unlike gcc, icc (version 12) does not pre-define an __AES__
+// pp-symbol when -maes or -xHost is on the command line. This feels
+// like a defect in icc (it defines __SSE4_2__ in analogous
+// circumstances), but until Intel fixes it, we're better off erring
+// on the side of caution and not generating instructions that are
+// going to raise SIGILL when executed. To get the AES-NI
+// instructions with icc, the caller must puts something like
+// -DR123_USE_AES_NI=1 or -D__AES__ on the command line. FWIW, the
+// AES-NI Whitepaper by Gueron says that icc has supported AES-NI from
+// 11.1 onwards.
+//
+#if defined(__AES__)
+#define R123_USE_AES_NI ((__ICC>=1101) && 1/*defined(__AES__)*/)
+#else
+#define R123_USE_AES_NI ((__ICC>=1101) && 0/*defined(__AES__)*/)
+#endif
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+/* There isn't really a good way to tell at compile time whether
+ openssl is available. Without a pre-compilation configure-like
+ tool, it's less error-prone to guess that it isn't available. Add
+ -DR123_USE_AES_OPENSSL=1 and any necessary LDFLAGS or LDLIBS to
+ play with openssl */
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#define R123_USE_ASM_GNU 1
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 0
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#define R123_USE_X86INTRIN_H 0
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 1
+#endif
+
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+#define R123_USE_EMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#define R123_USE_INTRIN_H 0
+#endif
+
+#ifndef R123_USE_MULHILO16_ASM
+#define R123_USE_MULHILO16_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 1
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+// If you add something, it must go in all the other XXfeatures.hpp
+// and in ../ut_features.cpp
+#endif
diff --git a/ext/random123/include/Random123/features/metalfeatures.h b/ext/random123/include/Random123/features/metalfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..bafe51a6840cb3ae5ca7fc1145cf5442db0657e5
--- /dev/null
+++ b/ext/random123/include/Random123/features/metalfeatures.h
@@ -0,0 +1,111 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/*
+ * Written by Tom Schoonjans <Tom.Schoonjans@me.com>
+ */
+
+#ifndef __metalfeatures_dot_hpp
+#define __metalfeatures_dot_hpp
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl __attribute__((always_inline))
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_METAL_THREAD_ADDRESS_SPACE
+#define R123_METAL_THREAD_ADDRESS_SPACE thread
+#endif
+
+#ifndef R123_METAL_CONSTANT_ADDRESS_SPACE
+#define R123_METAL_CONSTANT_ADDRESS_SPACE constant
+#endif
+
+#ifndef R123_ASSERT
+#define R123_ASSERT(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) expr
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO32_MULHI_INTRIN
+#define R123_USE_MULHILO32_MULHI_INTRIN 1
+#endif
+
+#if R123_USE_MULHILO32_MULHI_INTRIN
+#include <metal_integer>
+#define R123_MULHILO32_MULHI_INTRIN metal::mulhi
+#endif
+
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+#ifndef R123_USE_64BIT
+#define R123_USE_64BIT 0 /* Metal currently (Feb 2019, Specification-2) does not support 64-bit variable types */
+#endif
+
+#ifndef R123_ULONG_LONG
+/* the longest integer type in Metal (Feb 2019, Specification-2) is a
+ * 32-bit unsigned int. Let's hope for the best... */
+#define R123_ULONG_LONG unsigned int
+#endif
+
+#endif
diff --git a/ext/random123/include/Random123/features/msvcfeatures.h b/ext/random123/include/Random123/features/msvcfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..9eb9520912daf66869a6cf9fd027c37f06a8a3d4
--- /dev/null
+++ b/ext/random123/include/Random123/features/msvcfeatures.h
@@ -0,0 +1,200 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __msvcfeatures_dot_hpp
+#define __msvcfeatures_dot_hpp
+
+//#if _MSVC_FULL_VER <= 15
+//#error "We've only tested MSVC_FULL_VER==15."
+//#endif
+
+#if !defined(_M_IX86) && !defined(_M_X64)
+# error "This code has only been tested on x86 platforms."
+{ // maybe an unbalanced brace will terminate the compilation
+// You are invited to try Random123 on other architectures, by changing
+// the conditions that reach this error, but you should consider it a
+// porting exercise and expect to encounter bugs and deficiencies.
+// Please let the authors know of any successes (or failures).
+#endif
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static __inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) _forceinline decl
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) expr
+#endif
+
+// The basic idiom is:
+// #ifndef R123_SOMETHING
+// #if some condition
+// #define R123_SOMETHING 1
+// #else
+// #define R123_SOMETHING 0
+// #endif
+// #endif
+// This idiom allows an external user to override any decision
+// in this file with a command-line -DR123_SOMETHING=1 or -DR123_SOMETHINE=0
+
+// An alternative idiom is:
+// #ifndef R123_SOMETHING
+// #define R123_SOMETHING (some boolean expression)
+// #endif
+// where the boolean expression might contain previously-defined R123_SOMETHING_ELSE
+// pp-symbols.
+
+#ifndef R123_USE_AES_NI
+#if defined(_M_X64)
+#define R123_USE_AES_NI 1
+#else
+#define R123_USE_AES_NI 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE4_2
+#if defined(_M_X64)
+#define R123_USE_SSE4_2 1
+#else
+#define R123_USE_SSE4_2 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE4_1
+#if defined(_M_X64)
+#define R123_USE_SSE4_1 1
+#else
+#define R123_USE_SSE4_1 0
+#endif
+#endif
+
+#ifndef R123_USE_SSE
+#define R123_USE_SSE 1
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#define R123_USE_ASM_GNU 0
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 1
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#define R123_USE_X86INTRIN_H 0
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 0
+#endif
+
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+#define R123_USE_EMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#define R123_USE_INTRIN_H 1
+#endif
+
+#ifndef R123_USE_MULHILO16_ASM
+#define R123_USE_MULHILO16_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#if defined(_M_X64)
+#define R123_USE_MULHILO64_MSVC_INTRIN 1
+#else
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+#pragma warning(disable:4244)
+#pragma warning(disable:4996)
+
+// If you add something, it must go in all the other XXfeatures.hpp
+// and in ../ut_features.cpp
+#endif
diff --git a/ext/random123/include/Random123/features/nvccfeatures.h b/ext/random123/include/Random123/features/nvccfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..d1ff8bf521a05d45232e922c9a875439bc84f837
--- /dev/null
+++ b/ext/random123/include/Random123/features/nvccfeatures.h
@@ -0,0 +1,125 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __r123_nvcc_features_dot_h__
+#define __r123_nvcc_features_dot_h__
+
+#if !defined(CUDART_VERSION)
+#error "why are we in nvccfeatures.h if CUDART_VERSION is not defined"
+#endif
+
+#if CUDART_VERSION < 4010
+#error "CUDA versions earlier than 4.1 produce incorrect results for some templated functions in namespaces. Random123 isunsupported. See comments in nvccfeatures.h"
+// This test was added in Random123-1.08 (August, 2013) because we
+// discovered that Ftype(maxTvalue<T>()) with Ftype=double and
+// T=uint64_t in examples/uniform.hpp produces -1 for CUDA4.0 and
+// earlier. We can't be sure this bug doesn't also affect invocations
+// of other templated functions, e.g., essentially all of Random123.
+// Thus, we no longer trust CUDA versions earlier than 4.1 even though
+// we had previously tested and timed Random123 with CUDA 3.x and 4.0.
+// If you feel lucky or desperate, you can change #error to #warning, but
+// please take extra care to be sure that you are getting correct
+// results.
+#endif
+
+// nvcc falls through to gcc or msvc. So first define
+// a couple of things and then include either gccfeatures.h
+// or msvcfeatures.h
+
+//#ifdef __CUDA_ARCH__ allows Philox32 and Philox64 to be compiled
+//for both device and host functions in CUDA by setting compiler flags
+//for the device function
+#ifdef __CUDA_ARCH__
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE __device__
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 1
+#endif
+
+#ifndef R123_THROW
+// No exceptions in CUDA, at least upto 4.0
+#define R123_THROW(x) R123_ASSERT(0)
+#endif
+
+#ifndef R123_ASSERT
+#define R123_ASSERT(x) if((x)) ; else asm("trap;")
+#endif
+
+#else // ! __CUDA_ARCH__
+// If we're using nvcc not compiling for the CUDA architecture,
+// then we must be compiling for the host. In that case,
+// tell the philox code to use the mulhilo64 asm because
+// nvcc doesn't grok uint128_t.
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 1
+#endif
+
+#endif // __CUDA_ARCH__
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) expr
+#endif
+
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+#ifndef R123_USE_SSE4_2
+#define R123_USE_SSE4_2 0
+#endif
+
+#ifndef R123_USE_SSE4_1
+#define R123_USE_SSE4_1 0
+#endif
+
+#ifndef R123_USE_SSE
+#define R123_USE_SSE 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_ULONG_LONG
+// uint64_t, which is what we'd get without this, is
+// not the same as unsigned long long
+#define R123_ULONG_LONG unsigned long long
+#endif
+
+#if defined(__GNUC__)
+#include "gccfeatures.h"
+#elif defined(_MSC_FULL_VER)
+#include "msvcfeatures.h"
+#endif
+
+#endif
diff --git a/ext/random123/include/Random123/features/open64features.h b/ext/random123/include/Random123/features/open64features.h
new file mode 100644
index 0000000000000000000000000000000000000000..8da9f5f51efab021c644b632b4499f12fa0220d9
--- /dev/null
+++ b/ext/random123/include/Random123/features/open64features.h
@@ -0,0 +1,50 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __open64features_dot_hpp
+#define __open64features_dot_hpp
+
+/* The gcc features are mostly right. We just override a few and then include gccfeatures.h */
+
+/* Open64 4.2.3 and 4.2.4 accept the __uint128_t code without complaint
+ but produce incorrect code for 64-bit philox. The MULHILO64_ASM
+ seems to work fine */
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 1
+#endif
+
+#include "gccfeatures.h"
+
+#endif
diff --git a/ext/random123/include/Random123/features/openclfeatures.h b/ext/random123/include/Random123/features/openclfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..af03d3092318c6c27f1a65ce8104c1609b1e66e1
--- /dev/null
+++ b/ext/random123/include/Random123/features/openclfeatures.h
@@ -0,0 +1,89 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __openclfeatures_dot_hpp
+#define __openclfeatures_dot_hpp
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl __attribute__((always_inline))
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#define R123_ASSERT(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) expr
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 1
+#endif
+
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+// XXX ATI APP SDK 2.4 clBuildProgram SEGVs if one uses uint64_t instead of
+// ulong to mul_hi. And gets lots of complaints from stdint.h
+// on some machines.
+// But these typedefs mean we cannot include stdint.h with
+// these headers? Do we need R123_64T, R123_32T, R123_8T?
+typedef ulong uint64_t;
+typedef uint uint32_t;
+typedef uchar uint8_t;
+#define UINT64_C(x) ((ulong)(x##UL))
+
+#endif
diff --git a/ext/random123/include/Random123/features/pgccfeatures.h b/ext/random123/include/Random123/features/pgccfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..18ace1353b4e0e6201c823e17b5325c2a9b05afe
--- /dev/null
+++ b/ext/random123/include/Random123/features/pgccfeatures.h
@@ -0,0 +1,194 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+Copyright (c) 2013, Los Alamos National Security, LLC
+All rights reserved.
+
+Copyright 2013. Los Alamos National Security, LLC. This software was produced
+under U.S. Government contract DE-AC52-06NA25396 for Los Alamos National
+Laboratory (LANL), which is operated by Los Alamos National Security, LLC for
+the U.S. Department of Energy. The U.S. Government has rights to use,
+reproduce, and distribute this software. NEITHER THE GOVERNMENT NOR LOS
+ALAMOS NATIONAL SECURITY, LLC MAKES ANY WARRANTY, EXPRESS OR IMPLIED, OR
+ASSUMES ANY LIABILITY FOR THE USE OF THIS SOFTWARE. If software is modified
+to produce derivative works, such modified software should be clearly marked,
+so as not to confuse it with the version available from LANL.
+*/
+#ifndef __pgccfeatures_dot_hpp
+#define __pgccfeatures_dot_hpp
+
+#if !defined(__x86_64__) && !defined(__i386__)
+# error "This code has only been tested on x86 platforms."
+#include <including_a_nonexistent_file_will_stop_some_compilers_from_continuing_with_a_hopeless_task>
+{ /* maybe an unbalanced brace will terminate the compilation */
+ /* Feel free to try the Random123 library on other architectures by changing
+ the conditions that reach this error, but you should consider it a
+ porting exercise and expect to encounter bugs and deficiencies.
+ Please let the authors know of any successes (or failures). */
+#endif
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static inline
+#endif
+
+/* Found this example in PGI's emmintrin.h. */
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl __attribute__((__always_inline__))
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) (expr)
+#endif
+
+/* PGI through 13.2 doesn't appear to support AES-NI. */
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+/* PGI through 13.2 appears to support MMX, SSE, SSE3, SSE3, SSSE3, SSE4a, and
+ ABM, but not SSE4.1 or SSE4.2. */
+#ifndef R123_USE_SSE4_2
+#define R123_USE_SSE4_2 0
+#endif
+
+#ifndef R123_USE_SSE4_1
+#define R123_USE_SSE4_1 0
+#endif
+
+#ifndef R123_USE_SSE
+/* There's no point in trying to compile SSE code in Random123
+ unless SSE2 is available. */
+#ifdef __SSE2__
+#define R123_USE_SSE 1
+#else
+#define R123_USE_SSE 0
+#endif
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+/* There isn't really a good way to tell at compile time whether
+ openssl is available. Without a pre-compilation configure-like
+ tool, it's less error-prone to guess that it isn't available. Add
+ -DR123_USE_AES_OPENSSL=1 and any necessary LDFLAGS or LDLIBS to
+ play with openssl */
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#define R123_USE_ASM_GNU 1
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 0
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#define R123_USE_X86INTRIN_H 0
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 0
+#endif
+
+/* emmintrin.h from PGI #includes xmmintrin.h but then complains at link time
+ about undefined references to _mm_castsi128_ps(__m128i). Why? */
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+#define R123_USE_EMMINTRIN_H 1
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#ifdef __ABM__
+#define R123_USE_INTRIN_H 1
+#else
+#define R123_USE_INTRIN_H 0
+#endif
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MULHI_INTRIN
+#define R123_USE_MULHILO64_MULHI_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 1
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+/* If you add something, it must go in all the other XXfeatures.hpp
+ and in ../ut_features.cpp */
+#endif
diff --git a/ext/random123/include/Random123/features/sse.h b/ext/random123/include/Random123/features/sse.h
new file mode 100644
index 0000000000000000000000000000000000000000..3a49ebd8652e0d8deb50ca0daac50f56818ca6e5
--- /dev/null
+++ b/ext/random123/include/Random123/features/sse.h
@@ -0,0 +1,280 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef _Random123_sse_dot_h__
+#define _Random123_sse_dot_h__
+
+#if R123_USE_SSE
+
+#if R123_USE_X86INTRIN_H
+#include <x86intrin.h>
+#endif
+#if R123_USE_IA32INTRIN_H
+#include <ia32intrin.h>
+#endif
+#if R123_USE_XMMINTRIN_H
+#include <xmmintrin.h>
+#endif
+#if R123_USE_EMMINTRIN_H
+#include <emmintrin.h>
+#endif
+#if R123_USE_SMMINTRIN_H
+#include <smmintrin.h>
+#endif
+#if R123_USE_WMMINTRIN_H
+#include <wmmintrin.h>
+#endif
+#if R123_USE_INTRIN_H
+#include <intrin.h>
+#endif
+#ifdef __cplusplus
+#include <iostream>
+#include <limits>
+#include <stdexcept>
+#endif
+
+#if R123_USE_ASM_GNU
+
+/* bit25 of CX tells us whether AES is enabled. */
+R123_STATIC_INLINE int haveAESNI(){
+ unsigned int eax, ebx, ecx, edx;
+ __asm__ __volatile__ ("cpuid": "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) :
+ "a" (1));
+ return (ecx>>25) & 1;
+}
+#elif R123_USE_CPUID_MSVC
+R123_STATIC_INLINE int haveAESNI(){
+ int CPUInfo[4];
+ __cpuid(CPUInfo, 1);
+ return (CPUInfo[2]>>25)&1;
+}
+#else /* R123_USE_CPUID_??? */
+#warning "No R123_USE_CPUID_XXX method chosen. haveAESNI will always return false"
+R123_STATIC_INLINE int haveAESNI(){
+ return 0;
+}
+#endif /* R123_USE_ASM_GNU || R123_USE_CPUID_MSVC */
+
+// There is a lot of annoying and inexplicable variation in the
+// SSE intrinsics available in different compilation environments.
