Enable SVE Support for IP Metric Computation in FP16 and NY functions

src/simd/distances_sve.cc

@@ -96,6 +96,37 @@ fp16_vec_L2sqr_sve(const knowhere::fp16* x, const knowhere::fp16* y, size_t d) {
     return svaddv_f32(pg_32, total_sum);
 }
 
+float
+fp16_vec_inner_product_sve(const knowhere::fp16* x, const knowhere::fp16* y, size_t d) {
+    svfloat32_t sum1 = svdup_f32(0.0f);
+    svfloat32_t sum2 = svdup_f32(0.0f);
+    size_t i = 0;
+
+    svbool_t pg_16 = svptrue_b16();
+    svbool_t pg_32 = svptrue_b32();
+
+    while (i < d) {
+        if (d - i < svcnth())
+            pg_16 = svwhilelt_b16(i, d);
+
+        svfloat16_t a_fp16 = svld1_f16(pg_16, reinterpret_cast<const __fp16*>(x + i));
+        svfloat16_t b_fp16 = svld1_f16(pg_16, reinterpret_cast<const __fp16*>(y + i));
+
+        svfloat32_t a_fp32_low = svcvt_f32_f16_z(pg_32, svtrn1_f16(a_fp16, a_fp16));
+        svfloat32_t a_fp32_high = svcvt_f32_f16_z(pg_32, svtrn2_f16(a_fp16, a_fp16));
+        svfloat32_t b_fp32_low = svcvt_f32_f16_z(pg_32, svtrn1_f16(b_fp16, b_fp16));
+        svfloat32_t b_fp32_high = svcvt_f32_f16_z(pg_32, svtrn2_f16(b_fp16, b_fp16));
+
+        sum1 = svmla_f32_m(pg_32, sum1, a_fp32_low, b_fp32_low);
+        sum2 = svmla_f32_m(pg_32, sum2, a_fp32_high, b_fp32_high);
+
+        i += svcnth();
+    }
+
+    svfloat32_t total_sum = svadd_f32_m(pg_32, sum1, sum2);
+    return svaddv_f32(pg_32, total_sum);
+}
+
 float
 fvec_L1_sve(const float* x, const float* y, size_t d) {
     svfloat32_t sum = svdup_f32(0.0f);
@@ -308,6 +339,14 @@ fvec_L2sqr_ny_sve(float* dis, const float* x, const float* y, size_t d, size_t n
     }
 }
 
+void
+fvec_inner_products_ny_sve(float* ip, const float* x, const float* y, size_t d, size_t ny) {
+    for (size_t i = 0; i < ny; ++i) {
+        ip[i] = fvec_inner_product_sve(x, y, d);
+        y += d;
+    }
+}
+
 }  // namespace faiss
 
 #endif

src/simd/distances_sve.h

@@ -28,6 +28,9 @@ fvec_inner_product_sve(const float* x, const float* y, size_t d);
 float
 fp16_vec_L2sqr_sve(const knowhere::fp16* x, const knowhere::fp16* y, size_t d);
 
+float
+fp16_vec_inner_product_sve(const knowhere::fp16* x, const knowhere::fp16* y, size_t d);
+
 float
 fvec_L1_sve(const float* x, const float* y, size_t d);
 
@@ -60,5 +63,8 @@ fvec_L2sqr_batch_4_sve(const float* x, const float* y0, const float* y1, const f
 void
 fvec_L2sqr_ny_sve(float* dis, const float* x, const float* y, size_t d, size_t ny);
 
+void
+fvec_inner_products_ny_sve(float* ip, const float* x, const float* y, size_t d, size_t ny);
+
 }  // namespace faiss
 #endif

src/simd/hook.cc

@@ -447,13 +447,13 @@ fvec_hook(std::string& simd_type) {
 
         fvec_inner_product = fvec_inner_product_sve;
         fvec_L2sqr_ny = fvec_L2sqr_ny_sve;
-        fvec_inner_products_ny = fvec_inner_products_ny_neon;
+        fvec_inner_products_ny = fvec_inner_products_ny_sve;
 
         ivec_inner_product = ivec_inner_product_neon;
         ivec_L2sqr = ivec_L2sqr_neon;
 
         // fp16
-        fp16_vec_inner_product = fp16_vec_inner_product_neon;
+        fp16_vec_inner_product = fp16_vec_inner_product_sve;
         fp16_vec_L2sqr = fp16_vec_L2sqr_sve;
         fp16_vec_norm_L2sqr = fp16_vec_norm_L2sqr_sve;