faiss 0.6.2 → 0.6.3

data/vendor/faiss/faiss/utils/simd_impl/distances_avx512.cpp

@@ -9,6 +9,35 @@
 
 #include <immintrin.h>
 
+#include <memory>
+
+#include <faiss/impl/AuxIndexStructures.h>
+#include <faiss/impl/ResultHandler.h>
+#include <faiss/utils/distances_fused/distances_fused.h>
+#include <faiss/utils/simd_impl/exhaustive_L2sqr_blas_cmax.h>
+
+#ifndef FINTEGER
+#define FINTEGER long
+#endif
+
+extern "C" {
+
+int sgemm_(
+        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);
+}
+
 #define THE_SIMD_LEVEL SIMDLevel::AVX512
 #include <faiss/utils/simd_impl/distances_autovec-inl.h>
 // NOLINTNEXTLINE(facebook-hte-InlineHeader)
@@ -954,21 +983,6 @@ size_t fvec_L2sqr_ny_nearest<SIMDLevel::AVX512>(
             &fvec_L2sqr_ny_nearest_D8<SIMDLevel::AVX512>);
 }
 
-template <>
-size_t fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::AVX512>(
-        float* distances_tmp_buffer,
-        const float* x,
-        const float* y,
-        const float* y_sqlen,
-        size_t d,
-        size_t d_offset,
-        size_t ny) {
-    return fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::NONE>(
-            distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
-}
-
-// TODO: Following functions are not used in the current codebase. Check AVX2 ,
-// respective implementation has been used
 template <size_t DIM>
 size_t fvec_L2sqr_ny_nearest_y_transposed_D(
         float* /* distances_tmp_buffer */,
@@ -1082,6 +1096,33 @@ size_t fvec_L2sqr_ny_nearest_y_transposed_D(
     return current_min_index;
 }
 
+template <>
+size_t fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::AVX512>(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        size_t d,
+        size_t d_offset,
+        size_t ny) {
+    // optimized for a few special cases
+#define DISPATCH(dval)                                     \
+    case dval:                                             \
+        return fvec_L2sqr_ny_nearest_y_transposed_D<dval>( \
+                distances_tmp_buffer, x, y, y_sqlen, d_offset, ny);
+
+    switch (d) {
+        DISPATCH(1)
+        DISPATCH(2)
+        DISPATCH(4)
+        DISPATCH(8)
+        default:
+            return fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::NONE>(
+                    distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
+    }
+#undef DISPATCH
+}
+
 template <>
 int fvec_madd_and_argmin<SIMDLevel::AVX512>(
         size_t n,
@@ -1092,4 +1133,211 @@ int fvec_madd_and_argmin<SIMDLevel::AVX512>(
     return fvec_madd_and_argmin_sse(n, a, bf, b, c);
 }
 