+// The details seem to depend on the compiler, the version and
+// the target architecture. Rather than insisting on
+// R123_USE_feature tests for each of these in each of the
+// compilerfeatures.h files we just keep the complexity localized
+// to here...
+#if (defined(__ICC) && __ICC<1210) || (defined(_MSC_VER) && !defined(_WIN64))
+/* Is there an intrinsic to assemble an __m128i from two 64-bit words?
+ If not, use the 4x32-bit intrisic instead. N.B. It looks like Intel
+ added _mm_set_epi64x to icc version 12.1 in Jan 2012.
+*/
+R123_STATIC_INLINE __m128i _mm_set_epi64x(uint64_t v1, uint64_t v0){
+ union{
+ uint64_t u64;
+ uint32_t u32[2];
+ } u1, u0;
+ u1.u64 = v1;
+ u0.u64 = v0;
+ return _mm_set_epi32(u1.u32[1], u1.u32[0], u0.u32[1], u0.u32[0]);
+}
+#endif
+/* _mm_extract_lo64 abstracts the task of extracting the low 64-bit
+ word from an __m128i. The _mm_cvtsi128_si64 intrinsic does the job
+ on 64-bit platforms. Unfortunately, both MSVC and Open64 fail
+ assertions in ut_M128.cpp and ut_carray.cpp when we use the
+ _mm_cvtsi128_si64 intrinsic. (See
+ https://bugs.open64.net/show_bug.cgi?id=873 for the Open64 bug).
+ On 32-bit platforms, there's no MOVQ, so there's no intrinsic.
+ Finally, even if the intrinsic exists, it may be spelled with or
+ without the 'x'.
+*/
+#if !defined(__x86_64__) || defined(_MSC_VER) || defined(__OPEN64__)
+R123_STATIC_INLINE uint64_t _mm_extract_lo64(__m128i si){
+ union{
+ uint64_t u64[2];
+ __m128i m;
+ }u;
+ _mm_store_si128(&u.m, si);
+ return u.u64[0];
+}
+#elif defined(__llvm__) || defined(__ICC)
+R123_STATIC_INLINE uint64_t _mm_extract_lo64(__m128i si){
+ return (uint64_t)_mm_cvtsi128_si64(si);
+}
+#else /* GNUC, others */
+/* FWIW, gcc's emmintrin.h has had the 'x' spelling
+ since at least gcc-3.4.4. The no-'x' spelling showed up
+ around 4.2. */
+R123_STATIC_INLINE uint64_t _mm_extract_lo64(__m128i si){
+ return (uint64_t)_mm_cvtsi128_si64x(si);
+}
+#endif
+#if defined(__GNUC__) && __GNUC__ < 4
+/* the cast builtins showed up in gcc4. */
+R123_STATIC_INLINE __m128 _mm_castsi128_ps(__m128i si){
+ return (__m128)si;
+}
+#endif
+
+#ifdef __cplusplus
+
+struct r123m128i{
+ __m128i m;
+#if R123_USE_CXX11_UNRESTRICTED_UNIONS
+ // C++98 forbids a union member from having *any* constructors.
+ // C++11 relaxes this, and allows union members to have constructors
+ // as long as there is a "trivial" default construtor. So in C++11
+ // we can provide a r123m128i constructor with an __m128i argument, and still
+ // have the default (and hence trivial) default constructor.
+ r123m128i() = default;
+ r123m128i(__m128i _m): m(_m){}
+#endif
+ r123m128i& operator=(const __m128i& rhs){ m=rhs; return *this;}
+ r123m128i& operator=(R123_ULONG_LONG n){ m = _mm_set_epi64x(0, n); return *this;}
+#if R123_USE_CXX11_EXPLICIT_CONVERSIONS
+ // With C++11 we can attach explicit to the bool conversion operator
+ // to disambiguate undesired promotions. For g++, this works
+ // only in 4.5 and above.
+ explicit operator bool() const {return _bool();}
+#else
+ // Pre-C++11, we have to do something else. Google for the "safe bool"
+ // idiom for other ideas...
+ operator const void*() const{return _bool()?this:0;}
+#endif
+ operator __m128i() const {return m;}
+
+private:
+#if R123_USE_SSE4_1
+ bool _bool() const{ return !_mm_testz_si128(m,m); }
+#else
+ bool _bool() const{ return 0xf != _mm_movemask_ps(_mm_castsi128_ps(_mm_cmpeq_epi32(m, _mm_setzero_si128()))); }
+#endif
+};
+
+R123_STATIC_INLINE r123m128i& operator++(r123m128i& v){
+ __m128i& c = v.m;
+ __m128i zeroone = _mm_set_epi64x(R123_64BIT(0), R123_64BIT(1));
+ c = _mm_add_epi64(c, zeroone);
+ //return c;
+#if R123_USE_SSE4_1
+ __m128i zerofff = _mm_set_epi64x(0, ~(R123_64BIT(0)));
+ if( R123_BUILTIN_EXPECT(_mm_testz_si128(c,zerofff), 0) ){
+ __m128i onezero = _mm_set_epi64x(R123_64BIT(1), R123_64BIT(0));
+ c = _mm_add_epi64(c, onezero);
+ }
+#else
+ unsigned mask = _mm_movemask_ps( _mm_castsi128_ps(_mm_cmpeq_epi32(c, _mm_setzero_si128())));
+ // The low two bits of mask are 11 iff the low 64 bits of
+ // c are zero.
+ if( R123_BUILTIN_EXPECT((mask&0x3) == 0x3, 0) ){
+ __m128i onezero = _mm_set_epi64x(1,0);
+ c = _mm_add_epi64(c, onezero);
+ }
+#endif
+ return v;
+}
+
+R123_STATIC_INLINE r123m128i& operator+=(r123m128i& lhs, R123_ULONG_LONG n){
+ __m128i c = lhs.m;
+ __m128i incr128 = _mm_set_epi64x(0, n);
+ c = _mm_add_epi64(c, incr128);
+ // return c; // NO CARRY!
+
+ int64_t lo64 = _mm_extract_lo64(c);
+ if((uint64_t)lo64 < n)
+ c = _mm_add_epi64(c, _mm_set_epi64x(1,0));
+ lhs.m = c;
+ return lhs;
+}
+
+// We need this one because it's present, but never used in r123array1xm128i::incr
+R123_STATIC_INLINE bool operator<=(R123_ULONG_LONG, const r123m128i &){
+ throw std::runtime_error("operator<=(unsigned long long, r123m128i) is unimplemented.");}
+
+// The comparisons aren't implemented, but if we leave them out, and
+// somebody writes, e.g., M1 < M2, the compiler will do an implicit
+// conversion through void*. Sigh...
+R123_STATIC_INLINE bool operator<(const r123m128i&, const r123m128i&){
+ throw std::runtime_error("operator<(r123m128i, r123m128i) is unimplemented.");}
+R123_STATIC_INLINE bool operator<=(const r123m128i&, const r123m128i&){
+ throw std::runtime_error("operator<=(r123m128i, r123m128i) is unimplemented.");}
+R123_STATIC_INLINE bool operator>(const r123m128i&, const r123m128i&){
+ throw std::runtime_error("operator>(r123m128i, r123m128i) is unimplemented.");}
+R123_STATIC_INLINE bool operator>=(const r123m128i&, const r123m128i&){
+ throw std::runtime_error("operator>=(r123m128i, r123m128i) is unimplemented.");}
+
+R123_STATIC_INLINE bool operator==(const r123m128i &lhs, const r123m128i &rhs){
+ return 0xf==_mm_movemask_ps(_mm_castsi128_ps(_mm_cmpeq_epi32(lhs, rhs))); }
+R123_STATIC_INLINE bool operator!=(const r123m128i &lhs, const r123m128i &rhs){
+ return !(lhs==rhs);}
+R123_STATIC_INLINE bool operator==(R123_ULONG_LONG lhs, const r123m128i &rhs){
+ r123m128i LHS; LHS.m=_mm_set_epi64x(0, lhs); return LHS == rhs; }
+R123_STATIC_INLINE bool operator!=(R123_ULONG_LONG lhs, const r123m128i &rhs){
+ return !(lhs==rhs);}
+R123_STATIC_INLINE std::ostream& operator<<(std::ostream& os, const r123m128i& m){
+ union{
+ uint64_t u64[2];
+ __m128i m;
+ }u;
+ _mm_storeu_si128(&u.m, m.m);
+ return os << u.u64[0] << " " << u.u64[1];
+}
+
+R123_STATIC_INLINE std::istream& operator>>(std::istream& is, r123m128i& m){
+ uint64_t u64[2];
+ is >> u64[0] >> u64[1];
+ m.m = _mm_set_epi64x(u64[1], u64[0]);
+ return is;
+}
+
+template<typename T> inline T assemble_from_u32(uint32_t *p32); // forward declaration
+
+template <>
+inline r123m128i assemble_from_u32<r123m128i>(uint32_t *p32){
+ r123m128i ret;
+ ret.m = _mm_set_epi32(p32[3], p32[2], p32[1], p32[0]);
+ return ret;
+}
+
+#else
+
+typedef struct {
+ __m128i m;
+} r123m128i;
+
+#endif /* __cplusplus */
+
+#else /* !R123_USE_SSE */
+R123_STATIC_INLINE int haveAESNI(){
+ return 0;
+}
+#endif /* R123_USE_SSE */
+
+#endif /* _Random123_sse_dot_h__ */
diff --git a/ext/random123/include/Random123/features/sunprofeatures.h b/ext/random123/include/Random123/features/sunprofeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..c9cdc00f5e8f970898ae577b14fa910ceb135a91
--- /dev/null
+++ b/ext/random123/include/Random123/features/sunprofeatures.h
@@ -0,0 +1,172 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __sunprofeatures_dot_hpp
+#define __sunprofeatures_dot_hpp
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) expr
+#endif
+
+// The basic idiom is:
+// #ifndef R123_SOMETHING
+// #if some condition
+// #define R123_SOMETHING 1
+// #else
+// #define R123_SOMETHING 0
+// #endif
+// #endif
+// This idiom allows an external user to override any decision
+// in this file with a command-line -DR123_SOMETHING=1 or -DR123_SOMETHINE=0
+
+// An alternative idiom is:
+// #ifndef R123_SOMETHING
+// #define R123_SOMETHING (some boolean expression)
+// #endif
+// where the boolean expression might contain previously-defined R123_SOMETHING_ELSE
+// pp-symbols.
+
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+#ifndef R123_USE_SSE4_2
+#define R123_USE_SSE4_2 0
+#endif
+
+#ifndef R123_USE_SSE4_1
+#define R123_USE_SSE4_1 0
+#endif
+
+#ifndef R123_USE_SSE
+#define R123_USE_SSE 0
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#define R123_USE_ASM_GNU 0
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 0
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#define R123_USE_X86INTRIN_H 0
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 0
+#endif
+
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+#define R123_USE_EMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#define R123_USE_INTRIN_H 0
+#endif
+
+#ifndef R123_USE_MULHILO16_ASM
+#define R123_USE_MULHILO16_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#define R123_USE_MULHILO64_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef R123_USE_PHILOX_64BIT
+#define R123_USE_PHILOX_64BIT 0
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+// If you add something, it must go in all the other XXfeatures.hpp
+// and in ../ut_features.cpp
+#endif
diff --git a/ext/random123/include/Random123/features/xlcfeatures.h b/ext/random123/include/Random123/features/xlcfeatures.h
new file mode 100644
index 0000000000000000000000000000000000000000..ccb98ee5531c57253410eb4d1fe3692ec289bbca
--- /dev/null
+++ b/ext/random123/include/Random123/features/xlcfeatures.h
@@ -0,0 +1,210 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+Copyright (c) 2013, Los Alamos National Security, LLC
+All rights reserved.
+
+Copyright 2013. Los Alamos National Security, LLC. This software was produced
+under U.S. Government contract DE-AC52-06NA25396 for Los Alamos National
+Laboratory (LANL), which is operated by Los Alamos National Security, LLC for
+the U.S. Department of Energy. The U.S. Government has rights to use,
+reproduce, and distribute this software. NEITHER THE GOVERNMENT NOR LOS
+ALAMOS NATIONAL SECURITY, LLC MAKES ANY WARRANTY, EXPRESS OR IMPLIED, OR
+ASSUMES ANY LIABILITY FOR THE USE OF THIS SOFTWARE. If software is modified
+to produce derivative works, such modified software should be clearly marked,
+so as not to confuse it with the version available from LANL.
+*/
+#ifndef __xlcfeatures_dot_hpp
+#define __xlcfeatures_dot_hpp
+
+#if !defined(__x86_64__) && !defined(__i386__) && !defined(__powerpc__)
+# error "This code has only been tested on x86 and PowerPC platforms."
+#include <including_a_nonexistent_file_will_stop_some_compilers_from_continuing_with_a_hopeless_task>
+{ /* maybe an unbalanced brace will terminate the compilation */
+ /* Feel free to try the Random123 library on other architectures by changing
+ the conditions that reach this error, but you should consider it a
+ porting exercise and expect to encounter bugs and deficiencies.
