enhance: optimize get norms function (#950)

Signed-off-by: cqy123456 <[email protected]>
zilliztech · Nov 15, 2024 · 6b7d756 · 6b7d756
1 parent d3605fb
commit 6b7d756
Show file tree

Hide file tree

Showing 6 changed files with 106 additions and 8 deletions.
diff --git a/src/common/comp/brute_force.cc b/src/common/comp/brute_force.cc
@@ -114,7 +114,12 @@ BruteForce::Search(const DataSetPtr base_dataset, const DataSetPtr query_dataset
     int topk = cfg.k.value();
     auto labels = std::make_unique<int64_t[]>(nq * topk);
     auto distances = std::make_unique<float[]>(nq * topk);
-
+    std::unique_ptr<float[]> norms = nullptr;
+    if (is_cosine) {
+        ThreadPool::ScopedSearchOmpSetter setter(1);
+        norms = std::make_unique<float[]>(nb);
+        faiss::fvec_norms_L2(norms.get(), (float*)xb, dim, nb);
+    }
     auto pool = ThreadPool::GetGlobalSearchThreadPool();
     std::vector<folly::Future<Status>> futs;
     futs.reserve(nq);
@@ -139,7 +144,8 @@ BruteForce::Search(const DataSetPtr base_dataset, const DataSetPtr query_dataset
                     faiss::float_minheap_array_t buf{(size_t)1, (size_t)topk, cur_labels, cur_distances};
                     if (is_cosine) {
                         auto copied_query = CopyAndNormalizeVecs(cur_query, 1, dim);
-                        faiss::knn_cosine(copied_query.get(), (const float*)xb, nullptr, dim, 1, nb, &buf, id_selector);
+                        faiss::knn_cosine(copied_query.get(), (const float*)xb, norms.get(), dim, 1, nb, &buf,
+                                          id_selector);
                     } else {
                         faiss::knn_inner_product(cur_query, (const float*)xb, dim, 1, nb, &buf, id_selector);
                     }
@@ -248,6 +254,13 @@ BruteForce::SearchWithBuf(const DataSetPtr base_dataset, const DataSetPtr query_
     auto labels = ids;
     auto distances = dis;
 
+    std::unique_ptr<float[]> norms = nullptr;
+    if (is_cosine) {
+        ThreadPool::ScopedSearchOmpSetter setter(1);
+        norms = std::make_unique<float[]>(nb);
+        faiss::fvec_norms_L2(norms.get(), (float*)xb, dim, nb);
+    }
+
     auto pool = ThreadPool::GetGlobalSearchThreadPool();
     std::vector<folly::Future<Status>> futs;
     futs.reserve(nq);
@@ -272,7 +285,8 @@ BruteForce::SearchWithBuf(const DataSetPtr base_dataset, const DataSetPtr query_
                     faiss::float_minheap_array_t buf{(size_t)1, (size_t)topk, cur_labels, cur_distances};
                     if (is_cosine) {
                         auto copied_query = CopyAndNormalizeVecs(cur_query, 1, dim);
-                        faiss::knn_cosine(copied_query.get(), (const float*)xb, nullptr, dim, 1, nb, &buf, id_selector);
+                        faiss::knn_cosine(copied_query.get(), (const float*)xb, norms.get(), dim, 1, nb, &buf,
+                                          id_selector);
                     } else {
                         faiss::knn_inner_product(cur_query, (const float*)xb, dim, 1, nb, &buf, id_selector);
                     }
@@ -408,6 +422,13 @@ BruteForce::RangeSearch(const DataSetPtr base_dataset, const DataSetPtr query_da
     std::vector<std::vector<int64_t>> result_id_array(nq);
     std::vector<std::vector<float>> result_dist_array(nq);
 