+template <>
+void exhaustive_L2sqr_blas_cmax<SIMDLevel::AVX512>(
+        const float* x,
+        const float* y,
+        size_t d,
+        size_t nx,
+        size_t ny,
+        Top1BlockResultHandler<CMax<float, int64_t>>& res,
+        const float* y_norms) {
+    // BLAS does not like empty matrices
+    if (nx == 0 || ny == 0) {
+        return;
+    }
+
+    /* block sizes */
+    const size_t bs_x = distance_compute_blas_query_bs;
+    const size_t bs_y = distance_compute_blas_database_bs;
+    std::unique_ptr<float[]> ip_block(new float[bs_x * bs_y]);
+    std::unique_ptr<float[]> x_norms(new float[nx]);
+    std::unique_ptr<float[]> del2;
+
+    fvec_norms_L2sqr(x_norms.get(), x, d, nx);
+
+    if (!y_norms) {
+        float* y_norms2 = new float[ny];
+        del2.reset(y_norms2);
+        fvec_norms_L2sqr(y_norms2, y, d, ny);
+        y_norms = y_norms2;
+    }
+
+    for (size_t i0 = 0; i0 < nx; i0 += bs_x) {
+        size_t i1 = i0 + bs_x;
+        if (i1 > nx) {
+            i1 = nx;
+        }
+
+        res.begin_multiple(i0, i1);
+
+        for (size_t j0 = 0; j0 < ny; j0 += bs_y) {
+            size_t j1 = j0 + bs_y;
+            if (j1 > ny) {
+                j1 = ny;
+            }
+            /* compute the actual dot products */
+            {
+                float one = 1, zero = 0;
+                FINTEGER nyi = j1 - j0, nxi = i1 - i0, di = d;
+                sgemm_("Transpose",
+                       "Not transpose",
+                       &nyi,
+                       &nxi,
+                       &di,
+                       &one,
+                       y + j0 * d,
+                       &di,
+                       x + i0 * d,
+                       &di,
+                       &zero,
+                       ip_block.get(),
+                       &nyi);
+            }
+#pragma omp parallel for schedule(static) if ((i1 - i0) >= 16)
+            for (int64_t i = static_cast<int64_t>(i0);
+                 i < static_cast<int64_t>(i1);
+                 i++) {
+                float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
+
+                _mm_prefetch((const char*)ip_line, _MM_HINT_NTA);
+                _mm_prefetch((const char*)(ip_line + 16), _MM_HINT_NTA);
+
+                // constant
+                const __m512 mul_minus2 = _mm512_set1_ps(-2);
+
+                // Track 16 min distances + 16 min indices.
+                // All the distances tracked do not take x_norms[i]
+                // into account in order to get rid of extra
+                // _mm512_add_ps(x_norms[i], ...) instructions
+                // in distance computations.
+                __m512 min_distances =
+                        _mm512_set1_ps(res.dis_tab[i] - x_norms[i]);
+
+                // these indices are local and are relative to j0.
+                // so, value 0 means j0.
+                __m512i min_indices = _mm512_set1_epi32(0);
+
+                __m512i current_indices = _mm512_setr_epi32(
+                        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+                const __m512i indices_delta = _mm512_set1_epi32(16);
+
+                // current j index
+                size_t idx_j = 0;
+                size_t count = j1 - j0;
+
+                // process 32 elements per loop
+                for (; idx_j < (count / 32) * 32; idx_j += 32, ip_line += 32) {
+                    _mm_prefetch((const char*)(ip_line + 32), _MM_HINT_NTA);
+                    _mm_prefetch((const char*)(ip_line + 48), _MM_HINT_NTA);
+
+                    // load values for norms
+                    const __m512 y_norm_0 =
+                            _mm512_loadu_ps(y_norms + idx_j + j0 + 0);
+                    const __m512 y_norm_1 =
+                            _mm512_loadu_ps(y_norms + idx_j + j0 + 16);
+
+                    // load values for dot products
+                    const __m512 ip_0 = _mm512_loadu_ps(ip_line + 0);
+                    const __m512 ip_1 = _mm512_loadu_ps(ip_line + 16);
+
+                    // compute dis = y_norm[j] - 2 * dot(x_norm[i], y_norm[j]).
+                    // x_norm[i] was dropped off because it is a constant for a
+                    // given i.
+                    __m512 distances_0 =
+                            _mm512_fmadd_ps(ip_0, mul_minus2, y_norm_0);