+ Please let the authors know of any successes (or failures). */
+#endif
+
+#ifdef __cplusplus
+/* builtins are automatically available to xlc. To use them with xlc++,
+ one must include builtins.h. c.f
+ http://publib.boulder.ibm.com/infocenter/cellcomp/v101v121/index.jsp?topic=/com.ibm.xlcpp101.cell.doc/compiler_ref/compiler_builtins.html
+*/
+#include <builtins.h>
+#endif
+
+#ifndef R123_STATIC_INLINE
+#define R123_STATIC_INLINE static inline
+#endif
+
+#ifndef R123_FORCE_INLINE
+#define R123_FORCE_INLINE(decl) decl __attribute__((__always_inline__))
+#endif
+
+#ifndef R123_CUDA_DEVICE
+#define R123_CUDA_DEVICE
+#endif
+
+#ifndef R123_ASSERT
+#include <assert.h>
+#define R123_ASSERT(x) assert(x)
+#endif
+
+#ifndef R123_BUILTIN_EXPECT
+#define R123_BUILTIN_EXPECT(expr,likely) __builtin_expect(expr,likely)
+#endif
+
+#ifndef R123_USE_AES_NI
+#define R123_USE_AES_NI 0
+#endif
+
+#ifndef R123_USE_SSE4_2
+#define R123_USE_SSE4_2 0
+#endif
+
+#ifndef R123_USE_SSE4_1
+#define R123_USE_SSE4_1 0
+#endif
+
+#ifndef R123_USE_SSE
+#define R123_USE_SSE 0
+#endif
+
+#ifndef R123_USE_AES_OPENSSL
+/* There isn't really a good way to tell at compile time whether
+ openssl is available. Without a pre-compilation configure-like
+ tool, it's less error-prone to guess that it isn't available. Add
+ -DR123_USE_AES_OPENSSL=1 and any necessary LDFLAGS or LDLIBS to
+ play with openssl */
+#define R123_USE_AES_OPENSSL 0
+#endif
+
+#ifndef R123_USE_GNU_UINT128
+#define R123_USE_GNU_UINT128 0
+#endif
+
+#ifndef R123_USE_ASM_GNU
+#define R123_USE_ASM_GNU 1
+#endif
+
+#ifndef R123_USE_CPUID_MSVC
+#define R123_USE_CPUID_MSVC 0
+#endif
+
+#ifndef R123_USE_X86INTRIN_H
+#define R123_USE_X86INTRIN_H 0
+#endif
+
+#ifndef R123_USE_IA32INTRIN_H
+#define R123_USE_IA32INTRIN_H 0
+#endif
+
+#ifndef R123_USE_XMMINTRIN_H
+#define R123_USE_XMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_EMMINTRIN_H
+#define R123_USE_EMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_SMMINTRIN_H
+#define R123_USE_SMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_WMMINTRIN_H
+#define R123_USE_WMMINTRIN_H 0
+#endif
+
+#ifndef R123_USE_INTRIN_H
+#ifdef __ABM__
+#define R123_USE_INTRIN_H 1
+#else
+#define R123_USE_INTRIN_H 0
+#endif
+#endif
+
+#ifndef R123_USE_MULHILO32_ASM
+#define R123_USE_MULHILO32_ASM 0
+#endif
+
+#ifndef R123_USE_MULHILO64_MULHI_INTRIN
+#if (defined(__powerpc64__))
+#define R123_USE_MULHILO64_MULHI_INTRIN 1
+#else
+#define R123_USE_MULHILO64_MULHI_INTRIN 0
+#endif
+#endif
+
+#ifndef R123_MULHILO64_MULHI_INTRIN
+#define R123_MULHILO64_MULHI_INTRIN __mulhdu
+#endif
+
+#ifndef R123_USE_MULHILO32_MULHI_INTRIN
+#define R123_USE_MULHILO32_MULHI_INTRIN 0
+#endif
+
+#ifndef R123_MULHILO32_MULHI_INTRIN
+#define R123_MULHILO32_MULHI_INTRIN __mulhwu
+#endif
+
+#ifndef R123_USE_MULHILO64_ASM
+#if defined(__powerpc64__)
+#define R123_USE_MULHILO64_ASM (1 /*defined(__powerpc64__)*/ && !(R123_USE_MULHILO64_MULHI_INTRIN))
+#else
+#define R123_USE_MULHILO64_ASM (0 /*defined(__powerpc64__)*/ && !(R123_USE_MULHILO64_MULHI_INTRIN))
+#endif
+#endif
+
+#ifndef R123_USE_MULHILO64_MSVC_INTRIN
+#define R123_USE_MULHILO64_MSVC_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_CUDA_INTRIN
+#define R123_USE_MULHILO64_CUDA_INTRIN 0
+#endif
+
+#ifndef R123_USE_MULHILO64_OPENCL_INTRIN
+#define R123_USE_MULHILO64_OPENCL_INTRIN 0
+#endif
+
+#ifndef __STDC_CONSTANT_MACROS
+#define __STDC_CONSTANT_MACROS
+#endif
+#include <stdint.h>
+#ifndef UINT64_C
+#error UINT64_C not defined. You must define __STDC_CONSTANT_MACROS before you #include <stdint.h>
+#endif
+
+/* If you add something, it must go in all the other XXfeatures.hpp
+ and in ../ut_features.cpp */
+#endif
diff --git a/ext/random123/include/Random123/gsl_microrng.h b/ext/random123/include/Random123/gsl_microrng.h
new file mode 100644
index 0000000000000000000000000000000000000000..4f09412152687462506ce88650a5328f6787ae23
--- /dev/null
+++ b/ext/random123/include/Random123/gsl_microrng.h
@@ -0,0 +1,136 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef __r123_gslmicrorng_dot_h__
+#define __r123_gslmicrorng_dot_h__
+
+
+#include <gsl/gsl_rng.h>
+#include <string.h>
+
+/** The macro: GSL_MICRORNG(NAME, CBRNGNAME) is the GSL
+ analog analog of the C++ r123::MicroURNG template. It declares a gsl_rng
+ type named gsl_rng_NAME which uses the underlying CBRNGNAME
+ and can be invoked a limited number of times between calls to NAME_reset.
+
+ When the underlying CBRNG's \c ctr_t is an \ref arrayNxW "r123arrayNxW",
+ and the gsl_rng_NAME may called up to \c N*2^32 times
+ between calls to \c NAME_reset.
+
+ \c NAME_reset takes a gsl_rng_NAME type, a counter and a key as arguments.
+ It restarts the micro-rng with a new base counter and key.
+
+ Note that you must call NAME_reset before the first use
+ of a gsl_rng. NAME_reset is not called automatically by
+ gsl_rng_alloc().
+
+ @code
+ #include <Random123/threefry.h>
+ #include <Random123/gsl_microrng.h> // this file
+ GSL_MICRORNG(microcbrng, threefry4x64, 20) // creates gsl_rng_microcbrng
+
+ int main(int argc, char** argv) {
+ gsl_rng *r = gsl_rng_alloc(gsl_rng_microcbrng);
+ threefry4x64_ctr_t c = {{}};
+ threefry4x64_key_t k = {{}};
+
+ for (...) {
+ c.v[0] = ??; // some application variable
+ microcbrng_reset(r, c, k);
+ for (...) {
+ // gaussian calls r several times. It is safe for
+ // r to be used upto 2^20 times in this loop
+ something[i] = gsl_ran_gaussian(r, 1.5);
+ }
+ }
+ }
+ @endcode
+
+*/
+
+#define GSL_MICRORNG(NAME, CBRNGNAME) \
+const gsl_rng_type *gsl_rng_##NAME; \
+ \
+typedef struct{ \
+ CBRNGNAME##_ctr_t ctr; \
+ CBRNGNAME##_ctr_t r; \
+ CBRNGNAME##_key_t key; \
+ R123_ULONG_LONG n; \
+ int elem; \
+} NAME##_state; \
+ \
+static unsigned long int NAME##_get(void *vstate){ \
+ NAME##_state *st = (NAME##_state *)vstate; \
+ const int N=sizeof(st->ctr.v)/sizeof(st->ctr.v[0]); \
+ if( st->elem == 0 ){ \
+ CBRNGNAME##_ctr_t c = st->ctr; \
+ c.v[N-1] |= st->n<<(R123_W(CBRNGNAME##_ctr_t)-32); \
+ st->n++; \
+ st->r = CBRNGNAME(c, st->key); \
+ st->elem = N; \
+ } \
+ return 0xffffffff & st->r.v[--st->elem]; \
+} \
+ \
+static double \
+NAME##_get_double (void * vstate) \
+{ \
+ return NAME##_get (vstate)/4294967296.; \
+} \
+ \
+static void NAME##_set(void *vstate, unsigned long int s){ \
+ NAME##_state *st = (NAME##_state *)vstate; \
+ (void)s; /* ignored */ \
+ st->elem = 0; \
+ st->n = ~0; /* will abort if _reset is not called */ \
+} \
+ \
+static const gsl_rng_type NAME##_type = { \
+ #NAME, \
+ 0xffffffffUL, \
+ 0, \
+ sizeof(NAME##_state), \
+ &NAME##_set, \
+ &NAME##_get, \
+ &NAME##_get_double \
+}; \
+ \
+R123_STATIC_INLINE void NAME##_reset(const gsl_rng* gr, CBRNGNAME##_ctr_t c, CBRNGNAME##_key_t k) { \
+ NAME##_state* state = (NAME##_state *)gr->state; \
+ state->ctr = c; \
+ state->key = k; \
+ state->n = 0; \
+ state->elem = 0; \
+} \
+ \
+const gsl_rng_type *gsl_rng_##NAME = &NAME##_type
+
+#endif
diff --git a/ext/random123/include/Random123/philox.h b/ext/random123/include/Random123/philox.h
new file mode 100644
index 0000000000000000000000000000000000000000..7bf4d195772358a87b8fbb33667783b5caba61a4
--- /dev/null
+++ b/ext/random123/include/Random123/philox.h
@@ -0,0 +1,493 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef _philox_dot_h_
+#define _philox_dot_h_
+
+/** \cond HIDDEN_FROM_DOXYGEN */
+
+#include "features/compilerfeatures.h"
+#include "array.h"
+
+
+/*
+// Macros _Foo_tpl are code generation 'templates' They define
+// inline functions with names obtained by mangling Foo and the
+// macro arguments. E.g.,
+// _mulhilo_tpl(32, uint32_t, uint64_t)
+// expands to a definition of:
+// mulhilo32(uint32_t, uint32_t, uint32_t *, uint32_t *)
+// We then 'instantiate the template' to define
+// several different functions, e.g.,
+// mulhilo32
+// mulhilo64
+// These functions will be visible to user code, and may
+// also be used later in subsequent templates and definitions.
+
+// A template for mulhilo using a temporary of twice the word-width.
+// Gcc figures out that this can be reduced to a single 'mul' instruction,
+// despite the apparent use of double-wide variables, shifts, etc. It's
+// obviously not guaranteed that all compilers will be that smart, so
+// other implementations might be preferable, e.g., using an intrinsic
+// or an asm block. On the other hand, for 32-bit multiplies,
+// this *is* perfectly standard C99 - any C99 compiler should
+// understand it and produce correct code. For 64-bit multiplies,
+// it's only usable if the compiler recognizes that it can do
+// arithmetic on a 128-bit type. That happens to be true for gcc on
+// x86-64, and powerpc64 but not much else.
+*/
+#define _mulhilo_dword_tpl(W, Word, Dword) \
+R123_CUDA_DEVICE R123_STATIC_INLINE Word mulhilo##W(Word a, Word b, Word* hip){ \
+ Dword product = ((Dword)a)*((Dword)b); \
+ *hip = product>>W; \
+ return (Word)product; \
+}
+
+/*
+// A template for mulhilo using gnu-style asm syntax.
+// INSN can be "mulw", "mull" or "mulq".
+// FIXME - porting to other architectures, we'll need still-more conditional
+// branching here. Note that intrinsics are usually preferable.
+*/
+#ifdef __powerpc__
+#define _mulhilo_asm_tpl(W, Word, INSN) \
+R123_STATIC_INLINE Word mulhilo##W(Word ax, Word b, Word *hip){ \
+ Word dx = 0; \
+ __asm__("\n\t" \
+ INSN " %0,%1,%2\n\t" \
+ : "=r"(dx) \
+ : "r"(b), "r"(ax) \
+ ); \
+ *hip = dx; \
+ return ax*b; \
+}
+#else
+#define _mulhilo_asm_tpl(W, Word, INSN) \
+R123_STATIC_INLINE Word mulhilo##W(Word ax, Word b, Word *hip){ \
+ Word dx; \
+ __asm__("\n\t" \
+ INSN " %2\n\t" \
+ : "=a"(ax), "=d"(dx) \
+ : "r"(b), "0"(ax) \
+ ); \
+ *hip = dx; \
+ return ax; \
+}
+#endif /* __powerpc__ */
+
+/*
+// A template for mulhilo using MSVC-style intrinsics
+// For example,_umul128 is an msvc intrinsic, c.f.
+// http://msdn.microsoft.com/en-us/library/3dayytw9.aspx
+*/
+#define _mulhilo_msvc_intrin_tpl(W, Word, INTRIN) \
+R123_STATIC_INLINE Word mulhilo##W(Word a, Word b, Word* hip){ \
+ return INTRIN(a, b, hip); \
+}
+
+/* N.B. This really should be called _mulhilo_mulhi_intrin. It just
+ happens that CUDA was the first time we used the idiom. */
+#define _mulhilo_cuda_intrin_tpl(W, Word, INTRIN) \
+R123_CUDA_DEVICE R123_STATIC_INLINE Word mulhilo##W(Word a, Word b, R123_METAL_THREAD_ADDRESS_SPACE Word* hip){ \
+ *hip = INTRIN(a, b); \
+ return a*b; \
+}
+
+/*
+// A template for mulhilo using only word-size operations and
+// C99 operators (no adc, no mulhi). It
+// requires four multiplies and a dozen or so shifts, adds
+// and tests. It's *SLOW*. It can be used to
+// implement philoxNx32 on platforms that completely lack
+// 64-bit types, e.g., Metal.
+// On 32-bit platforms, it could be used to
+// implement philoxNx64, but on such platforms both the philoxNx32
+// and the threefryNx64 cbrngs are going to have much better
+// performance. It is enabled below by R123_USE_MULHILO64_C99,
+// but that is currently (Feb 2019) only set by
+// features/metalfeatures.h headers. It can, of course, be
+// set with a compile-time -D option.
+*/
+#define _mulhilo_c99_tpl(W, Word) \
+R123_STATIC_INLINE Word mulhilo##W(Word a, Word b, R123_METAL_THREAD_ADDRESS_SPACE Word *hip){ \
+ const unsigned WHALF = W/2; \
+ const Word LOMASK = ((((Word)1)<<WHALF)-1); \
+ Word lo = a*b; /* full low multiply */ \
+ Word ahi = a>>WHALF; \
+ Word alo = a& LOMASK; \
+ Word bhi = b>>WHALF; \
+ Word blo = b& LOMASK; \
+ \
+ Word ahbl = ahi*blo; \
+ Word albh = alo*bhi; \
+ \
+ Word ahbl_albh = ((ahbl&LOMASK) + (albh&LOMASK)); \
+ Word hi = ahi*bhi + (ahbl>>WHALF) + (albh>>WHALF); \
+ hi += ahbl_albh >> WHALF; /* carry from the sum of lo(ahbl) + lo(albh) ) */ \
+ /* carry from the sum with alo*blo */ \
+ hi += ((lo >> WHALF) < (ahbl_albh&LOMASK)); \
+ *hip = hi; \
+ return lo; \
+}
+
+/*
+// A template for mulhilo on a platform that can't do it
+// We could put a C version here, but is it better to run *VERY*
+// slowly or to just stop and force the user to find another CBRNG?