+    std::unique_ptr<float[]> norms = nullptr;
+    if (is_cosine) {
+        ThreadPool::ScopedSearchOmpSetter setter(1);
+        norms = std::make_unique<float[]>(nb);
+        faiss::fvec_norms_L2(norms.get(), (float*)xb, dim, nb);
+    }
+
     std::vector<folly::Future<Status>> futs;
     futs.reserve(nq);
     for (int i = 0; i < nq; ++i) {
@@ -452,8 +473,8 @@ BruteForce::RangeSearch(const DataSetPtr base_dataset, const DataSetPtr query_da
                     auto cur_query = (const float*)xq + dim * index;
                     if (is_cosine) {
                         auto copied_query = CopyAndNormalizeVecs(cur_query, 1, dim);
-                        faiss::range_search_cosine(copied_query.get(), (const float*)xb, nullptr, dim, 1, nb, radius,
-                                                   &res, id_selector);
+                        faiss::range_search_cosine(copied_query.get(), (const float*)xb, norms.get(), dim, 1, nb,
+                                                   radius, &res, id_selector);
                     } else {
                         faiss::range_search_inner_product(cur_query, (const float*)xb, dim, 1, nb, radius, &res,
                                                           id_selector);
@@ -678,6 +699,13 @@ BruteForce::AnnIterator(const DataSetPtr base_dataset, const DataSetPtr query_da
     faiss::MetricType faiss_metric_type = result.value();
     bool is_cosine = IsMetricType(metric_str, metric::COSINE);
 
+    std::unique_ptr<float[]> norms = nullptr;
+    if (is_cosine) {
+        ThreadPool::ScopedSearchOmpSetter setter(1);
+        norms = std::make_unique<float[]>(nb);
+        faiss::fvec_norms_L2(norms.get(), (float*)xb, dim, nb);
+    }
+
     auto pool = ThreadPool::GetGlobalSearchThreadPool();
     auto vec = std::vector<IndexNode::IteratorPtr>(nq, nullptr);
     std::vector<folly::Future<Status>> futs;
@@ -703,7 +731,7 @@ BruteForce::AnnIterator(const DataSetPtr base_dataset, const DataSetPtr query_da
                     auto cur_query = (const float*)xq + dim * index;
                     if (is_cosine) {
                         auto copied_query = CopyAndNormalizeVecs(cur_query, 1, dim);
-                        faiss::all_cosine(copied_query.get(), (const float*)xb, nullptr, dim, 1, nb, distances_ids,
+                        faiss::all_cosine(copied_query.get(), (const float*)xb, norms.get(), dim, 1, nb, distances_ids,
                                           id_selector);
                     } else {
                         faiss::all_inner_product(cur_query, (const float*)xb, dim, 1, nb, distances_ids, id_selector);

diff --git a/src/simd/distances_avx.cc b/src/simd/distances_avx.cc
@@ -461,6 +461,43 @@ ivec_L2sqr_avx(const int8_t* x, const int8_t* y, size_t d) {
     return res;
 }
 
+float
+fvec_norm_L2sqr_avx(const float* x, size_t d) {
+    __m256 msum_0 = _mm256_setzero_ps();
+    __m256 msum_1 = _mm256_setzero_ps();
+    while (d >= 16) {
+        auto mx_0 = _mm256_loadu_ps(x);
+        auto mx_1 = _mm256_loadu_ps(x + 8);
+        msum_0 = _mm256_fmadd_ps(mx_0, mx_0, msum_0);
+        msum_1 = _mm256_fmadd_ps(mx_1, mx_1, msum_1);
+        x += 16;
+        d -= 16;
+    }
+    msum_0 = msum_0 + msum_1;
+    if (d >= 8) {
+        auto mx = _mm256_loadu_ps(x);
+        msum_0 = _mm256_fmadd_ps(mx, mx, msum_0);
+        x += 8;
+        d -= 8;
+    }
+    if (d > 0) {
+        __m128 rest_0 = _mm_setzero_ps();
+        __m128 rest_1 = _mm_setzero_ps();
+        if (d >= 4) {
+            rest_0 = _mm_loadu_ps(x);
+            x += 4;
+            d -= 4;
+        }
+        if (d >= 0) {
+            rest_1 = masked_read(d, x);
+        }
+        auto mx = _mm256_set_m128(rest_0, rest_1);
+        msum_0 = _mm256_fmadd_ps(mx, mx, msum_0);
+    }
+    auto res = _mm256_reduce_add_ps(msum_0);
+    return res;
+}
+
 float
 fp16_vec_norm_L2sqr_avx(const knowhere::fp16* x, size_t d) {
     __m256 msum_0 = _mm256_setzero_ps();

diff --git a/src/simd/distances_avx.h b/src/simd/distances_avx.h
@@ -79,6 +79,9 @@ ivec_inner_product_avx(const int8_t* x, const int8_t* y, size_t d);
 int32_t
 ivec_L2sqr_avx(const int8_t* x, const int8_t* y, size_t d);
 
