sequenzo 0.1.17__cp39-cp39-win_amd64.whl → 0.1.18__cp39-cp39-win_amd64.whl

sequenzo/dissimilarity_measures/src/OMspellDistance.cpp

@@ -3,11 +3,12 @@
 #include <vector>
 #include <cmath>
 #include <iostream>
-#include <xsimd/xsimd.hpp>
 #include "utils.h"
+#include "dp_utils.h"
 #ifdef _OPENMP
     #include <omp.h>
 #endif
+#include <xsimd/xsimd.hpp>
 
 namespace py = pybind11;
 
@@ -85,83 +86,119 @@ public:
         }
     }
 
-    double getIndel(int i, int j, int state) {
+    // 对齐分配函数 moved to dp_utils.h
+
+    double getIndel(int i, int j, int state){
         auto ptr_indel = indellist.mutable_unchecked<1>();
         auto ptr_dur = seqdur.mutable_unchecked<2>();
 
-        xsimd::batch<double, xsimd::default_arch> state_vec(ptr_indel(state));
-        xsimd::batch<double, xsimd::default_arch> timecost_vec(timecost);
-
-        xsimd::batch<double, xsimd::default_arch> dur_vec(ptr_dur(i, j));
-        xsimd::batch<double, xsimd::default_arch> result = state_vec + timecost_vec * dur_vec;
-
-        return result.get(0);
+        return ptr_indel(state) + timecost * ptr_dur(i, j);
     }
 
-
-    double getSubCost(int i_state, int j_state, int i_x, int i_y, int j_x, int j_y) {
+    double getSubCost(int i_state, int j_state, int i_x, int i_y, int j_x, int j_y){
         auto ptr_dur = seqdur.mutable_unchecked<2>();
 
-        if (i_state == j_state) {
+        if(i_state == j_state){
             double diffdur = ptr_dur(i_x, i_y) - ptr_dur(j_x, j_y);
-            return std::abs(timecost * diffdur);
-        } else {
-            auto ptr_sm = sm.mutable_unchecked<2>();
 
-            double d1 = ptr_dur(i_x, i_y);
-            double d2 = ptr_dur(j_x, j_y);
-
-            xsimd::batch<double, xsimd::default_arch> d1_vec = xsimd::batch<double, xsimd::default_arch>::broadcast(d1);
-            xsimd::batch<double, xsimd::default_arch> d2_vec = xsimd::batch<double, xsimd::default_arch>::broadcast(d2);
-            xsimd::batch<double, xsimd::default_arch> cost = xsimd::batch<double, xsimd::default_arch>::broadcast(timecost);
-            xsimd::batch<double, xsimd::default_arch> sum = (d1_vec + d2_vec) * cost;
+            return abs(timecost * diffdur);
+        }else{
+            auto ptr_sm = sm.mutable_unchecked<2>();
 
-            return ptr_sm(i_state, j_state) + sum.get(0);
+            return ptr_sm(i_state, j_state) +
+                    (ptr_dur(i_x, i_y) + ptr_dur(j_x, j_y)) * timecost;
         }
     }
 
-
-    double compute_distance(int is, int js) {
+    double compute_distance(int is, int js, double* prev, double* curr) {
         try {
             auto ptr_seq = sequences.unchecked<2>();
             auto ptr_len = seqlength.unchecked<1>();
+            auto ptr_sm = sm.unchecked<2>();
+            auto ptr_dur = seqdur.unchecked<2>();
+            auto ptr_indel = indellist.unchecked<1>();
 
             int i_state = 0, j_state = 0;
-            double maxpossiblecost;
             int mm = ptr_len(is);
             int nn = ptr_len(js);
-            int mSuf = mm + 1, nSuf = nn + 1;
-
-            std::vector<double> prev(fmatsize, 0.0);
-            std::vector<double> curr(fmatsize, 0.0);
+            int mSuf = mm + 1;
+            int nSuf = nn + 1;
 
             prev[0] = 0;
             curr[0] = 0;
 