+*/
+#define _mulhilo_fail_tpl(W, Word) \
+R123_STATIC_INLINE Word mulhilo##W(Word a, Word b, Word *hip){ \
+ R123_STATIC_ASSERT(0, "mulhilo" #W " is not implemented on this machine\n"); \
+}
+
+/*
+// N.B. There's an MSVC intrinsic called _emul,
+// which *might* compile into better code than
+// _mulhilo_dword_tpl
+*/
+#if R123_USE_MULHILO32_ASM
+#ifdef __powerpc__
+_mulhilo_asm_tpl(32, uint32_t, "mulhwu")
+#else
+_mulhilo_asm_tpl(32, uint32_t, "mull")
+#endif /* __powerpc__ */
+#else
+#if R123_USE_64BIT
+_mulhilo_dword_tpl(32, uint32_t, uint64_t)
+#elif R123_USE_MULHILO32_MULHI_INTRIN
+_mulhilo_cuda_intrin_tpl(32, uint32_t, R123_MULHILO32_MULHI_INTRIN)
+#else
+_mulhilo_c99_tpl(32, uint32_t)
+#endif
+#endif
+
+#if R123_USE_PHILOX_64BIT
+#if R123_USE_MULHILO64_ASM
+#ifdef __powerpc64__
+_mulhilo_asm_tpl(64, uint64_t, "mulhdu")
+#else
+_mulhilo_asm_tpl(64, uint64_t, "mulq")
+#endif /* __powerpc64__ */
+#elif R123_USE_MULHILO64_MSVC_INTRIN
+_mulhilo_msvc_intrin_tpl(64, uint64_t, _umul128)
+#elif R123_USE_MULHILO64_CUDA_INTRIN
+_mulhilo_cuda_intrin_tpl(64, uint64_t, __umul64hi)
+#elif R123_USE_MULHILO64_OPENCL_INTRIN
+_mulhilo_cuda_intrin_tpl(64, uint64_t, mul_hi)
+#elif R123_USE_MULHILO64_MULHI_INTRIN
+_mulhilo_cuda_intrin_tpl(64, uint64_t, R123_MULHILO64_MULHI_INTRIN)
+#elif R123_USE_GNU_UINT128
+_mulhilo_dword_tpl(64, uint64_t, __uint128_t)
+#elif R123_USE_MULHILO64_C99
+_mulhilo_c99_tpl(64, uint64_t)
+#else
+_mulhilo_fail_tpl(64, uint64_t)
+#endif
+#endif
+
+/*
+// The multipliers and Weyl constants are "hard coded".
+// To change them, you can #define them with different
+// values before #include-ing this file.
+// This isn't terribly elegant, but it works for C as
+// well as C++. A nice C++-only solution would be to
+// use template parameters in the style of <random>
+*/
+#ifndef PHILOX_M2x64_0
+#define PHILOX_M2x64_0 R123_64BIT(0xD2B74407B1CE6E93)
+#endif
+
+#ifndef PHILOX_M4x64_0
+#define PHILOX_M4x64_0 R123_64BIT(0xD2E7470EE14C6C93)
+#endif
+
+#ifndef PHILOX_M4x64_1
+#define PHILOX_M4x64_1 R123_64BIT(0xCA5A826395121157)
+#endif
+
+#ifndef PHILOX_M2x32_0
+#define PHILOX_M2x32_0 ((uint32_t)0xd256d193)
+#endif
+
+#ifndef PHILOX_M4x32_0
+#define PHILOX_M4x32_0 ((uint32_t)0xD2511F53)
+#endif
+#ifndef PHILOX_M4x32_1
+#define PHILOX_M4x32_1 ((uint32_t)0xCD9E8D57)
+#endif
+
+#ifndef PHILOX_W64_0
+#define PHILOX_W64_0 R123_64BIT(0x9E3779B97F4A7C15) /* golden ratio */
+#endif
+#ifndef PHILOX_W64_1
+#define PHILOX_W64_1 R123_64BIT(0xBB67AE8584CAA73B) /* sqrt(3)-1 */
+#endif
+
+#ifndef PHILOX_W32_0
+#define PHILOX_W32_0 ((uint32_t)0x9E3779B9)
+#endif
+#ifndef PHILOX_W32_1
+#define PHILOX_W32_1 ((uint32_t)0xBB67AE85)
+#endif
+
+/** \endcond */
+#ifndef PHILOX2x32_DEFAULT_ROUNDS
+#define PHILOX2x32_DEFAULT_ROUNDS 10
+#endif
+
+#ifndef PHILOX2x64_DEFAULT_ROUNDS
+#define PHILOX2x64_DEFAULT_ROUNDS 10
+#endif
+
+#ifndef PHILOX4x32_DEFAULT_ROUNDS
+#define PHILOX4x32_DEFAULT_ROUNDS 10
+#endif
+
+#ifndef PHILOX4x64_DEFAULT_ROUNDS
+#define PHILOX4x64_DEFAULT_ROUNDS 10
+#endif
+/** \cond HIDDEN_FROM_DOXYGEN */
+
+/* The ignored fourth argument allows us to instantiate the
+ same macro regardless of N. */
+#define _philox2xWround_tpl(W, T) \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(struct r123array2x##W _philox2x##W##round(struct r123array2x##W ctr, struct r123array1x##W key)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE struct r123array2x##W _philox2x##W##round(struct r123array2x##W ctr, struct r123array1x##W key){ \
+ T hi; \
+ T lo = mulhilo##W(PHILOX_M2x##W##_0, ctr.v[0], &hi); \
+ struct r123array2x##W out = {{hi^key.v[0]^ctr.v[1], lo}}; \
+ return out; \
+}
+#define _philox2xWbumpkey_tpl(W) \
+R123_CUDA_DEVICE R123_STATIC_INLINE struct r123array1x##W _philox2x##W##bumpkey( struct r123array1x##W key) { \
+ key.v[0] += PHILOX_W##W##_0; \
+ return key; \
+}
+
+#define _philox4xWround_tpl(W, T) \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(struct r123array4x##W _philox4x##W##round(struct r123array4x##W ctr, struct r123array2x##W key)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE struct r123array4x##W _philox4x##W##round(struct r123array4x##W ctr, struct r123array2x##W key){ \
+ T hi0; \
+ T hi1; \
+ T lo0 = mulhilo##W(PHILOX_M4x##W##_0, ctr.v[0], &hi0); \
+ T lo1 = mulhilo##W(PHILOX_M4x##W##_1, ctr.v[2], &hi1); \
+ struct r123array4x##W out = {{hi1^ctr.v[1]^key.v[0], lo1, \
+ hi0^ctr.v[3]^key.v[1], lo0}}; \
+ return out; \
+}
+
+#define _philox4xWbumpkey_tpl(W) \
+R123_CUDA_DEVICE R123_STATIC_INLINE struct r123array2x##W _philox4x##W##bumpkey( struct r123array2x##W key) { \
+ key.v[0] += PHILOX_W##W##_0; \
+ key.v[1] += PHILOX_W##W##_1; \
+ return key; \
+}
+
+/** \endcond */
+#define _philoxNxW_tpl(N, Nhalf, W, T) \
+/** @ingroup PhiloxNxW */ \
+enum r123_enum_philox##N##x##W { philox##N##x##W##_rounds = PHILOX##N##x##W##_DEFAULT_ROUNDS }; \
+typedef struct r123array##N##x##W philox##N##x##W##_ctr_t; \
+typedef struct r123array##Nhalf##x##W philox##N##x##W##_key_t; \
+typedef struct r123array##Nhalf##x##W philox##N##x##W##_ukey_t; \
+R123_CUDA_DEVICE R123_STATIC_INLINE philox##N##x##W##_key_t philox##N##x##W##keyinit(philox##N##x##W##_ukey_t uk) { return uk; } \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(philox##N##x##W##_ctr_t philox##N##x##W##_R(unsigned int R, philox##N##x##W##_ctr_t ctr, philox##N##x##W##_key_t key)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE philox##N##x##W##_ctr_t philox##N##x##W##_R(unsigned int R, philox##N##x##W##_ctr_t ctr, philox##N##x##W##_key_t key) { \
+ R123_ASSERT(R<=16); \
+ if(R>0){ ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>1){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>2){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>3){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>4){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>5){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>6){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>7){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>8){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>9){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>10){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>11){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>12){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>13){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>14){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ if(R>15){ key = _philox##N##x##W##bumpkey(key); ctr = _philox##N##x##W##round(ctr, key); } \
+ return ctr; \
+}
+
+_philox2xWbumpkey_tpl(32)
+_philox4xWbumpkey_tpl(32)
+_philox2xWround_tpl(32, uint32_t) /* philox2x32round */
+_philox4xWround_tpl(32, uint32_t) /* philo4x32round */
+
+_philoxNxW_tpl(2, 1, 32, uint32_t) /* philox2x32bijection */
+_philoxNxW_tpl(4, 2, 32, uint32_t) /* philox4x32bijection */
+#if R123_USE_PHILOX_64BIT
+/** \cond HIDDEN_FROM_DOXYGEN */
+_philox2xWbumpkey_tpl(64)
+_philox4xWbumpkey_tpl(64)
+_philox2xWround_tpl(64, uint64_t) /* philo2x64round */
+_philox4xWround_tpl(64, uint64_t) /* philo4x64round */
+/** \endcond */
+_philoxNxW_tpl(2, 1, 64, uint64_t) /* philox2x64bijection */
+_philoxNxW_tpl(4, 2, 64, uint64_t) /* philox4x64bijection */
+#endif /* R123_USE_PHILOX_64BIT */
+
+#define philox2x32(c,k) philox2x32_R(philox2x32_rounds, c, k)
+#define philox4x32(c,k) philox4x32_R(philox4x32_rounds, c, k)
+#if R123_USE_PHILOX_64BIT
+#define philox2x64(c,k) philox2x64_R(philox2x64_rounds, c, k)
+#define philox4x64(c,k) philox4x64_R(philox4x64_rounds, c, k)
+#endif /* R123_USE_PHILOX_64BIT */
+
+#if defined(__cplusplus)
+
+#define _PhiloxNxW_base_tpl(CType, KType, N, W) \
+namespace r123{ \
+template<unsigned int ROUNDS> \
+struct Philox##N##x##W##_R{ \
+ typedef CType ctr_type; \
+ typedef KType key_type; \
+ typedef KType ukey_type; \
+ static const R123_METAL_CONSTANT_ADDRESS_SPACE unsigned int rounds=ROUNDS; \
+ inline R123_CUDA_DEVICE R123_FORCE_INLINE(ctr_type operator()(ctr_type ctr, key_type key) const){ \
+ R123_STATIC_ASSERT(ROUNDS<=16, "philox is only unrolled up to 16 rounds\n"); \
+ return philox##N##x##W##_R(ROUNDS, ctr, key); \
+ } \
+}; \
+typedef Philox##N##x##W##_R<philox##N##x##W##_rounds> Philox##N##x##W; \
+ } // namespace r123
+
+_PhiloxNxW_base_tpl(r123array2x32, r123array1x32, 2, 32) // Philox2x32_R<R>
+_PhiloxNxW_base_tpl(r123array4x32, r123array2x32, 4, 32) // Philox4x32_R<R>
+#if R123_USE_PHILOX_64BIT
+_PhiloxNxW_base_tpl(r123array2x64, r123array1x64, 2, 64) // Philox2x64_R<R>
+_PhiloxNxW_base_tpl(r123array4x64, r123array2x64, 4, 64) // Philox4x64_R<R>
+#endif
+
+/* The _tpl macros don't quite work to do string-pasting inside comments.
+ so we just write out the boilerplate documentation four times... */
+
+/**
+@defgroup PhiloxNxW Philox Classes and Typedefs
+
+The PhiloxNxW classes export the member functions, typedefs and
+operator overloads required by a @ref CBRNG "CBRNG" class.
+
+As described in
+<a href="http://dl.acm.org/citation.cfm?doid=2063405"><i>Parallel Random Numbers: As Easy as 1, 2, 3</i> </a>.
+The Philox family of counter-based RNGs use integer multiplication, xor and permutation of W-bit words
+to scramble its N-word input key. Philox is a mnemonic for Product HI LO Xor).
+
+
+@class r123::Philox2x32_R
+@ingroup PhiloxNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Philox round
+function will be applied.
+
+As of November 2011, the authors know of no statistical flaws with
+ROUNDS=6 or more for Philox2x32.
+
+@typedef r123::Philox2x32
+@ingroup PhiloxNxW
+ Philox2x32 is equivalent to Philox2x32_R<10>. With 10 rounds,
+ Philox2x32 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+
+
+@class r123::Philox2x64_R
+@ingroup PhiloxNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Philox round
+function will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=6 or more for Philox2x64.
+
+@typedef r123::Philox2x64
+@ingroup PhiloxNxW
+ Philox2x64 is equivalent to Philox2x64_R<10>. With 10 rounds,
+ Philox2x64 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+
+
+@class r123::Philox4x32_R
+@ingroup PhiloxNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Philox round
+function will be applied.
+
+In November 2011, the authors recorded some suspicious p-values (approximately 1.e-7) from
+some very long (longer than the default BigCrush length) SimpPoker tests. Despite
+the fact that even longer tests reverted to "passing" p-values, a cloud remains over
+Philox4x32 with 7 rounds. The authors know of no statistical flaws with
+ROUNDS=8 or more for Philox4x32.
+
+@typedef r123::Philox4x32
+@ingroup PhiloxNxW
+ Philox4x32 is equivalent to Philox4x32_R<10>. With 10 rounds,
+ Philox4x32 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+
+
+@class r123::Philox4x64_R
+@ingroup PhiloxNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Philox round
+function will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=7 or more for Philox4x64.
+
+@typedef r123::Philox4x64
+@ingroup PhiloxNxW
+ Philox4x64 is equivalent to Philox4x64_R<10>. With 10 rounds,
+ Philox4x64 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+*/
+
+#endif /* __cplusplus */
+
+#endif /* _philox_dot_h_ */
diff --git a/ext/random123/include/Random123/threefry.h b/ext/random123/include/Random123/threefry.h
new file mode 100644
index 0000000000000000000000000000000000000000..390ceffe6865e6d23d7c69b38fb8f022abc532f6
--- /dev/null
+++ b/ext/random123/include/Random123/threefry.h
@@ -0,0 +1,870 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef _threefry_dot_h_
+#define _threefry_dot_h_
+#include "features/compilerfeatures.h"
+#include "array.h"
+
+/** \cond HIDDEN_FROM_DOXYGEN */
+/* Significant parts of this file were copied from
+ from:
+ Skein_FinalRnd/ReferenceImplementation/skein.h
+ Skein_FinalRnd/ReferenceImplementation/skein_block.c
+
+ in http://csrc.nist.gov/groups/ST/hash/sha-3/Round3/documents/Skein_FinalRnd.zip
+
+ This file has been modified so that it may no longer perform its originally
+ intended function. If you're looking for a Skein or Threefish source code,
+ please consult the original file.