+float
+fvec_norm_L2sqr_avx(const float* x, size_t d);
+
 float
 fp16_vec_norm_L2sqr_avx(const knowhere::fp16* x, size_t d);
 

diff --git a/src/simd/distances_avx512.cc b/src/simd/distances_avx512.cc
@@ -487,6 +487,33 @@ ivec_L2sqr_avx512(const int8_t* x, const int8_t* y, size_t d) {
     return res;
 }
 
+float
+fvec_norm_L2sqr_avx512(const float* x, size_t d) {
+    __m512 m512_res = _mm512_setzero_ps();
+    __m512 m512_res_0 = _mm512_setzero_ps();
+    while (d >= 32) {
+        auto mx_0 = _mm512_loadu_ps(x);
+        auto mx_1 = _mm512_loadu_ps(x + 16);
+        m512_res = _mm512_fmadd_ps(mx_0, mx_0, m512_res);
+        m512_res_0 = _mm512_fmadd_ps(mx_1, mx_1, m512_res_0);
+        x += 32;
+        d -= 32;
+    }
+    m512_res = m512_res + m512_res_0;
+    if (d >= 16) {
+        auto mx = _mm512_loadu_ps(x);
+        m512_res = _mm512_fmadd_ps(mx, mx, m512_res);
+        x += 16;
+        d -= 16;
+    }
+    if (d > 0) {
+        const __mmask16 mask = (1U << d) - 1U;
+        auto mx = _mm512_maskz_loadu_ps(mask, x);
+        m512_res = _mm512_fmadd_ps(mx, mx, m512_res);
+    }
+    return _mm512_reduce_add_ps(m512_res);
+}
+
 float
 fp16_vec_norm_L2sqr_avx512(const knowhere::fp16* x, size_t d) {
     __m512 m512_res = _mm512_setzero_ps();

diff --git a/src/simd/distances_avx512.h b/src/simd/distances_avx512.h
@@ -78,6 +78,9 @@ ivec_inner_product_avx512(const int8_t* x, const int8_t* y, size_t d);
 int32_t
 ivec_L2sqr_avx512(const int8_t* x, const int8_t* y, size_t d);
 
+float
+fvec_norm_L2sqr_avx512(const float* x, size_t d);
+
 float
 fp16_vec_norm_L2sqr_avx512(const knowhere::fp16* x, size_t d);
 

diff --git a/src/simd/hook.cc b/src/simd/hook.cc
@@ -185,7 +185,7 @@ fvec_hook(std::string& simd_type) {
         fvec_L1 = fvec_L1_avx512;
         fvec_Linf = fvec_Linf_avx512;
 
-        fvec_norm_L2sqr = fvec_norm_L2sqr_sse;
+        fvec_norm_L2sqr = fvec_norm_L2sqr_avx512;
         fvec_L2sqr_ny = fvec_L2sqr_ny_sse;
         fvec_inner_products_ny = fvec_inner_products_ny_sse;
         fvec_madd = fvec_madd_avx512;
@@ -213,7 +213,7 @@ fvec_hook(std::string& simd_type) {
         fvec_L1 = fvec_L1_avx;
         fvec_Linf = fvec_Linf_avx;
 
-        fvec_norm_L2sqr = fvec_norm_L2sqr_sse;
+        fvec_norm_L2sqr = fvec_norm_L2sqr_avx;
         fvec_L2sqr_ny = fvec_L2sqr_ny_sse;
         fvec_inner_products_ny = fvec_inner_products_ny_sse;
         fvec_madd = fvec_madd_avx;