+                    __m512 distances_1 =
+                            _mm512_fmadd_ps(ip_1, mul_minus2, y_norm_1);
+
+                    // compare the new distances to the min distances
+                    // for each of the first group of 16 AVX512 components.
+                    const __mmask16 comparison_0 = _mm512_cmp_ps_mask(
+                            min_distances, distances_0, _CMP_LE_OS);
+
+                    // update min distances and indices with closest vectors if
+                    // needed.
+                    min_distances = _mm512_mask_blend_ps(
+                            comparison_0, distances_0, min_distances);
+                    min_indices = _mm512_mask_blend_epi32(
+                            comparison_0, current_indices, min_indices);
+                    current_indices =
+                            _mm512_add_epi32(current_indices, indices_delta);
+
+                    // compare the new distances to the min distances
+                    // for each of the second group of 16 AVX512 components.
+                    const __mmask16 comparison_1 = _mm512_cmp_ps_mask(
+                            min_distances, distances_1, _CMP_LE_OS);
+
+                    // update min distances and indices with closest vectors if
+                    // needed.
+                    min_distances = _mm512_mask_blend_ps(
+                            comparison_1, distances_1, min_distances);
+                    min_indices = _mm512_mask_blend_epi32(
+                            comparison_1, current_indices, min_indices);
+                    current_indices =
+                            _mm512_add_epi32(current_indices, indices_delta);
+                }
+
+                // dump values and find the minimum distance / minimum index
+                float min_distances_scalar[16];
+                uint32_t min_indices_scalar[16];
+                _mm512_storeu_ps(min_distances_scalar, min_distances);
+                _mm512_storeu_si512(
+                        (__m512i*)(min_indices_scalar), min_indices);
+
+                float current_min_distance = res.dis_tab[i];
+                uint32_t current_min_index = res.ids_tab[i];
+
+                // if two indices are represented with equal distance values,
+                // then the index with the min value is returned.
+                for (size_t jv = 0; jv < 16; jv++) {
+                    // add missing x_norms[i]
+                    float distance_candidate =
+                            min_distances_scalar[jv] + x_norms[i];
+
+                    // negative values can occur for identical vectors
+                    //    due to roundoff errors.
+                    if (distance_candidate < 0) {
+                        distance_candidate = 0;
+                    }
+
+                    int64_t index_candidate = min_indices_scalar[jv] + j0;
+
+                    if (current_min_distance > distance_candidate) {
+                        current_min_distance = distance_candidate;
+                        current_min_index = index_candidate;
+                    } else if (
+                            current_min_distance == distance_candidate &&
+                            current_min_index > index_candidate) {
+                        current_min_index = index_candidate;
+                    }
+                }
+
+                // process leftovers
+                for (; idx_j < count; idx_j++, ip_line++) {
+                    float ip = *ip_line;
+                    float dis = x_norms[i] + y_norms[idx_j + j0] - 2 * ip;
+                    // negative values can occur for identical vectors
+                    //    due to roundoff errors.
+                    if (dis < 0) {
+                        dis = 0;
+                    }
+
+                    if (current_min_distance > dis) {
+                        current_min_distance = dis;
+                        current_min_index = idx_j + j0;
+                    }
+                }
+
+                //
+                res.add_result(i, current_min_distance, current_min_index);
+            }
+        }
+        // Does nothing for SingleBestResultHandler, but
+        // keeping the call for the consistency.
+        res.end_multiple();
+        InterruptCallback::check();
+    }
+}
+
 } // namespace faiss