+
+ The original file had the following header:
+**************************************************************************
+**
+** Interface declarations and internal definitions for Skein hashing.
+**
+** Source code author: Doug Whiting, 2008.
+**
+** This algorithm and source code is released to the public domain.
+**
+***************************************************************************
+
+*/
+
+/* See comment at the top of philox.h for the macro pre-process
+ strategy. */
+
+/* Rotation constants: */
+enum r123_enum_threefry64x4 {
+ /* These are the R_256 constants from the Threefish reference sources
+ with names changed to R_64x4... */
+ R_64x4_0_0=14, R_64x4_0_1=16,
+ R_64x4_1_0=52, R_64x4_1_1=57,
+ R_64x4_2_0=23, R_64x4_2_1=40,
+ R_64x4_3_0= 5, R_64x4_3_1=37,
+ R_64x4_4_0=25, R_64x4_4_1=33,
+ R_64x4_5_0=46, R_64x4_5_1=12,
+ R_64x4_6_0=58, R_64x4_6_1=22,
+ R_64x4_7_0=32, R_64x4_7_1=32
+};
+
+enum r123_enum_threefry64x2 {
+ /*
+ // Output from skein_rot_search: (srs64_B64-X1000)
+ // Random seed = 1. BlockSize = 128 bits. sampleCnt = 1024. rounds = 8, minHW_or=57
+ // Start: Tue Mar 1 10:07:48 2011
+ // rMin = 0.136. #0325[*15] [CRC=455A682F. hw_OR=64. cnt=16384. blkSize= 128].format
+ */
+ R_64x2_0_0=16,
+ R_64x2_1_0=42,
+ R_64x2_2_0=12,
+ R_64x2_3_0=31,
+ R_64x2_4_0=16,
+ R_64x2_5_0=32,
+ R_64x2_6_0=24,
+ R_64x2_7_0=21
+ /* 4 rounds: minHW = 4 [ 4 4 4 4 ]
+ // 5 rounds: minHW = 8 [ 8 8 8 8 ]
+ // 6 rounds: minHW = 16 [ 16 16 16 16 ]
+ // 7 rounds: minHW = 32 [ 32 32 32 32 ]
+ // 8 rounds: minHW = 64 [ 64 64 64 64 ]
+ // 9 rounds: minHW = 64 [ 64 64 64 64 ]
+ //10 rounds: minHW = 64 [ 64 64 64 64 ]
+ //11 rounds: minHW = 64 [ 64 64 64 64 ] */
+};
+
+enum r123_enum_threefry32x4 {
+ /* Output from skein_rot_search: (srs-B128-X5000.out)
+ // Random seed = 1. BlockSize = 64 bits. sampleCnt = 1024. rounds = 8, minHW_or=28
+ // Start: Mon Aug 24 22:41:36 2009
+ // ...
+ // rMin = 0.472. #0A4B[*33] [CRC=DD1ECE0F. hw_OR=31. cnt=16384. blkSize= 128].format */
+ R_32x4_0_0=10, R_32x4_0_1=26,
+ R_32x4_1_0=11, R_32x4_1_1=21,
+ R_32x4_2_0=13, R_32x4_2_1=27,
+ R_32x4_3_0=23, R_32x4_3_1= 5,
+ R_32x4_4_0= 6, R_32x4_4_1=20,
+ R_32x4_5_0=17, R_32x4_5_1=11,
+ R_32x4_6_0=25, R_32x4_6_1=10,
+ R_32x4_7_0=18, R_32x4_7_1=20
+
+ /* 4 rounds: minHW = 3 [ 3 3 3 3 ]
+ // 5 rounds: minHW = 7 [ 7 7 7 7 ]
+ // 6 rounds: minHW = 12 [ 13 12 13 12 ]
+ // 7 rounds: minHW = 22 [ 22 23 22 23 ]
+ // 8 rounds: minHW = 31 [ 31 31 31 31 ]
+ // 9 rounds: minHW = 32 [ 32 32 32 32 ]
+ //10 rounds: minHW = 32 [ 32 32 32 32 ]
+ //11 rounds: minHW = 32 [ 32 32 32 32 ] */
+
+};
+
+enum r123_enum_threefry32x2 {
+ /* Output from skein_rot_search (srs32x2-X5000.out)
+ // Random seed = 1. BlockSize = 64 bits. sampleCnt = 1024. rounds = 8, minHW_or=28
+ // Start: Tue Jul 12 11:11:33 2011
+ // rMin = 0.334. #0206[*07] [CRC=1D9765C0. hw_OR=32. cnt=16384. blkSize= 64].format */
+ R_32x2_0_0=13,
+ R_32x2_1_0=15,
+ R_32x2_2_0=26,
+ R_32x2_3_0= 6,
+ R_32x2_4_0=17,
+ R_32x2_5_0=29,
+ R_32x2_6_0=16,
+ R_32x2_7_0=24
+
+ /* 4 rounds: minHW = 4 [ 4 4 4 4 ]
+ // 5 rounds: minHW = 6 [ 6 8 6 8 ]
+ // 6 rounds: minHW = 9 [ 9 12 9 12 ]
+ // 7 rounds: minHW = 16 [ 16 24 16 24 ]
+ // 8 rounds: minHW = 32 [ 32 32 32 32 ]
+ // 9 rounds: minHW = 32 [ 32 32 32 32 ]
+ //10 rounds: minHW = 32 [ 32 32 32 32 ]
+ //11 rounds: minHW = 32 [ 32 32 32 32 ] */
+ };
+
+enum r123_enum_threefry_wcnt {
+ WCNT2=2,
+ WCNT4=4
+};
+
+#if R123_USE_64BIT
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(uint64_t RotL_64(uint64_t x, unsigned int N));
+R123_CUDA_DEVICE R123_STATIC_INLINE uint64_t RotL_64(uint64_t x, unsigned int N)
+{
+ return (x << (N & 63)) | (x >> ((64-N) & 63));
+}
+#endif
+
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(uint32_t RotL_32(uint32_t x, unsigned int N));
+R123_CUDA_DEVICE R123_STATIC_INLINE uint32_t RotL_32(uint32_t x, unsigned int N)
+{
+ return (x << (N & 31)) | (x >> ((32-N) & 31));
+}
+
+#define SKEIN_MK_64(hi32,lo32) ((lo32) + (((uint64_t) (hi32)) << 32))
+#define SKEIN_KS_PARITY64 SKEIN_MK_64(0x1BD11BDA,0xA9FC1A22)
+#define SKEIN_KS_PARITY32 0x1BD11BDA
+
+/** \endcond */
+
+#ifndef THREEFRY2x32_DEFAULT_ROUNDS
+#define THREEFRY2x32_DEFAULT_ROUNDS 20
+#endif
+
+#ifndef THREEFRY2x64_DEFAULT_ROUNDS
+#define THREEFRY2x64_DEFAULT_ROUNDS 20
+#endif
+
+#ifndef THREEFRY4x32_DEFAULT_ROUNDS
+#define THREEFRY4x32_DEFAULT_ROUNDS 20
+#endif
+
+#ifndef THREEFRY4x64_DEFAULT_ROUNDS
+#define THREEFRY4x64_DEFAULT_ROUNDS 20
+#endif
+
+#define _threefry2x_tpl(W) \
+typedef struct r123array2x##W threefry2x##W##_ctr_t; \
+typedef struct r123array2x##W threefry2x##W##_key_t; \
+typedef struct r123array2x##W threefry2x##W##_ukey_t; \
+R123_CUDA_DEVICE R123_STATIC_INLINE threefry2x##W##_key_t threefry2x##W##keyinit(threefry2x##W##_ukey_t uk) { return uk; } \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry2x##W##_ctr_t threefry2x##W##_R(unsigned int Nrounds, threefry2x##W##_ctr_t in, threefry2x##W##_key_t k)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE \
+threefry2x##W##_ctr_t threefry2x##W##_R(unsigned int Nrounds, threefry2x##W##_ctr_t in, threefry2x##W##_key_t k){ \
+ threefry2x##W##_ctr_t X; \
+ uint##W##_t ks[2+1]; \
+ int i; /* avoid size_t to avoid need for stddef.h */ \
+ R123_ASSERT(Nrounds<=32); \
+ ks[2] = SKEIN_KS_PARITY##W; \
+ for (i=0;i < 2; i++) \
+ { \
+ ks[i] = k.v[i]; \
+ X.v[i] = in.v[i]; \
+ ks[2] ^= k.v[i]; \
+ } \
+ \
+ /* Insert initial key before round 0 */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; \
+ \
+ if(Nrounds>0){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>1){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>2){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>3){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>3){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; \
+ X.v[1] += 1; /* X.v[2-1] += r */ \
+ } \
+ if(Nrounds>4){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>5){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>6){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>7){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>7){ \
+ /* InjectKey(r=2) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[0]; \
+ X.v[1] += 2; \
+ } \
+ if(Nrounds>8){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>9){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>10){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>11){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>11){ \
+ /* InjectKey(r=3) */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; \
+ X.v[1] += 3; \
+ } \
+ if(Nrounds>12){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>13){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>14){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>15){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>15){ \
+ /* InjectKey(r=4) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; \
+ X.v[1] += 4; \
+ } \
+ if(Nrounds>16){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>17){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>18){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>19){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>19){ \
+ /* InjectKey(r=5) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[0]; \
+ X.v[1] += 5; \
+ } \
+ if(Nrounds>20){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>21){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>22){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>23){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>23){ \
+ /* InjectKey(r=6) */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; \
+ X.v[1] += 6; \
+ } \
+ if(Nrounds>24){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>25){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>26){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>27){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>27){ \
+ /* InjectKey(r=7) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; \
+ X.v[1] += 7; \
+ } \
+ if(Nrounds>28){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>29){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>30){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>31){ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
+ if(Nrounds>31){ \
+ /* InjectKey(r=8) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[0]; \
+ X.v[1] += 8; \
+ } \
+ return X; \
+} \
+ /** @ingroup ThreefryNxW */ \
+enum r123_enum_threefry2x##W { threefry2x##W##_rounds = THREEFRY2x##W##_DEFAULT_ROUNDS }; \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry2x##W##_ctr_t threefry2x##W(threefry2x##W##_ctr_t in, threefry2x##W##_key_t k)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE \
+threefry2x##W##_ctr_t threefry2x##W(threefry2x##W##_ctr_t in, threefry2x##W##_key_t k){ \
+ return threefry2x##W##_R(threefry2x##W##_rounds, in, k); \
+}
+
+
+#define _threefry4x_tpl(W) \
+typedef struct r123array4x##W threefry4x##W##_ctr_t; \
+typedef struct r123array4x##W threefry4x##W##_key_t; \
+typedef struct r123array4x##W threefry4x##W##_ukey_t; \
+R123_CUDA_DEVICE R123_STATIC_INLINE threefry4x##W##_key_t threefry4x##W##keyinit(threefry4x##W##_ukey_t uk) { return uk; } \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry4x##W##_ctr_t threefry4x##W##_R(unsigned int Nrounds, threefry4x##W##_ctr_t in, threefry4x##W##_key_t k)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE \
+threefry4x##W##_ctr_t threefry4x##W##_R(unsigned int Nrounds, threefry4x##W##_ctr_t in, threefry4x##W##_key_t k){ \
+ threefry4x##W##_ctr_t X; \
+ uint##W##_t ks[4+1]; \
+ int i; /* avoid size_t to avoid need for stddef.h */ \
+ R123_ASSERT(Nrounds<=72); \
+ ks[4] = SKEIN_KS_PARITY##W; \
+ for (i=0;i < 4; i++) \
+ { \
+ ks[i] = k.v[i]; \
+ X.v[i] = in.v[i]; \
+ ks[4] ^= k.v[i]; \
+ } \
+ \
+ /* Insert initial key before round 0 */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
+ \
+ if(Nrounds>0){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>1){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>2){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>3){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>3){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
+ X.v[4-1] += 1; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>4){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>5){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>6){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>7){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>7){ \
+ /* InjectKey(r=2) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
+ X.v[4-1] += 2; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>8){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>9){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>10){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>11){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>11){ \
+ /* InjectKey(r=3) */ \
+ X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
+ X.v[4-1] += 3; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>12){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>13){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>14){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>15){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>15){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
+ X.v[4-1] += 4; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>16){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>17){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>18){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>19){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>19){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
+ X.v[4-1] += 5; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>20){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>21){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>22){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>23){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>23){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
+ X.v[4-1] += 6; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>24){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>25){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>26){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>27){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>27){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
+ X.v[4-1] += 7; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>28){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>29){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>30){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>31){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>31){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
+ X.v[4-1] += 8; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>32){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>33){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>34){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>35){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>35){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
+ X.v[4-1] += 9; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>36){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>37){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>38){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>39){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>39){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
+ X.v[4-1] += 10; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>40){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>41){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>42){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>43){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>43){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
+ X.v[4-1] += 11; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>44){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>45){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>46){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>47){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>47){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
+ X.v[4-1] += 12; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>48){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>49){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>50){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>51){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>51){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
+ X.v[4-1] += 13; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>52){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>53){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>54){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>55){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>55){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
+ X.v[4-1] += 14; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>56){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>57){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>58){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>59){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>59){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
+ X.v[4-1] += 15; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>60){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>61){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>62){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>63){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>63){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
+ X.v[4-1] += 16; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>64){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>65){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>66){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>67){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>67){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
+ X.v[4-1] += 17; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ if(Nrounds>68){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>69){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>70){ \
+ X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
+ X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
+ } \
+ if(Nrounds>71){ \
+ X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
+ X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
+ } \
+ if(Nrounds>71){ \
+ /* InjectKey(r=1) */ \
+ X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
+ X.v[4-1] += 18; /* X.v[WCNT4-1] += r */ \
+ } \
+ \
+ return X; \
+} \
+ \
+ /** @ingroup ThreefryNxW */ \
+enum r123_enum_threefry4x##W { threefry4x##W##_rounds = THREEFRY4x##W##_DEFAULT_ROUNDS }; \
+R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry4x##W##_ctr_t threefry4x##W(threefry4x##W##_ctr_t in, threefry4x##W##_key_t k)); \
+R123_CUDA_DEVICE R123_STATIC_INLINE \
+threefry4x##W##_ctr_t threefry4x##W(threefry4x##W##_ctr_t in, threefry4x##W##_key_t k){ \
+ return threefry4x##W##_R(threefry4x##W##_rounds, in, k); \
+}
+
+#if R123_USE_64BIT
+_threefry2x_tpl(64)
+_threefry4x_tpl(64)
+#endif
+_threefry2x_tpl(32)
+_threefry4x_tpl(32)
+
+/* gcc4.5 and 4.6 seem to optimize a macro-ized threefryNxW better
+ than a static inline function. Why? */
+#define threefry2x32(c,k) threefry2x32_R(threefry2x32_rounds, c, k)
+#define threefry4x32(c,k) threefry4x32_R(threefry4x32_rounds, c, k)
+#define threefry2x64(c,k) threefry2x64_R(threefry2x64_rounds, c, k)
+#define threefry4x64(c,k) threefry4x64_R(threefry4x64_rounds, c, k)
+
+#if defined(__cplusplus)
+#define _threefryNxWclass_tpl(NxW) \
+namespace r123{ \
+template<unsigned int ROUNDS> \
+ struct Threefry##NxW##_R{ \
+ typedef threefry##NxW##_ctr_t ctr_type; \
+ typedef threefry##NxW##_key_t key_type; \
+ typedef threefry##NxW##_key_t ukey_type; \
+ static const R123_METAL_CONSTANT_ADDRESS_SPACE unsigned int rounds=ROUNDS; \
+ inline R123_CUDA_DEVICE R123_FORCE_INLINE(ctr_type operator()(ctr_type ctr, key_type key)){ \
+ R123_STATIC_ASSERT(ROUNDS<=72, "threefry is only unrolled up to 72 rounds\n"); \
+ return threefry##NxW##_R(ROUNDS, ctr, key); \
+ } \
+}; \
+ typedef Threefry##NxW##_R<threefry##NxW##_rounds> Threefry##NxW; \
+} // namespace r123
+
+_threefryNxWclass_tpl(2x32)
+_threefryNxWclass_tpl(4x32)
+#if R123_USE_64BIT
+_threefryNxWclass_tpl(2x64)
+_threefryNxWclass_tpl(4x64)
+#endif
+
+/* The _tpl macros don't quite work to do string-pasting inside comments.