-            for (int ii = 1; ii < nSuf; ii++) {
-                j_state = ptr_seq(js, ii - 1);
-                prev[ii] = prev[ii-1] + getIndel(js, ii-1, j_state);
+            // initialize first row: cumulative insertions into js along columns
+            for (int jj = 1; jj < nSuf; jj++) {
+                int bj = ptr_seq(js, jj - 1);
+                prev[jj] = prev[jj - 1] + (ptr_indel(bj) + timecost * ptr_dur(js, jj - 1));
             }
 
+            using batch_t = xsimd::batch<double>;
+            constexpr std::size_t B = batch_t::size;
+
             for (int i = 1; i < mSuf; i++) {
                 i_state = ptr_seq(is, i - 1);
-                curr[0] = prev[0] + getIndel(is, i - 1, i_state);
-
-                for (int j = 1; j < nSuf; j++) {
-                    j_state = ptr_seq(js, j - 1);
+                // per-row deletion cost (depends only on i_state and i position)
+                double dur_i = ptr_dur(is, i - 1);
+                double del_cost_i = ptr_indel(i_state) + timecost * dur_i;
+
+                // first column: cumulative deletions D[i][0] = D[i-1][0] + del_cost_i
+                curr[0] = prev[0] + del_cost_i;
+
+                int j = 1;
+                for (; j + (int)B <= nSuf; j += (int)B) {
+                    const double* prev_ptr = prev + j;
+                    const double* prevm1_ptr = prev + (j - 1);
+
+                    batch_t prevj = batch_t::load_unaligned(prev_ptr);
+                    batch_t prevjm1 = batch_t::load_unaligned(prevm1_ptr);
+
+                    alignas(64) double subs[B];
+                    alignas(64) double ins[B];
+                    for (std::size_t b = 0; b < B; ++b) {
+                        int jj_idx = j + (int)b - 1;
+                        int bj = ptr_seq(js, jj_idx);
+                        double dur_j = ptr_dur(js, jj_idx);
+
+                        if (i_state == bj) {
+                            subs[b] = std::abs(timecost * (dur_i - dur_j));
+                        } else {
+                            subs[b] = ptr_sm(i_state, bj) + (dur_i + dur_j) * timecost;
+                        }
+                        ins[b] = ptr_indel(bj) + timecost * dur_j;
+                    }
 
-                    xsimd::batch<double, xsimd::default_arch> minimum_batch = prev[j] + getIndel(is, i - 1, i_state);
-                    xsimd::batch<double, xsimd::default_arch> j_indel_batch = curr[j - 1] + getIndel(js, j - 1, j_state);
-                    xsimd::batch<double, xsimd::default_arch> sub_batch = prev[j - 1] + getSubCost(i_state, j_state, is, i - 1, js, j - 1);
+                    batch_t sub_batch = batch_t::load_unaligned(subs);
+                    batch_t cand_del = prevj + batch_t(del_cost_i);
+                    batch_t cand_sub = prevjm1 + sub_batch;
+                    batch_t vert = xsimd::min(cand_del, cand_sub);
+
+                    double running = curr[j - 1] + ins[0];
+                    for (std::size_t b = 0; b < B; ++b) {
+                        double v = vert.get(b);
+                        double c = std::min(v, running);
+                        curr[j + (int)b] = c;
+                        if (b + 1 < B) running = c + ins[b + 1];
+                    }
+                }
 