data/vendor/faiss/faiss/utils/simd_impl/distances_rvv.cpp

@@ -12,12 +12,24 @@
 #ifdef COMPILE_SIMD_RISCV_RVV
 
 #include <faiss/utils/extra_distances.h>
+#include <riscv_vector.h>
 
 namespace faiss {
 
 template <>
 float fvec_norm_L2sqr<SIMDLevel::RISCV_RVV>(const float* x, size_t d) {
-    return fvec_norm_L2sqr<SIMDLevel::NONE>(x, d);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t acc = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        acc = __riscv_vfmacc_vv_f32m8_tu(acc, vx, vx, vl);
+        i += vl;
+    }
+    vfloat32m1_t sum = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    sum = __riscv_vfredusum_vs_f32m8_f32m1(acc, sum, vlmax);
+    return __riscv_vfmv_f_s_f32m1_f32(sum);
 }
 
 template <>
@@ -25,7 +37,20 @@ float fvec_L2sqr<SIMDLevel::RISCV_RVV>(
         const float* x,
         const float* y,
         size_t d) {
-    return fvec_L2sqr<SIMDLevel::NONE>(x, y, d);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t acc = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        vfloat32m8_t vy = __riscv_vle32_v_f32m8(y + i, vl);
+        vx = __riscv_vfsub_vv_f32m8(vx, vy, vl);
+        acc = __riscv_vfmacc_vv_f32m8_tu(acc, vx, vx, vl);
+        i += vl;
+    }
+    vfloat32m1_t sum = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    sum = __riscv_vfredusum_vs_f32m8_f32m1(acc, sum, vlmax);
+    return __riscv_vfmv_f_s_f32m1_f32(sum);
 }
 
 template <>
@@ -33,12 +58,38 @@ float fvec_inner_product<SIMDLevel::RISCV_RVV>(
         const float* x,
         const float* y,
         size_t d) {
-    return fvec_inner_product<SIMDLevel::NONE>(x, y, d);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t acc = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        vfloat32m8_t vy = __riscv_vle32_v_f32m8(y + i, vl);
+        acc = __riscv_vfmacc_vv_f32m8_tu(acc, vx, vy, vl);
+        i += vl;
+    }
+    vfloat32m1_t sum = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    sum = __riscv_vfredusum_vs_f32m8_f32m1(acc, sum, vlmax);
+    return __riscv_vfmv_f_s_f32m1_f32(sum);
 }
 
 template <>
 float fvec_L1<SIMDLevel::RISCV_RVV>(const float* x, const float* y, size_t d) {
-    return fvec_L1<SIMDLevel::NONE>(x, y, d);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t acc = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        vfloat32m8_t vy = __riscv_vle32_v_f32m8(y + i, vl);
+        vx = __riscv_vfsub_vv_f32m8(vx, vy, vl);
+        vx = __riscv_vfsgnjx_vv_f32m8(vx, vx, vl);
+        acc = __riscv_vfadd_vv_f32m8_tu(acc, acc, vx, vl);
+        i += vl;
+    }
+    vfloat32m1_t sum = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    sum = __riscv_vfredusum_vs_f32m8_f32m1(acc, sum, vlmax);
+    return __riscv_vfmv_f_s_f32m1_f32(sum);
 }
 
 template <>
@@ -46,7 +97,21 @@ float fvec_Linf<SIMDLevel::RISCV_RVV>(
         const float* x,
         const float* y,
         size_t d) {
-    return fvec_Linf<SIMDLevel::NONE>(x, y, d);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t vmax = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        vfloat32m8_t vy = __riscv_vle32_v_f32m8(y + i, vl);
+        vx = __riscv_vfsub_vv_f32m8(vx, vy, vl);
+        vx = __riscv_vfsgnjx_vv_f32m8(vx, vx, vl);
+        vmax = __riscv_vfmax_vv_f32m8_tu(vmax, vmax, vx, vl);
+        i += vl;
+    }
+    vfloat32m1_t max = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    max = __riscv_vfredmax_vs_f32m8_f32m1(vmax, max, vlmax);
+    return __riscv_vfmv_f_s_f32m1_f32(max);
 }
 
 template <>
@@ -61,8 +126,10 @@ void fvec_inner_product_batch_4<SIMDLevel::RISCV_RVV>(
         float& dis1,
         float& dis2,
         float& dis3) {
-    fvec_inner_product_batch_4<SIMDLevel::NONE>(
-            x, y0, y1, y2, y3, d, dis0, dis1, dis2, dis3);
+    dis0 = fvec_inner_product<SIMDLevel::RISCV_RVV>(x, y0, d);
+    dis1 = fvec_inner_product<SIMDLevel::RISCV_RVV>(x, y1, d);
+    dis2 = fvec_inner_product<SIMDLevel::RISCV_RVV>(x, y2, d);
+    dis3 = fvec_inner_product<SIMDLevel::RISCV_RVV>(x, y3, d);
 }
 