+ so we just write out the boilerplate documentation four times... */
+
+/**
+@defgroup ThreefryNxW Threefry Classes and Typedefs
+
+The ThreefryNxW classes export the member functions, typedefs and
+operator overloads required by a @ref CBRNG "CBRNG" class.
+
+As described in
+<a href="http://dl.acm.org/citation.cfm?doid=2063405"><i>Parallel Random Numbers: As Easy as 1, 2, 3</i> </a>,
+the Threefry family is closely related to the Threefish block cipher from
+<a href="http://www.skein-hash.info/"> Skein Hash Function</a>.
+Threefry is \b not suitable for cryptographic use.
+
+Threefry uses integer addition, bitwise rotation, xor and permutation of words to randomize its output.
+
+@class r123::Threefry2x32_R
+@ingroup ThreefryNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Threefry round
+function will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=13 or more for Threefry2x32.
+
+@typedef r123::Threefry2x32
+@ingroup ThreefryNxW
+ Threefry2x32 is equivalent to Threefry2x32_R<20>. With 20 rounds,
+ Threefry2x32 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+@class r123::Threefry2x64_R
+@ingroup ThreefryNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Threefry round
+function will be applied.
+
+In November 2011, the authors discovered that 13 rounds of
+Threefry2x64 sequenced by strided, interleaved key and counter
+increments failed a very long (longer than the default BigCrush
+length) WeightDistrub test. At the same time, it was confirmed that
+14 rounds passes much longer tests (up to 5x10^12 samples) of a
+similar nature. The authors know of no statistical flaws with
+ROUNDS=14 or more for Threefry2x64.
+
+@typedef r123::Threefry2x64
+@ingroup ThreefryNxW
+ Threefry2x64 is equivalent to Threefry2x64_R<20>. With 20 rounds,
+ Threefry2x64 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+
+
+@class r123::Threefry4x32_R
+@ingroup ThreefryNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Threefry round
+function will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=12 or more for Threefry4x32.
+
+@typedef r123::Threefry4x32
+@ingroup ThreefryNxW
+ Threefry4x32 is equivalent to Threefry4x32_R<20>. With 20 rounds,
+ Threefry4x32 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+
+
+
+@class r123::Threefry4x64_R
+@ingroup ThreefryNxW
+
+exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
+
+The template argument, ROUNDS, is the number of times the Threefry round
+function will be applied.
+
+As of September 2011, the authors know of no statistical flaws with
+ROUNDS=12 or more for Threefry4x64.
+
+@typedef r123::Threefry4x64
+@ingroup ThreefryNxW
+ Threefry4x64 is equivalent to Threefry4x64_R<20>. With 20 rounds,
+ Threefry4x64 has a considerable safety margin over the minimum number
+ of rounds with no known statistical flaws, but still has excellent
+ performance.
+*/
+
+#endif
+
+#endif
diff --git a/ext/random123/include/Random123/u01fixedpt.h b/ext/random123/include/Random123/u01fixedpt.h
new file mode 100644
index 0000000000000000000000000000000000000000..2058f8b57efcbb14a82f7d14def1066c24732dc9
--- /dev/null
+++ b/ext/random123/include/Random123/u01fixedpt.h
@@ -0,0 +1,200 @@
+/*
+Copyright 2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#ifndef _random123_ufixed01_dot_h_
+#define _random123_ufixed01_dot_h_
+
+#include "features/compilerfeatures.h"
+
+/** @defgroup u01fixedpt The u01fixedpt conversion functions
+
+ These functions convert unsigned W-bit integers to uniformly
+ spaced real values (float or double) between 0.0 and 1.0 with
+ mantissas of M bits.
+
+ PLEASE THINK CAREFULLY BEFORE USING THESE FUNCTIONS. THEY MAY
+ NOT BE WHAT YOU WANT. YOU MAY BE MUCH BETTER SERVED BY THE
+ FUNCTIONS IN ./uniform.hpp.
+
+ These functions produce a finite number *uniformly spaced* values
+ in the range from 0.0 to 1.0 with uniform probability. The price
+ of uniform spacing is that they may not utilize the entire space
+ of possible outputs. E.g., u01fixedpt_closed_open_32_24 will never
+ produce a non-zero value less than 2^-24, even though such values
+ are representable in single-precision floating point.
+
+ There are 12 functions, corresponding to the following choices:
+
+ - W = 32 or 64
+ - M = 24 (float) or 53 (double)
+ - open0 or closed0 : whether the output is open or closed at 0.0
+ - open1 or closed1 : whether the output is open or closed at 1.0
+
+ The W=64 M=24 cases are not implemented. To obtain an M=24 float
+ from a uint64_t, use a cast (possibly with right-shift and bitwise
+ and) to convert some of the bits of the uint64_t to a uint32_t and
+ then use u01fixedpt_x_y_32_float. Note that the 64-bit random integers
+ produced by the Random123 library are random in "all the bits", so
+ with a little extra effort you can obtain two floats this way --
+ one from the high bits and one from the low bits of the 64-bit
+ value.
+
+ If the output is open at one end, then the extreme
+ value (0.0 or 1.0) will never be returned. Conversely, if the output
+ is closed at one end, then the extreme value is a possible
+ return value.
+
+ The values returned are as follows. All values are returned
+ with equal frequency, except as noted in the closed_closed case:
+
+ closed_open: Let P=min(M,W)
+ there are 2^P possible output values:
+ {0, 1, 2, ..., 2^P-1}/2^P
+
+ open_closed: Let P=min(M,W)
+ there are 2^P possible values:
+ {1, 2, ..., 2^P}/2^P
+
+ open_open: Let P=min(M, W+1)
+ there are 2^(P-1) possible values:
+ {1, 3, 5, ..., 2^P-1}/2^P
+
+ closed_closed: Let P=min(M, W-1)
+ there are 1+2^P possible values:
+ {0, 1, 2, ... 2^P}/2^P
+ The extreme values (0.0 and 1.0) are
+ returned with half the frequency of
+ all others.
+
+ On x86 hardware, especially on 32bit machines, the use of
+ internal 80bit x87-style floating point may result in
+ 'bonus' precision, which may cause closed intervals to not
+ be really closed, i.e. the conversions below might not
+ convert UINT{32,64}_MAX to 1.0. This sort of issue is
+ likely to occur when storing the output of a u01fixedpt_*_32_float
+ function in a double, though one can imagine getting extra
+ precision artifacts when going from 64_53 as well. Other
+ artifacts may exist on some GPU hardware. The tests in
+ kat_u01_main.h try to expose such issues, but caveat emptor.
+
+ @cond HIDDEN_FROM_DOXYGEN
+ */
+
+/* Hex floats were standardized by C in 1999, but weren't standardized
+ by C++ until 2011. So, we're obliged to write out our constants in
+ decimal, even though they're most naturally expressed in binary.
+ We cross our fingers and hope that the compiler does the compile-time
+ constant arithmetic properly.
+*/
+#define R123_0x1p_31f (1.f/(1024.f*1024.f*1024.f*2.f))
+#define R123_0x1p_24f (128.f*R123_0x1p_31f)
+#define R123_0x1p_23f (256.f*R123_0x1p_31f)
+#define R123_0x1p_32 (1./(1024.*1024.*1024.*4.))
+#define R123_0x1p_63 (2.*R123_0x1p_32*R123_0x1p_32)
+#define R123_0x1p_53 (1024.*R123_0x1p_63)
+#define R123_0x1p_52 (2048.*R123_0x1p_63)
+
+/** @endcond */
+
+#ifndef R123_USE_U01_DOUBLE
+#define R123_USE_U01_DOUBLE 1
+#endif
+
+#ifdef __cplusplus
+extern "C"{
+#endif
+
+/* narrowing conversions: uint32_t to float */
+R123_CUDA_DEVICE R123_STATIC_INLINE float u01fixedpt_closed_closed_32_float(uint32_t i){
+ /* N.B. we ignore the high bit, so output is not monotonic */
+ return ((i&0x7fffffc0) + (i&0x40))*R123_0x1p_31f; /* 0x1.p-31f */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE float u01fixedpt_closed_open_32_float(uint32_t i){
+ return (i>>8)*R123_0x1p_24f; /* 0x1.0p-24f; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE float u01fixedpt_open_closed_32_float(uint32_t i){
+ return (1+(i>>8))*R123_0x1p_24f; /* *0x1.0p-24f; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE float u01fixedpt_open_open_32_float(uint32_t i){
+ return (0.5f+(i>>9))*R123_0x1p_23f; /* 0x1.p-23f; */
+}
+
+#if R123_USE_U01_DOUBLE
+/* narrowing conversions: uint64_t to double */
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_closed_closed_64_double(uint64_t i){
+ /* N.B. we ignore the high bit, so output is not monotonic */
+ return ((i&R123_64BIT(0x7ffffffffffffe00)) + (i&0x200))*R123_0x1p_63; /* 0x1.p-63; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_closed_open_64_double(uint64_t i){
+ return (i>>11)*R123_0x1p_53; /* 0x1.0p-53; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_open_closed_64_double(uint64_t i){
+ return (1+(i>>11))*R123_0x1p_53; /* 0x1.0p-53; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_open_open_64_double(uint64_t i){
+ return (0.5+(i>>12))*R123_0x1p_52; /* 0x1.0p-52; */
+}
+
+/* widening conversions: u32 to double */
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_closed_closed_32_double(uint32_t i){
+ /* j = i+(i&1) takes on 2^31+1 possible values with a 'trapezoid' distribution:
+ p_j = 1 0 2 0 2 .... 2 0 2 0 1
+ j = 0 1 2 3 4 .... 2^32
+ by converting to double *before* doing the add, we don't wrap the high bit.
+ */
+ return (((double)(i&1)) + i)*R123_0x1p_32; /* 0x1.p-32; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_closed_open_32_double(uint32_t i){
+ return i*R123_0x1p_32; /* 0x1.p-32; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_open_closed_32_double(uint32_t i){
+ return (1.+i)*R123_0x1p_32; /* 0x1.p-32; */
+}
+
+R123_CUDA_DEVICE R123_STATIC_INLINE double u01fixedpt_open_open_32_double(uint32_t i){
+ return (0.5+i)*R123_0x1p_32; /* 0x1.p-32; */
+}
+#endif /* R123_USE_U01_DOUBLE */
+
+#ifdef __cplusplus
+}
+#endif
+
+/** @} */
+#endif
diff --git a/ext/random123/include/Random123/uniform.hpp b/ext/random123/include/Random123/uniform.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..a815066ae8d302f7fba8c57c1feec56979fc5bd8
--- /dev/null
+++ b/ext/random123/include/Random123/uniform.hpp
@@ -0,0 +1,310 @@
+/*
+Copyright 2010-2011, D. E. Shaw Research.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+* Redistributions of source code must retain the above copyright
+ notice, this list of conditions, and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions, and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+* Neither the name of D. E. Shaw Research nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef __r123_uniform_dot_hpp
+#define __r123_uniform_dot_hpp
+
+/** @defgroup uniform Uniform distribution scalar conversion functions
+
+This file provides some simple functions that can be used to convert
+integers of various widths to floats and doubles with various
+characteristics. It can be used to generate real-valued, uniformly
+distributed random variables from the random integers produced by
+the Random123 CBRNGs.
+
+There are three templated functions:
+
+ - u01: output is as dense as possible in (0,1}, never 0.0. May
+ return 1.0 if and only if the number of output mantissa bits
+ is less than the width of the input.
+
+ - uneg11: output is as dense as possible in {-1,1}, never 0.0. May
+ return 1.0 or -1.0 if and only if the number of output mantissa bits
+ is less than the width of the input.
+
+ - u01fixedpt: output is "fixed point", equispaced, open at both ends,
+ and is never 0.0, 0.5 nor 1.0.
+
+The behavior of u01 and uneg11 depend on the pre-processor symbol:
+R123_UNIFORM_FLOAT_STORE. When #defined to a non-zero value, u01
+and uneg11 declare a volatile intermediate result, with the
+intention of forcing architectures that have "extra bits" in their
+floating point registers to more closely conform to IEEE
+arithmetic. When compiled this way, u01 and uneg11 will be
+significantly slower, as they will incur a memory write and read on
+every call. Without it, they may fail the "known answer test"
+implemented in ut_uniform_IEEEkat.cpp even though they perform
+perfectly reasonable int to float conversions. We have used
+this option to get 32-bit x86 to produce the same results as
+64-bit x86-64 code, but we do not recommend it for normal
+use.
+
+Three additional functions are defined when C++11 or newer is in use:
+
+ - u01all
+ - uneg11all
+ - u01fixedptall
+
+These functions apply the corresponding conversion to every
+element of their argument, which must be a staticly sized
+array, e.g., an r123array or a std::array of an integer type.
+
+This file may not be as portable, and has not been tested as
+rigorously as other files in the library, e.g., the generators.
+Nevertheless, we hope it is useful and we encourage developers to
+copy it and modify it for their own use. We invite comments and
+improvements.
+*/
+
+#include <Random123/features/compilerfeatures.h>
+#include <limits>
+#if R123_USE_CXX11_TYPE_TRAITS
+#include <type_traits>
+#endif
+#if __cplusplus >= 201103L
+#include <array>
+#endif
+
+namespace r123{
+/**
+@{
+@cond HIDDEN_FROM_DOXYGEN
+*/
+
+#if R123_USE_CXX11_TYPE_TRAITS
+using std::make_signed;
+using std::make_unsigned;
+#else
+// Sigh... We could try to find another <type_traits>, e.g., from
+// boost or TR1. Or we can do it ourselves in the r123 namespace.