-                    xsimd::batch<double> result = xsimd::min(xsimd::min(minimum_batch, j_indel_batch), sub_batch);
-                    curr[j] = result.get(0);
+                // tail scalar handling
+                for (; j < nSuf; ++j) {
+                    j_state = ptr_seq(js, j - 1);
+                    double minimum = prev[j] + del_cost_i;
+                    double j_indel = curr[j - 1] + (ptr_indel(j_state) + timecost * ptr_dur(js, j - 1));
+                    double sub = prev[j - 1] + (
+                        (i_state == j_state)
+                        ? std::abs(timecost * (dur_i - ptr_dur(js, j - 1)))
+                        : (ptr_sm(i_state, j_state) + (dur_i + ptr_dur(js, j - 1)) * timecost)
+                    );
+                    curr[j] = std::min({ minimum, j_indel, sub });
                 }
 
                 std::swap(prev, curr);
             }
 
-            maxpossiblecost = std::abs(nn - mm) * indel + maxscost * std::min(mm, nn);
+            double maxpossiblecost = std::abs(nn - mm) * indel + maxscost * std::min(mm, nn);
             double ml = double(mm) * indel;
             double nl = double(nn) * indel;
 
@@ -172,26 +209,16 @@ public:
         }
     }
 
-
     py::array_t<double> compute_all_distances() {
         try {
-            auto buffer = dist_matrix.mutable_unchecked<2>();
-
-            #pragma omp parallel for schedule(dynamic)
-            for (int i = 0; i < nseq; i++) {
-                for (int j = i; j < nseq; j++) {
-                    buffer(i, j) = compute_distance(i, j);
+            return dp_utils::compute_all_distances(
+                nseq,
+                fmatsize,
+                dist_matrix,
+                [this](int i, int j, double* prev, double* curr) {
+                    return this->compute_distance(i, j, prev, curr);
                 }
-            }
-
-            #pragma omp parallel for schedule(dynamic)
-            for (int i = 0; i < nseq; i++) {
-                for (int j = i + 1; j < nseq; j++) {
-                    buffer(j, i) = buffer(i, j);
-                }
-            }
-
-            return dist_matrix;
+            );
         } catch (const std::exception& e) {
             py::print("Error in compute_all_distances: ", e.what());
             throw;
@@ -202,15 +229,24 @@ public:
         try {
             auto buffer = refdist_matrix.mutable_unchecked<2>();
 
-            #pragma omp parallel for schedule(static)
-            for (int rseq = rseq1; rseq < rseq2; rseq ++) {
-                for (int is = 0; is < nseq; is ++) {
-                    if(is == rseq){
-                        buffer(is, rseq-rseq1) = 0;
-                    }else{
-                        buffer(is, rseq-rseq1) = compute_distance(is, rseq);
+            #pragma omp parallel
+            {
+                double* prev = dp_utils::aligned_alloc_double(static_cast<size_t>(fmatsize));
+                double* curr = dp_utils::aligned_alloc_double(static_cast<size_t>(fmatsize));
+
+                #pragma omp for schedule(static)
+                for (int rseq = rseq1; rseq < rseq2; rseq ++) {
+                    for (int is = 0; is < nseq; is ++) {
+                        double cmpres = 0;
+                        if(is != rseq){
+                            cmpres = compute_distance(is, rseq, prev, curr);
+                        }
+
+                        buffer(is, rseq - rseq1) = cmpres;
                     }
                 }
+                dp_utils::aligned_free_double(prev);
+                dp_utils::aligned_free_double(curr);
             }
 
             return refdist_matrix;