 template <>
@@ -77,8 +144,10 @@ void fvec_L2sqr_batch_4<SIMDLevel::RISCV_RVV>(
         float& dis1,
         float& dis2,
         float& dis3) {
-    fvec_L2sqr_batch_4<SIMDLevel::NONE>(
-            x, y0, y1, y2, y3, d, dis0, dis1, dis2, dis3);
+    dis0 = fvec_L2sqr<SIMDLevel::RISCV_RVV>(x, y0, d);
+    dis1 = fvec_L2sqr<SIMDLevel::RISCV_RVV>(x, y1, d);
+    dis2 = fvec_L2sqr<SIMDLevel::RISCV_RVV>(x, y2, d);
+    dis3 = fvec_L2sqr<SIMDLevel::RISCV_RVV>(x, y3, d);
 }
 
 template <>
@@ -90,8 +159,33 @@ void fvec_L2sqr_ny_transposed<SIMDLevel::RISCV_RVV>(
         size_t d,
         size_t d_offset,
         size_t ny) {
-    fvec_L2sqr_ny_transposed<SIMDLevel::NONE>(
-            dis, x, y, y_sqlen, d, d_offset, ny);
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t acc = __riscv_vfmv_v_f_f32m8(0.0f, vlmax);
+    size_t i = 0;
+    while (i < d) {
+        size_t vl = __riscv_vsetvl_e32m8(d - i);
+        vfloat32m8_t vx = __riscv_vle32_v_f32m8(x + i, vl);
+        acc = __riscv_vfmacc_vv_f32m8_tu(acc, vx, vx, vl);
+        i += vl;
+    }
+    vfloat32m1_t sum = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    sum = __riscv_vfredusum_vs_f32m8_f32m1(acc, sum, vlmax);
+    float x_sqlen = __riscv_vfmv_f_s_f32m1_f32(sum);
+    i = 0;
+    while (i < ny) {
+        size_t vl = __riscv_vsetvl_e32m8(ny - i);
+        acc = __riscv_vfmv_v_f_f32m8(0.0f, vl);
+        for (size_t j = 0; j < d; j++) {
+            vfloat32m8_t vy = __riscv_vle32_v_f32m8(y + j * d_offset + i, vl);
+            acc = __riscv_vfmacc_vf_f32m8(acc, x[j], vy, vl);
+        }
+        vfloat32m8_t vres = __riscv_vle32_v_f32m8(y_sqlen + i, vl);
+        vres = __riscv_vfadd_vf_f32m8(vres, x_sqlen, vl);
+        acc = __riscv_vfmul_vf_f32m8(acc, 2.0f, vl);
+        vres = __riscv_vfsub_vv_f32m8(vres, acc, vl);
+        __riscv_vse32_v_f32m8(dis + i, vres, vl);
+        i += vl;
+    }
 }
 
 template <>
@@ -101,7 +195,10 @@ void fvec_inner_products_ny<SIMDLevel::RISCV_RVV>(
         const float* y,
         size_t d,
         size_t ny) {
-    fvec_inner_products_ny<SIMDLevel::NONE>(ip, x, y, d, ny);
+    for (size_t i = 0; i < ny; i++) {
+        ip[i] = fvec_inner_product<SIMDLevel::RISCV_RVV>(x, y, d);
+        y += d;
+    }
 }
 