+// It's not clear which will cause less headache...
+template <typename T> struct make_signed{};
+template <typename T> struct make_unsigned{};
+#define R123_MK_SIGNED_UNSIGNED(ST, UT) \
+template<> struct make_signed<ST>{ typedef ST type; }; \
+template<> struct make_signed<UT>{ typedef ST type; }; \
+template<> struct make_unsigned<ST>{ typedef UT type; }; \
+template<> struct make_unsigned<UT>{ typedef UT type; }
+
+R123_MK_SIGNED_UNSIGNED(int8_t, uint8_t);
+R123_MK_SIGNED_UNSIGNED(int16_t, uint16_t);
+R123_MK_SIGNED_UNSIGNED(int32_t, uint32_t);
+R123_MK_SIGNED_UNSIGNED(int64_t, uint64_t);
+#if R123_USE_GNU_UINT128
+R123_MK_SIGNED_UNSIGNED(__int128_t, __uint128_t);
+#endif
+#undef R123_MK_SIGNED_UNSIGNED
+#endif
+
+#if defined(__CUDACC__) || defined(_LIBCPP_HAS_NO_CONSTEXPR)
+// Amazing! cuda thinks numeric_limits::max() is a __host__ function, so
+// we can't use it in a device function.
+//
+// The LIBCPP_HAS_NO_CONSTEXP test catches situations where the libc++
+// library thinks that the compiler doesn't support constexpr, but we
+// think it does. As a consequence, the library declares
+// numeric_limits::max without constexpr. This workaround should only
+// affect a narrow range of compiler/library pairings.
+//
+// In both cases, we find max() by computing ~(unsigned)0 right-shifted
+// by is_signed.
+template <typename T>
+R123_CONSTEXPR R123_STATIC_INLINE R123_CUDA_DEVICE T maxTvalue(){
+ typedef typename make_unsigned<T>::type uT;
+ return (~uT(0)) >> std::numeric_limits<T>::is_signed;
+ }
+#else
+template <typename T>
+R123_CONSTEXPR R123_STATIC_INLINE T maxTvalue(){
+ return std::numeric_limits<T>::max();
+}
+#endif
+/** @endcond
+ @}
+ */
+
+//! Return a uniform real value in (0, 1]
+/**
+ @ingroup uniform
+ Input is a W-bit integer (signed or unsigned). It is cast to
+ a W-bit unsigned integer, multiplied by Ftype(2^-W) and added to
+ Ftype(2^(-W-1)). A good compiler should optimize it down to an
+ int-to-float conversion followed by a multiply and an add, which
+ might be fused, depending on the architecture.
+
+ If the input is a uniformly distributed integer, and if Ftype
+ arithmetic follows IEEE754 round-to-nearest rules, then the
+ result is a uniformly distributed floating point number in (0, 1].
+
+- The result is never exactly 0.0.
+- The smallest value returned is 2^-(W-1).
+- Let M be the number of mantissa bits in Ftype (typically 24 or 53).
+ - If W>M then the largest value retured is 1.0.
+ - If W<=M then the largest value returned is Ftype(1.0 - 2^(-W-1)).
+*/
+template <typename Ftype, typename Itype>
+R123_CUDA_DEVICE R123_STATIC_INLINE Ftype u01(Itype in){
+ typedef typename make_unsigned<Itype>::type Utype;
+ R123_CONSTEXPR Ftype factor = Ftype(1.)/(maxTvalue<Utype>() + Ftype(1.));
+ R123_CONSTEXPR Ftype halffactor = Ftype(0.5)*factor;
+#if R123_UNIFORM_FLOAT_STORE
+ volatile Ftype x = Utype(in)*factor; return x+halffactor;
+#else
+ return Utype(in)*factor + halffactor;
+#endif
+}
+
+//! Return a signed value in [-1,1]
+/**
+ @ingroup uniform
+ The argument is converted to a W-bit signed integer, multiplied by Ftype(2^-(W-1)) and
+ then added to Ftype(2^-W). A good compiler should optimize
+ it down to an int-to-float conversion followed by a multiply and
+ an add, which might be fused, depending on the architecture.
+
+ If the input is a uniformly distributed integer, and if Ftype
+ arithmetic follows IEEE754 round-to-nearest rules, then the
+ output is a uniformly distributed floating point number in [-1, 1].
+
+- The result is never exactly 0.0.
+- The smallest absolute value returned is 2^-W
+- Let M be the number of mantissa bits in Ftype.
+ - If W>M then the largest value retured is 1.0 and the smallest is -1.0.
+ - If W<=M then the largest value returned is the Ftype(1.0 - 2^-W)
+ and the smallest value returned is -Ftype(1.0 - 2^-W).
+*/
+template <typename Ftype, typename Itype>
+R123_CUDA_DEVICE R123_STATIC_INLINE Ftype uneg11(Itype in){
+ typedef typename make_signed<Itype>::type Stype;
+ R123_CONSTEXPR Ftype factor = Ftype(1.)/(maxTvalue<Stype>() + Ftype(1.));
+ R123_CONSTEXPR Ftype halffactor = Ftype(0.5)*factor;
+#if R123_UNIFORM_FLOAT_STORE
+ volatile Ftype x = Stype(in)*factor; return x+halffactor;
+#else
+ return Stype(in)*factor + halffactor;
+#endif
+}
+
+//! Return a value in (0,1) chosen from a set of equally spaced fixed-point values
+/**
+ @ingroup uniform
+ Let:
+ - W = width of Itype, e.g., 32 or 64, regardless of signedness.
+ - M = mantissa bits of Ftype, e.g., 24, 53 or 64
+ - B = min(M, W)
+
+ Then the 2^(B-1) possible output values are: 2^-B*{1, 3, 5, ..., 2^B - 1}
+
+ The smallest output is: 2^-B
+
+ The largest output is: 1 - 2^-B
+
+ The output is never exactly 0.0, nor 0.5, nor 1.0.
+
+ The 2^(B-1) possible outputs:
+ - are equally likely,
+ - are uniformly spaced by 2^-(B-1),
+ - are balanced around 0.5
+*/
+template <typename Ftype, typename Itype>
+R123_CUDA_DEVICE R123_STATIC_INLINE Ftype u01fixedpt(Itype in){
+ typedef typename make_unsigned<Itype>::type Utype;
+ R123_CONSTEXPR int excess = std::numeric_limits<Utype>::digits - std::numeric_limits<Ftype>::digits;
+ if(excess>=0){
+ R123_CONSTEXPR int ex_nowarn = (excess>=0) ? excess : 0;
+ R123_CONSTEXPR Ftype factor = Ftype(1.)/(Ftype(1.) + ((maxTvalue<Utype>()>>ex_nowarn)));
+ return (1 | (Utype(in)>>ex_nowarn)) * factor;
+ }else
+ return u01<Ftype>(in);
+}
+
+#if R123_USE_CXX11_STD_ARRAY
+
+//! Apply u01 to every item in an r123array, returning a std::array
+/** @ingroup uniform
+ * Only in C++11 and newer.
+ * The argument type may be any integer collection with a constexpr static_size member,
+ * e.g., an r123array or a std::array of an integer type.
+ */
+template <typename Ftype, typename CollType>
+static inline
+std::array<Ftype, CollType::static_size> u01all(CollType in)
+{
+ std::array<Ftype, CollType::static_size> ret;
+ size_t i=0;
+ for(auto e : in){
+ ret[i++] = u01<Ftype>(e);
+ }
+ return ret;
+}
+
+//! Apply uneg11 to every item in an r123array, returning a std::array
+/** @ingroup uniform
+ * Only in C++11 and newer.
+ * The argument type may be any integer collection with a constexpr static_size member,
+ * e.g., an r123array or a std::array of an integer type.
+ */
+template <typename Ftype, typename CollType>
+static inline
+std::array<Ftype, CollType::static_size> uneg11all(CollType in)
+{
+ std::array<Ftype, CollType::static_size> ret;
+ size_t i=0;
+ for(auto e : in){
+ ret[i++] = uneg11<Ftype>(e);
+ }
+ return ret;
+}
+
+//! Apply u01fixedpt to every item in an r123array, returning a std::array
+/** @ingroup uniform
+ * Only in C++11 and newer.
+ * The argument type may be any integer collection with a constexpr static_size member,
+ * e.g., an r123array or a std::array of an integer type.
+*/
+template <typename Ftype, typename CollType>
+static inline
+std::array<Ftype, CollType::static_size> u01fixedptall(CollType in)
+{
+ std::array<Ftype, CollType::static_size> ret;
+ size_t i=0;
+ for(auto e : in){
+ ret[i++] = u01fixedpt<Ftype>(e);
+ }
+ return ret;
+}
+#endif // __cplusplus >= 201103L
+
+} // namespace r123
+
+#endif
+
diff --git a/test/unit/test_morph_expr.cpp b/test/unit/test_morph_expr.cpp
index 60d104e992458d8b64e0ad760c0f0e17b4d3317f..5a06a27586ea7c5e24b260f12830b9b3ea21ee39 100644
--- a/test/unit/test_morph_expr.cpp
+++ b/test/unit/test_morph_expr.cpp
@@ -24,6 +24,22 @@ namespace arb {
}
}
+::testing::AssertionResult mlocation_eq(mlocation a, mlocation b) {
+ if (a.branch!=b.branch) {
+ return ::testing::AssertionFailure()
+ << "cables " << a << " and " << b << " differ";
+ }
+
+ using FP = testing::internal::FloatingPoint<double>;
+ if (FP(a.pos).AlmostEquals(FP(b.pos))) {
+ return ::testing::AssertionSuccess();
+ }
+ else {
+ return ::testing::AssertionFailure()
+ << "mlocations " << a << " and " << b << " differ";
+ }
+}
+
::testing::AssertionResult cable_eq(mcable a, mcable b) {
if (a.branch!=b.branch) {
return ::testing::AssertionFailure()
@@ -54,6 +70,20 @@ namespace arb {
return ::testing::AssertionSuccess();
}
+::testing::AssertionResult mloctionlist_eq(const mlocation_list& as, const mlocation_list& bs) {
+ if (as.size()!=bs.size()) {
+ return ::testing::AssertionFailure()
+ << "cablelists " << as << " and " << bs << " differ";
+ }
+
+ for (auto i: util::count_along(as)) {
+ auto result = mlocation_eq(as[i], bs[i]);
+ if (!result) return ::testing::AssertionFailure()
+ << "mlocation lists " << as << " and " << bs << " differ";
+ }
+ return ::testing::AssertionSuccess();
+}
+
TEST(region, expr_repn) {
using util::to_string;
@@ -268,9 +298,91 @@ TEST(locset, thingify) {
// In the absence of a spherical root, there is no branch 4.
EXPECT_THROW(thingify(begb4, mp), no_such_branch);
}
+ {
+ mprovider mp(morphology(sm, false));
+
+ auto all = reg::all();
+ auto ls0 = thingify(ls::uniform(all, 0, 9, 12), mp);
+ auto ls1 = thingify(ls::uniform(all, 0, 9, 12), mp);
+ auto ls2 = thingify(ls::uniform(all, 10, 19, 12), mp);
+ auto ls3 = thingify(ls::uniform(all, 0, 9, 13), mp);
+ auto ls4 = thingify(ls::uniform(all, 5, 6, 12), mp);
+ auto ls5 = thingify(ls::uniform(all, 2, 5, 12), mp);
+ auto ls6 = thingify(ls::uniform(all, 5, 11, 12), mp);
+
+ EXPECT_EQ(ls0, ls1);
+
+ bool found_none = true;
+ for (auto l: ls2) {
+ auto it = std::find(ls0.begin(), ls0.end(), l);
+ if (it != ls0.end()) {
+ found_none = false;
+ }
+ }
+ EXPECT_TRUE(found_none);
+
+ found_none = true;
+ for (auto l: ls3) {
+ auto it = std::find(ls0.begin(), ls0.end(), l);
+ if (it != ls0.end()) {
+ found_none = false;
+ }
+ }
+ EXPECT_TRUE(found_none);
+
+ bool found_all = true;
+ for (auto l: ls4) {
+ auto it = std::find(ls0.begin(), ls0.end(), l);
+ if (it == ls0.end()) {
+ found_all = false;
+ }
+ }
+ EXPECT_TRUE(found_all);
+
+ int found = 0;
+ for (auto l: ls5) {
+ auto it = std::find(ls4.begin(), ls4.end(), l);
+ if (it != ls4.end()) found++;
+ }
+ EXPECT_TRUE(found == 1);
+
+ found = 0;
+ for (auto l: ls6) {
+ auto it = std::find(ls4.begin(), ls4.end(), l);
+ if (it != ls4.end()) found++;
+ }
+ EXPECT_TRUE(found == 2);
+ }
+ {
+ mprovider mp(morphology(sm, false));
+ auto sub_reg = join(reg::cable(0, 0.2, 0.7), reg::cable(1, 0.1, 1), reg::cable(3, 0.5, 0.6));
+
+ auto ls0 = thingify(ls::uniform(sub_reg, 0, 10000, 72), mp);
+ for (auto l: ls0) {
+ switch(l.branch) {
+ case 0: {
+ if (l.pos < 0.2 || l.pos > 0.7) FAIL();
+ break;
+ }
+ case 1: {
+ if (l.pos < 0.1 || l.pos > 1) FAIL();
+ break;
+ }
+ case 3: {
+ if (l.pos < 0.5 || l.pos > 0.6) FAIL();
+ break;
+ }
+ default: {
+ FAIL();
+ break;
+ }
+ }
+ SUCCEED();
+ }
+ }
}
-TEST(region, thingify) {
+TEST(region, thingify_simple_morphologies) {
using pvec = std::vector<msize_t>;
using svec = std::vector<msample>;
using cl = mcable_list;
@@ -352,16 +464,45 @@ TEST(region, thingify) {
sample_tree sm(samples, parents);
mprovider mp(morphology(sm, true));
+ using ls::location;
using reg::tagged;
+ using reg::distal_interval;
+ using reg::proximal_interval;
using reg::branch;
+ using reg::cable;
using reg::all;
+ locset mid0_ = location(0,0.5);
+ locset start1_ = location(1,0);
+ locset end1_ = location(1,1);
+
+ auto reg0_ = distal_interval(start1_, 45);
+ auto reg1_ = distal_interval(mid0_, 74);
+ auto reg2_ = proximal_interval(end1_, 45);
+ auto reg3_ = proximal_interval(end1_, 91);
+ auto reg4_ = distal_interval(end1_, 0);
+ auto reg5_ = distal_interval(start1_, 0);
+ auto reg6_ = proximal_interval(start1_, 0);
+
EXPECT_EQ(thingify(tagged(1), mp), (mcable_list{{0,0,1}}));
EXPECT_EQ(thingify(tagged(2), mp), (mcable_list{{2,0,1}}));
EXPECT_EQ(thingify(tagged(3), mp), (mcable_list{{1,0,1}}));
EXPECT_EQ(thingify(join(tagged(1), tagged(2), tagged(3)), mp), (mcable_list{{0,0,1}, {1,0,1}, {2,0,1}}));
EXPECT_EQ(thingify(join(tagged(1), tagged(2), tagged(3)), mp), thingify(all(), mp));
+ EXPECT_EQ(thingify(reg0_, mp), (mcable_list{{1,0,0.5}}));
+ EXPECT_EQ(thingify(reg1_, mp), (mcable_list{{0,0.5,1}, {1,0,0.8}, {2,0,0.8}}));
+ EXPECT_EQ(thingify(reg2_, mp), (mcable_list{{1,0.5,1}}));
+ EXPECT_EQ(thingify(reg3_, mp), (mcable_list{{0, 0.75, 1}, {1,0,1}}));
+ EXPECT_EQ(thingify(reg4_, mp), (mcable_list{{1,1,1}}));
+ EXPECT_EQ(thingify(reg5_, mp), (mcable_list{{0,1,1}}));
+ EXPECT_EQ(thingify(reg6_, mp), (mcable_list{{0,1,1}}));
}
+}
+
+TEST(region, thingify_moderate_morphologies) {
+ using pvec = std::vector<msize_t>;
+ using svec = std::vector<msample>;
+ using cl = mcable_list;
// Test multi-level morphologies.
