Merge branch 'haol/distance-cpp-remove-omp' into 'mqdb-dev'

remove OMP and use manual sgemm to avoid debug error

Closes #17

See merge request mqdb/faiss!34

faiss/utils/distances.cpp

@@ -32,6 +32,8 @@
 #define FINTEGER long
 #endif
 
+// Remove all OpenMP to avoid CPU interference
+
 extern "C" {
 
 /* declare BLAS functions, see http://www.netlib.org/clapack/cblas/ */
@@ -52,6 +54,41 @@ int sgemm_(
         FINTEGER* ldc);
 }
 
+int sgemm_manual(
+        const char* transa,
+        const char* transb,
+        FINTEGER* m,
+        FINTEGER* n,
+        FINTEGER* k,
+        const float* alpha,
+        const float* a,
+        FINTEGER* lda,
+        const float* b,
+        FINTEGER* ldb,
+        float* beta,
+        float* c,
+        FINTEGER* ldc
+        ){
+    for (int i = 0; i < *m; ++i) {
+        for (int j = 0; j < *n; ++j) {
+            float result = 0.0;
+            for (int l = 0; l < (*k); ++l) {
+                result += (*alpha) * a[i * (*lda) + l] * b[l * (*ldb) + j];
+            }
+            c[i * (*ldc) + j] = (*beta) * c[i * (*ldc) + j] + result;
+        }
+    }
+    return 0;
+}
+
+
+#ifdef NDEBUG
+#define sgemm_func sgemm_
+#else
+// use sgemm_manual in debug mode to avoid sleep error in ClickHouse
+#define sgemm_func sgemm_manual
+#endif
+
 namespace faiss {
 
 /***************************************************************************
@@ -64,7 +101,7 @@ void fvec_norms_L2(
         const float* __restrict x,
         size_t d,
         size_t nx) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t i = 0; i < nx; i++) {
         nr[i] = sqrtf(fvec_norm_L2sqr(x + i * d, d));
     }
@@ -75,13 +112,13 @@ void fvec_norms_L2sqr(
         const float* __restrict x,
         size_t d,
         size_t nx) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t i = 0; i < nx; i++)
         nr[i] = fvec_norm_L2sqr(x + i * d, d);
 }
 
 void fvec_renorm_L2(size_t d, size_t nx, float* __restrict x) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t i = 0; i < nx; i++) {
         float* __restrict xi = x + i * d;
 
@@ -289,7 +326,7 @@ void exhaustive_L2sqr_blas_default_impl(
                        ip_block.get(),
                        &nyi);
             }
-#pragma omp parallel for
+//#pragma omp parallel for
             for (int64_t i = i0; i < i1; i++) {
                 float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
 
@@ -370,7 +407,7 @@ void exhaustive_L2sqr_blas_cmax_avx2(
             {
                 float one = 1, zero = 0;
                 FINTEGER nyi = j1 - j0, nxi = i1 - i0, di = d;
-                sgemm_("Transpose",
+                sgemm_func("Transpose",
                        "Not transpose",
                        &nyi,
                        &nxi,
@@ -384,7 +421,7 @@ void exhaustive_L2sqr_blas_cmax_avx2(
                        ip_block.get(),
                        &nyi);
             }
-#pragma omp parallel for
+//#pragma omp parallel for
             for (int64_t i = i0; i < i1; i++) {
                 float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
 
@@ -824,7 +861,7 @@ void pairwise_indexed_L2sqr(
         const float* y,
         const int64_t* iy,
         float* dis) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t j = 0; j < n; j++) {
         if (ix[j] >= 0 && iy[j] >= 0) {
             dis[j] = fvec_L2sqr(x + d * ix[j], y + d * iy[j], d);
@@ -840,7 +877,7 @@ void pairwise_indexed_inner_product(
         const float* y,
         const int64_t* iy,
         float* dis) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t j = 0; j < n; j++) {
         if (ix[j] >= 0 && iy[j] >= 0) {
             dis[j] = fvec_inner_product(x + d * ix[j], y + d * iy[j], d);
@@ -865,7 +902,7 @@ void knn_inner_products_by_idx(
         ld_ids = ny;
     }
 
-#pragma omp parallel for if (nx > 100)
+//#pragma omp parallel for if (nx > 100)
     for (int64_t i = 0; i < nx; i++) {
         const float* x_ = x + i * d;
         const int64_t* idsi = ids + i * ld_ids;
@@ -901,7 +938,7 @@ void knn_L2sqr_by_idx(
     if (ld_ids < 0) {
         ld_ids = ny;
     }
-#pragma omp parallel for if (nx > 100)
+//#pragma omp parallel for if (nx > 100)
     for (int64_t i = 0; i < nx; i++) {
         const float* x_ = x + i * d;
         const int64_t* __restrict idsi = ids + i * ld_ids;
@@ -941,11 +978,11 @@ void pairwise_L2sqr(
     // store in beginning of distance matrix to avoid malloc
     float* b_norms = dis;
 
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t i = 0; i < nb; i++)
         b_norms[i] = fvec_norm_L2sqr(xb + i * ldb, d);
 
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t i = 1; i < nq; i++) {
         float q_norm = fvec_norm_L2sqr(xq + i * ldq, d);
         for (int64_t j = 0; j < nb; j++)
@@ -984,7 +1021,7 @@ void inner_product_to_L2sqr(
         const float* nr2,
         size_t n1,
         size_t n2) {
-#pragma omp parallel for
+//#pragma omp parallel for
     for (int64_t j = 0; j < n1; j++) {
         float* disj = dis + j * n2;
         for (size_t i = 0; i < n2; i++)