 template <>
@@ -111,7 +208,39 @@ void fvec_L2sqr_ny<SIMDLevel::RISCV_RVV>(
         const float* y,
         size_t d,
         size_t ny) {
-    fvec_L2sqr_ny<SIMDLevel::NONE>(dis, x, y, d, ny);
+    for (size_t i = 0; i < ny; i++) {
+        dis[i] = fvec_L2sqr<SIMDLevel::RISCV_RVV>(x, y, d);
+        y += d;
+    }
+}
+
+// Index of the first element equal to the minimum of values[0..n), or n when
+// there is none (e.g. n == 0). Shared by the *_nearest and madd_and_argmin
+// kernels so the vfmin/vfredmin/vfirst sequence lives in one place.
+static size_t rvv_argmin(const float* values, size_t n) {
+    size_t vlmax = __riscv_vsetvlmax_e32m8();
+    vfloat32m8_t vmin = __riscv_vfmv_v_f_f32m8(__builtin_inff(), vlmax);
+    size_t i = 0;
+    while (i < n) {
+        size_t vl = __riscv_vsetvl_e32m8(n - i);
+        vfloat32m8_t vd = __riscv_vle32_v_f32m8(values + i, vl);
+        vmin = __riscv_vfmin_vv_f32m8_tu(vmin, vmin, vd, vl);
+        i += vl;
+    }
+    vfloat32m1_t rmin = __riscv_vfmv_s_f_f32m1(__builtin_inff(), 1);
+    rmin = __riscv_vfredmin_vs_f32m8_f32m1(vmin, rmin, vlmax);
+    float min_val = __riscv_vfmv_f_s_f32m1_f32(rmin);
+    i = 0;
+    while (i < n) {
+        size_t vl = __riscv_vsetvl_e32m8(n - i);
+        vfloat32m8_t vd = __riscv_vle32_v_f32m8(values + i, vl);
+        long j = __riscv_vfirst_m_b4(
+                __riscv_vmfeq_vf_f32m8_b4(vd, min_val, vl), vl);
+        if (j >= 0)
+            return i + static_cast<size_t>(j);
+        i += vl;
+    }
+    return n;
 }
 
 template <>
@@ -121,8 +250,9 @@ size_t fvec_L2sqr_ny_nearest<SIMDLevel::RISCV_RVV>(
         const float* y,
         size_t d,
         size_t ny) {
-    return fvec_L2sqr_ny_nearest<SIMDLevel::NONE>(
-            distances_tmp_buffer, x, y, d, ny);
+    fvec_L2sqr_ny<SIMDLevel::RISCV_RVV>(distances_tmp_buffer, x, y, d, ny);
+    const size_t j = rvv_argmin(distances_tmp_buffer, ny);
+    return j < ny ? j : 0;
 }
 
 template <>
@@ -134,8 +264,10 @@ size_t fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::RISCV_RVV>(
         size_t d,
         size_t d_offset,
         size_t ny) {
-    return fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::NONE>(
+    fvec_L2sqr_ny_transposed<SIMDLevel::RISCV_RVV>(
             distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
+    const size_t j = rvv_argmin(distances_tmp_buffer, ny);
+    return j < ny ? j : 0;
 }
 
 template <>
@@ -145,7 +277,15 @@ void fvec_madd<SIMDLevel::RISCV_RVV>(
         float bf,
         const float* b,
         float* c) {
-    fvec_madd<SIMDLevel::NONE>(n, a, bf, b, c);
+    size_t i = 0;
+    while (i < n) {
+        size_t vl = __riscv_vsetvl_e32m8(n - i);
+        vfloat32m8_t va = __riscv_vle32_v_f32m8(a + i, vl);
+        vfloat32m8_t vb = __riscv_vle32_v_f32m8(b + i, vl);
+        va = __riscv_vfmacc_vf_f32m8(va, bf, vb, vl);
+        __riscv_vse32_v_f32m8(c + i, va, vl);
+        i += vl;
+    }
 }
 
 template <>
@@ -155,7 +295,9 @@ int fvec_madd_and_argmin<SIMDLevel::RISCV_RVV>(
         float bf,
         const float* b,
         float* c) {
-    return fvec_madd_and_argmin<SIMDLevel::NONE>(n, a, bf, b, c);
+    fvec_madd<SIMDLevel::RISCV_RVV>(n, a, bf, b, c);
+    const size_t j = rvv_argmin(c, n);
+    return j < n ? static_cast<int>(j) : -1;
 }
 
 #define DEFINE_VECTOR_DISTANCE_RVV_FALLBACK(metric)                 \