sequenzo/dissimilarity_measures/src/dp_utils.h

@@ -0,0 +1,160 @@
+#pragma once
+
+#include <pybind11/pybind11.h>
+#include <pybind11/numpy.h>
+#ifdef _OPENMP
+    #include <omp.h>
+#endif
+#include <cstdlib>
+#include <new>
+
+namespace dp_utils {
+
+// Cross-platform aligned allocation for double buffers
+#ifdef _WIN32
+inline double* aligned_alloc_double(size_t size, size_t align = 64) {
+    return reinterpret_cast<double*>(_aligned_malloc(size * sizeof(double), align));
+}
+inline void aligned_free_double(double* ptr) {
+    _aligned_free(ptr);
+}
+#else
+inline double* aligned_alloc_double(size_t size, size_t align = 64) {
+    void* ptr = nullptr;
+    if (posix_memalign(&ptr, align, size * sizeof(double)) != 0) throw std::bad_alloc();
+    return reinterpret_cast<double*>(ptr);
+}
+inline void aligned_free_double(double* ptr) { free(ptr); }
+#endif
+
+// Generic pairwise symmetric computation helper
+// ComputeFn signature: double(int i, int j, double* prev, double* curr)
+template <typename ComputeFn>
+inline pybind11::array_t<double> compute_all_distances(
+    int nseq,
+    int fmatsize,
+    pybind11::array_t<double>& dist_matrix,
+    ComputeFn&& compute_fn
+) {
+    auto buffer = dist_matrix.mutable_unchecked<2>();
+
+    #pragma omp parallel
+    {
+        double* prev = aligned_alloc_double(static_cast<size_t>(fmatsize));
+        double* curr = aligned_alloc_double(static_cast<size_t>(fmatsize));
+
+        #pragma omp for schedule(static)
+        for (int i = 0; i < nseq; i++) {
+            for (int j = i; j < nseq; j++) {
+                buffer(i, j) = compute_fn(i, j, prev, curr);
+            }
+        }
+
+        aligned_free_double(prev);
+        aligned_free_double(curr);
+    }
+
+    #pragma omp parallel for schedule(static)
+    for (int i = 0; i < nseq; i++) {
+        for (int j = i + 1; j < nseq; j++) {
+            buffer(j, i) = buffer(i, j);
+        }
+    }
+
+    return dist_matrix;
+}
+
+// Generic pairwise symmetric computation helper (no buffers)
+// ComputeFn signature: double(int i, int j)
+template <typename ComputeFn>
+inline pybind11::array_t<double> compute_all_distances_simple(
+    int nseq,
+    pybind11::array_t<double>& dist_matrix,
+    ComputeFn&& compute_fn
+) {
+    auto buffer = dist_matrix.mutable_unchecked<2>();
+
+    #pragma omp parallel
+    {
+        #pragma omp for schedule(static)
+        for (int i = 0; i < nseq; i++) {
+            for (int j = i; j < nseq; j++) {
+                buffer(i, j) = compute_fn(i, j);
+            }
+        }
+    }
+
+    #pragma omp parallel for schedule(static)
+    for (int i = 0; i < nseq; ++i) {
+        for (int j = i + 1; j < nseq; ++j) {
+            buffer(j, i) = buffer(i, j);
+        }
+    }
+
+    return dist_matrix;
+}
+
+// Generic reference-sequence computation helper (no buffers)
+// ComputeFn signature: double(int is, int rseq)
+template <typename ComputeFn>
+inline pybind11::array_t<double> compute_refseq_distances_simple(
+    int nseq,
+    int rseq1,
+    int rseq2,
+    pybind11::array_t<double>& refdist_matrix,
+    ComputeFn&& compute_fn
+) {
+    auto buffer = refdist_matrix.mutable_unchecked<2>();
+
+    #pragma omp parallel
+    {
+        #pragma omp for schedule(guided)
+        for (int rseq = rseq1; rseq < rseq2; rseq++) {
+            for (int is = 0; is < nseq; is++) {
+                buffer(is, rseq - rseq1) = (is == rseq) ? 0.0 : compute_fn(is, rseq);
+            }
+        }
+    }
+
+    return refdist_matrix;
+}
+
+// Generic reference-sequence computation helper (with DP buffers)
+// ComputeFn signature: double(int is, int rseq, double* prev, double* curr)
+template <typename ComputeFn>
+inline pybind11::array_t<double> compute_refseq_distances_buffered(
+    int nseq,
+    int rseq1,
+    int rseq2,
+    int fmatsize,
+    pybind11::array_t<double>& refdist_matrix,
+    ComputeFn&& compute_fn
+) {
+    auto buffer = refdist_matrix.mutable_unchecked<2>();
+
+    #pragma omp parallel
+    {
+        double* prev = aligned_alloc_double(static_cast<size_t>(fmatsize));
+        double* curr = aligned_alloc_double(static_cast<size_t>(fmatsize));
+
+        #pragma omp for schedule(static)
+        for (int rseq = rseq1; rseq < rseq2; rseq++) {
+            for (int is = 0; is < nseq; is++) {
+                double cmpres = 0.0;
+                if (is != rseq) {
+                    cmpres = compute_fn(is, rseq, prev, curr);
+                }
+                buffer(is, rseq - rseq1) = cmpres;
+            }
+        }
+
+        aligned_free_double(prev);
+        aligned_free_double(curr);
+    }
+
+    return refdist_matrix;
+}
+
+} // namespace dp_utils
+
+