//
@@ -382,21 +523,28 @@ TEST(region, thingify) {
{
pvec parents = {mnpos, 0, 1, 0, 3, 4, 4, 6};
svec samples = {
- {{ 0, 0, 0, 2}, 1},
- {{ 10, 0, 0, 2}, 3},
- {{100, 0, 0, 2}, 3},
- {{ 0, 10, 0, 2}, 2},
- {{ 0,100, 0, 2}, 2},
+ {{ 0, 0, 0, 1}, 1},
+ {{ 10, 0, 0, 1}, 3},
+ {{100, 0, 0, 3}, 3},
+ {{ 0, 10, 0, 1}, 2},
+ {{ 0,100, 0, 5}, 2},
{{100,100, 0, 2}, 4},
- {{ 0,200, 0, 2}, 3},
- {{ 0,300, 0, 2}, 3},
+ {{ 0,200, 0, 1}, 3},
+ {{ 0,300, 0, 3}, 3},
};
sample_tree sm(samples, parents);
// Without spherical root
mprovider mp(morphology(sm, false));
+ using ls::location;
using reg::tagged;
+ using reg::distal_interval;
+ using reg::proximal_interval;
+ using reg::radius_lt;
+ using reg::radius_le;
+ using reg::radius_gt;
+ using reg::radius_ge;
using reg::branch;
using reg::all;
using reg::cable;
@@ -419,11 +567,10 @@ TEST(region, thingify) {
mcable b3_{3,0,1};
cl all_ = {b0_,b1_,b2_,b3_};
- mcable end1_{1,1,1};
- mcable root_{0,0,0};
+ mcable c_end1_{1,1,1};
+ mcable c_root_{0,0,0};
EXPECT_EQ(thingify(all(), mp), all_);
- EXPECT_EQ(thingify(soma, mp), empty_);
EXPECT_EQ(thingify(axon, mp), (cl{b1_}));
EXPECT_EQ(thingify(dend, mp), (cl{b0_,b3_}));
EXPECT_EQ(thingify(apic, mp), (cl{b2_}));
@@ -432,9 +579,60 @@ TEST(region, thingify) {
// Test that intersection correctly generates zero-length cables at
// parent-child interfaces.
- EXPECT_EQ(thingify(intersect(apic, dend), mp), (cl{end1_}));
- EXPECT_EQ(thingify(intersect(apic, axon), mp), (cl{end1_}));
- EXPECT_EQ(thingify(intersect(axon, dend), mp), (cl{root_, end1_}));
+ EXPECT_EQ(thingify(intersect(apic, dend), mp), (cl{c_end1_}));
+ EXPECT_EQ(thingify(intersect(apic, axon), mp), (cl{c_end1_}));
+ EXPECT_EQ(thingify(intersect(axon, dend), mp), (cl{c_root_, c_end1_}));
+
+ // Test distal and proximal interavls
+ auto start0_ = location(0, 0 );
+ auto quar_1_ = location(1, 0.25);
+ auto mid1_ = location(1, 0.5 );
+ auto end1_ = location(1, 1 );
+ auto mid2_ = location(2, 0.5 );
+ auto end2_ = location(2, 1 );
+ auto mid3_ = location(3, 0.5 );
+ auto loc_3_0_ = location(3, 0.4 );
+ auto loc_3_1_ = location(3, 0.65);
+ auto mid_3_ = location(3, 0.5 );
+ auto reg_a_ = join(cable(0,0.1,0.4), cable(2,0,1), cable(3,0.1,0.4));
+ auto reg_b_ = join(cable(0,0.1,0.4), cable(2,0,1), cable(3,0.1,0.3));
+ auto reg_c_ = join(cable(0,0,0.7), cable(2,0,0.5), cable(3,0.1,0.4), cable(3,0.9,1));
+ auto reg_d_ = join(cable(0,0,0.7), cable(2,0,0.5), cable(3,0.1,0.9));
+
+ // Distal from point and/or interval
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(start0_, 1000), mp), (mcable_list{{0,0,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(quar_1_, 150), mp), (mcable_list{{1,0.25,1}, {2,0,0.75}, {3,0,0.375}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(mid1_, 1000), mp), (mcable_list{{1,0.5,1}, {2,0,1}, {3,0,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(mid1_, 150), mp), (mcable_list{{1,0.5,1}, {2,0,1}, {3,0,0.5}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(end1_, 100), mp), (mcable_list{{2,0,1},{3,0,0.5}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(join(quar_1_, mid1_), 150), mp), (mcable_list{{1,0.25,1}, {2,0,1}, {3,0,0.5}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(join(quar_1_, loc_3_1_), 150), mp), (mcable_list{{1,0.25,1}, {2,0,0.75}, {3,0,0.375}, {3,0.65,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(distal_interval(join(quar_1_, loc_3_1_), 150), mp), (mcable_list{{1,0.25,1}, {2,0,0.75}, {3,0,0.375}, {3,0.65,1}})));
+
+ // Proximal from point and/or interval
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(mid3_, 100), mp), (mcable_list{{3,0,0.5}})));
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(mid3_, 150), mp), (mcable_list{{1,0.5,1}, {3,0,0.5}})));
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(end2_, 150), mp), (mcable_list{{1,0.5,1}, {2,0,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(end2_, 500), mp), (mcable_list{{1,0,1}, {2,0,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(loc_3_0_, 100), mp), (mcable_list{{1,0.8,1}, {3,0,0.4}})));
+ EXPECT_TRUE(cablelist_eq(thingify(proximal_interval(join(loc_3_0_, mid2_), 120), mp), (mcable_list{{1,0.3,1}, {2,0,0.5}, {3, 0, 0.4}})));
+
+ // Test radius_lt and radius_gt
+ EXPECT_TRUE(cablelist_eq(thingify(radius_lt(all(), 2), mp), (mcable_list{{0,0,0.55}, {1,0,0.325}, {3,0.375,0.75}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_lt(all(), 3), mp), (mcable_list{{0,0,1}, {1,0,0.55}, {2,6.0/9.0,1}, {3,0.25,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_gt(all(), 2), mp), (mcable_list{{0,0.55,1}, {1,0.325,1}, {2,0,1}, {3,0,0.375}, {3,0.75,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_gt(all(), 3), mp), (mcable_list{{1,0.55,1}, {2,0,6.0/9.0}, {3,0,0.25}})));
+
+ EXPECT_TRUE(cablelist_eq(thingify(radius_le(all(), 2), mp), (mcable_list{{0,0,0.55}, {1,0,0.325}, {2,1,1}, {3,0.375,0.75}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_le(all(), 3), mp), (mcable_list{{0,0,1}, {1,0,0.55}, {2,6.0/9.0,1}, {3,0.25,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_ge(all(), 2), mp), (mcable_list{{0,0.55,1}, {1,0.325,1}, {2,0,1}, {3,0,0.375}, {3,0.75,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_ge(all(), 3), mp), (mcable_list{{1,0.55,1}, {2,0,6.0/9.0}, {3,0,0.25}})));
+
+ EXPECT_TRUE(cablelist_eq(thingify(radius_lt(reg_a_, 2), mp), (mcable_list{{0,0.1,0.4},{3,0.375,0.4}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_gt(reg_a_, 2), mp), (mcable_list{{2,0,1},{3,0.1,0.375}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_lt(reg_b_, 2), mp), (mcable_list{{0,0.1,0.4}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_gt(reg_c_, 2), mp), (mcable_list{{0,0.55,0.7},{2,0,0.5},{3,0.1,0.375},{3,0.9,1}})));
+ EXPECT_TRUE(cablelist_eq(thingify(radius_gt(reg_d_, 2), mp), (mcable_list{{0,0.55,0.7},{2,0,0.5},{3,0.1,0.375},{3,0.75,0.9}})));
// Test some more interesting intersections and unions.
@@ -570,6 +768,11 @@ TEST(region, thingify) {
EXPECT_EQ(thingify(join(lhs, rhs), mp), ror);
}
+}
+TEST(region, thingify_complex_morphologies) {
+ using pvec = std::vector<msize_t>;
+ using svec = std::vector<msample>;
+ using cl = mcable_list;
{
pvec parents = {mnpos, 0, 1, 0, 3, 4, 5, 5, 7, 7, 4, 10};
svec samples = {
@@ -588,8 +791,6 @@ TEST(region, thingify) {
};
sample_tree sm(samples, parents);
auto m = morphology(sm, false);
- std::cout << m.branch_parent(7);
-
{
auto in = cl{{0,0,0},{1,0,0.5},{1,1,1},{2,0,1},{2,1,1},{3,1,1},{4,0,1},{5,1,1},{7,0,1}};
auto out = reg::remove_covered_points(in, m);
@@ -604,5 +805,104 @@ TEST(region, thingify) {
auto expected = cl{{1,0,0.5},{3,1,1},{4,0,1},{5,1,1},{7,0,1}};
EXPECT_TRUE(cablelist_eq(out, expected));
}
+ {
+ mprovider mp(m);
+ using reg::cable;
+ using ls::most_distal;
+ using ls::most_proximal;
+
+ auto reg_a_ = join(cable(0,0.1,0.4), cable(0,0,0.9), cable(1,0.1,0.4));
+ auto reg_b_ = join(cable(0,0.1,0.4), cable(0,0,0.9), cable(1,0.1,0.4), cable(1,0.2,0.5));
+ auto reg_c_ = join(cable(0,0.1,0.4), cable(0,0,0.9), cable(1,0.1,0.4), cable(2,0.2,0.5));
+ auto reg_d_ = join(cable(2,0,0.9), cable(3,0.1,0.1), cable(4,0.1,0.6));
+ auto reg_e_ = join(cable(2,0,0.9), cable(4,0.1,0.1), cable(5,0.1,0.6));
+ auto reg_f_ = join(cable(7,0,1), cable(2,0,0.9), cable(4,0.1,0.1), cable(5,0.1,0.6));
+
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_a_), mp), mlocation_list{{0,0.9},{1,0.4}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_b_), mp), mlocation_list{{0,0.9},{1,0.5}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_c_), mp), mlocation_list{{0,0.9},{2,0.5}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_d_), mp), mlocation_list{{3,0.1},{4,0.6}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_e_), mp), mlocation_list{{5,0.6}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_distal(reg_f_), mp), mlocation_list{{5,0.6},{7,1}}));
+
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_a_), mp), mlocation_list{{0,0}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_b_), mp), mlocation_list{{0,0}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_c_), mp), mlocation_list{{0,0}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_d_), mp), mlocation_list{{2,0}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_e_), mp), mlocation_list{{2,0}}));
+ EXPECT_TRUE(mloctionlist_eq(thingify(most_proximal(reg_f_), mp), mlocation_list{{2,0}}));
+ }
+ }
+ {
+ pvec parents = {mnpos, 0, 1, 1, 2, 3, 0, 6, 7, 8, 7};
+ svec samples = {
+ {{ 0, 10, 10, 1}, 1},
+ {{ 0, 30, 30, 1}, 2},
+ {{ 0, 60,-20, 1}, 2},
+ {{ 0, 90, 70, 1}, 2},
+ {{ 0, 80,-10, 1}, 2},
+ {{ 0,100,-40, 1}, 2},
+ {{ 0,-50,-50, 1}, 2},
+ {{ 0, 20,-30, 2}, 2},
+ {{ 0, 40,-80, 2}, 2},
+ {{ 0,-30,-80, 3}, 2},
+ {{ 0, 90,-70, 5}, 2}
+ };
+ sample_tree sm(samples, parents);
+
+ // Without spherical root
+ mprovider mp(morphology(sm, false));
+
+ using reg::all;
+ using reg::z_dist_from_soma_lt;
+ using reg::z_dist_from_soma_le;
+ using reg::z_dist_from_soma_gt;
+ using reg::z_dist_from_soma_ge;
+ using reg::cable;
+
+ // Test projection
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_lt(0), mp), (mcable_list{})));
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_ge(0), mp), thingify(all(), mp)));
+
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_le(100), mp), thingify(all(), mp)));
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_gt(100), mp), (mcable_list{})));
+
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_le(90), mp), thingify(all(), mp)));
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_gt(90), mp), (mcable_list{})));
+
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_lt(20), mp),
+ (mcable_list{{0,0,1},
+ {1,0,0.578250901781922829},
+ {2,0.61499300915417734997,0.8349970039232188642},
+ {3,0,0.179407353580315756}
+ })));
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_ge(20), mp),
+ (mcable_list{{0,1,1},
+ {1,0.578250901781922829,1},
+ {2,0,0.61499300915417734997},
+ {2,0.8349970039232188642,1},
+ {3,0.179407353580315756,1},
+ {4,0,1},
+ {5,0,1}
+ })));
+ EXPECT_TRUE(cablelist_eq(thingify(join(z_dist_from_soma_lt(20), z_dist_from_soma_ge(20)), mp), thingify(all(), mp)));
+
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_le(50), mp),
+ (mcable_list{{0,0,1},
+ {1,0,1},
+ {2,0,0.2962417607888518767},
+ {2,0.4499900130773962142,1},
+ {3,0,0.4485183839507893905},
+ {3,0.7691110303704736343,1},
+ {4,0,0.0869615364994152821},
+ {5,0,0.25}
+ })));
+ EXPECT_TRUE(cablelist_eq(thingify(z_dist_from_soma_gt(50), mp),
+ (mcable_list{{2,0.2962417607888518767,0.4499900130773962142},
+ {3,0.4485183839507893905,0.7691110303704736343},
+ {4,0.0869615364994152821,1},
+ {5,0.25,1}})));
+
+ EXPECT_TRUE(cablelist_eq(thingify(join(z_dist_from_soma_le(50), z_dist_from_soma_gt(50)), mp), thingify(all(), mp)));
}
}