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/common/xsimd_common_arithmetic.hpp

@@ -34,6 +34,13 @@ namespace xsimd
                                  { return x << y; },
                                  self, other);
         }
+        template <size_t shift, class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value, void>::type*/>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(shift < bits, "Shift must be less than the number of bits in T");
+            return bitwise_lshift(self, shift, A {});
+        }
 
         // bitwise_rshift
         template <class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value, void>::type*/>
@@ -43,6 +50,13 @@ namespace xsimd
                                  { return x >> y; },
                                  self, other);
         }
+        template <size_t shift, class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value, void>::type*/>
+        XSIMD_INLINE batch<T, A> bitwise_rshift(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(shift < bits, "Shift must be less than the number of bits in T");
+            return bitwise_rshift(self, shift, A {});
+        }
 
         // decr
         template <class A, class T>
@@ -127,18 +141,16 @@ namespace xsimd
             return { res_r, res_i };
         }
 
-        // hadd
-        template <class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value, void>::type*/>
-        XSIMD_INLINE T hadd(batch<T, A> const& self, requires_arch<common>) noexcept
+        // fmas
+        template <class A, class T>
+        XSIMD_INLINE batch<T, A> fmas(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z, requires_arch<common>) noexcept
         {
-            alignas(A::alignment()) T buffer[batch<T, A>::size];
-            self.store_aligned(buffer);
-            T res = 0;
-            for (T val : buffer)
+            struct even_lane
             {
-                res += val;
-            }
-            return res;
+                static constexpr bool get(unsigned const i, unsigned) noexcept { return (i & 1u) == 0; }
+            };
+            const auto mask = make_batch_bool_constant<T, even_lane, A>();
+            return fma(x, y, select(mask, neg(z), z));
         }
 
         // incr
@@ -168,16 +180,30 @@ namespace xsimd
         template <class A, class T, class STy>
         XSIMD_INLINE batch<T, A> rotl(batch<T, A> const& self, STy other, requires_arch<common>) noexcept
         {
-            constexpr auto N = std::numeric_limits<T>::digits;
-            return (self << other) | (self >> (N - other));
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            return (self << other) | (self >> (bits - other));
+        }
+        template <size_t count, class A, class T>
+        XSIMD_INLINE batch<T, A> rotl(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(count < bits, "Count amount must be less than the number of bits in T");
+            return bitwise_lshift<count>(self) | bitwise_rshift<bits - count>(self);
         }
 
         // rotr
         template <class A, class T, class STy>
         XSIMD_INLINE batch<T, A> rotr(batch<T, A> const& self, STy other, requires_arch<common>) noexcept
         {
-            constexpr auto N = std::numeric_limits<T>::digits;
-            return (self >> other) | (self << (N - other));
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            return (self >> other) | (self << (bits - other));
+        }
+        template <size_t count, class A, class T>
+        XSIMD_INLINE batch<T, A> rotr(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(count < bits, "Count must be less than the number of bits in T");
+            return bitwise_rshift<count>(self) | bitwise_lshift<bits - count>(self);
         }
 
         // sadd
@@ -191,10 +217,9 @@ namespace xsimd
         {
             if (std::is_signed<T>::value)
             {
-                auto mask = (other >> (8 * sizeof(T) - 1));
                 auto self_pos_branch = min(std::numeric_limits<T>::max() - other, self);
                 auto self_neg_branch = max(std::numeric_limits<T>::min() - other, self);
-                return other + select(batch_bool<T, A>(mask.data), self_neg_branch, self_pos_branch);
+                return other + select(other >= 0, self_pos_branch, self_neg_branch);
             }
             else
             {

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/common/xsimd_common_complex.hpp

@@ -78,11 +78,15 @@ namespace xsimd
             using batch_type = complex_batch_type_t<batch<T, A>>;
             using real_batch = typename batch_type::real_batch;
             using real_value_type = typename real_batch::value_type;
+#ifdef __FAST_MATH__
+            return { self };
+#else
             auto cond = xsimd::isinf(real(self)) || xsimd::isinf(imag(self));
             return select(cond,
                           batch_type(constants::infinity<real_batch>(),
                                      copysign(real_batch(real_value_type(0)), imag(self))),
                           batch_type(self));
+#endif
         }
 
         template <class A, class T>

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/common/xsimd_common_details.hpp

@@ -47,6 +47,8 @@ namespace xsimd
     template <class T, class A>
     XSIMD_INLINE batch<T, A> fms(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z) noexcept;
     template <class T, class A>
+    XSIMD_INLINE batch<T, A> fmas(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z) noexcept;
+    template <class T, class A>
     XSIMD_INLINE batch<T, A> frexp(const batch<T, A>& x, const batch<as_integer_t<T>, A>& e) noexcept;
     template <class T, class A, uint64_t... Coefs>
     XSIMD_INLINE batch<T, A> horner(const batch<T, A>& self) noexcept;
@@ -75,6 +77,8 @@ namespace xsimd
     template <class T, class A>
     XSIMD_INLINE T reduce_add(batch<T, A> const&) noexcept;
     template <class T, class A>
+    XSIMD_INLINE T reduce_mul(batch<T, A> const&) noexcept;
+    template <class T, class A>
     XSIMD_INLINE batch<T, A> select(batch_bool<T, A> const&, batch<T, A> const&, batch<T, A> const&) noexcept;
     template <class T, class A>
     XSIMD_INLINE batch<std::complex<T>, A> select(batch_bool<T, A> const&, batch<std::complex<T>, A> const&, batch<std::complex<T>, A> const&) noexcept;
@@ -90,6 +94,9 @@ namespace xsimd
     XSIMD_INLINE std::pair<batch<T, A>, batch<T, A>> sincos(batch<T, A> const& self) noexcept;
     template <class T, class A>
     XSIMD_INLINE batch<T, A> sqrt(batch<T, A> const& self) noexcept;
+    template <class T, class A, class Vt, Vt... Values>
+    XSIMD_INLINE typename std::enable_if<std::is_arithmetic<T>::value, batch<T, A>>::type
+    swizzle(batch<T, A> const& x, batch_constant<Vt, A, Values...> mask) noexcept;
     template <class T, class A>
     XSIMD_INLINE batch<T, A> tan(batch<T, A> const& self) noexcept;
     template <class T, class A>

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/common/xsimd_common_logical.hpp

@@ -124,12 +124,22 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE batch_bool<float, A> isinf(batch<float, A> const& self, requires_arch<common>) noexcept
         {
+#ifdef __FAST_MATH__
+            (void)self;
+            return { false };
+#else
             return abs(self) == std::numeric_limits<float>::infinity();
+#endif
         }
         template <class A>
         XSIMD_INLINE batch_bool<double, A> isinf(batch<double, A> const& self, requires_arch<common>) noexcept
         {
+#ifdef __FAST_MATH__
+            (void)self;
+            return { false };
+#else
             return abs(self) == std::numeric_limits<double>::infinity();
+#endif
         }
 
         // isfinite