fix(debug): fix algo AlgoInt8x8x* performance

issue caused by store overhead

GitOrigin-RevId: 9a7c0dabfa7d5e5a45509ef1b890708810b5ae8e

dnn/src/x86/matrix_mul/int8/kernel_avx2_4x16x2.h

@@ -19,10 +19,10 @@ namespace x86 {
 namespace matmul_avx2_4x16x2 {
 
 template <typename CType>
-DNN_AVX2_TARGET void store_overflow(void* ptr, __m256i a);
+DNN_AVX2_TARGET void inline store_overflow(void* ptr, __m256i a);
 
 template <>
-void store_overflow<int16_t>(void* ptr, __m256i a) {
+void inline store_overflow<int16_t>(void* ptr, __m256i a) {
     static __m256i idx = _mm256_setr_epi32(0, 2, 4, 6, 0, 0, 0, 0);
     a = _mm256_shufflelo_epi16(a, 0x08);
     a = _mm256_shufflehi_epi16(a, 0x08);
@@ -31,15 +31,15 @@ void store_overflow<int16_t>(void* ptr, __m256i a) {
 }
 
 template <>
-void store_overflow<int32_t>(void* ptr, __m256i a) {
+void inline store_overflow<int32_t>(void* ptr, __m256i a) {
     _mm256_storeu_si256((__m256i*)(ptr), a);
 }
 
 template <typename CType>
-DNN_AVX2_TARGET void store_overflow(void* ptr, __m256i a, int remain);
+DNN_AVX2_TARGET void inline store_overflow(void* ptr, __m256i a, int remain);
 
 template <>
-void store_overflow<int16_t>(void* ptr, __m256i a, int remain) {
+void inline store_overflow<int16_t>(void* ptr, __m256i a, int remain) {
     __m128i mask = _mm_continue_mask(remain * sizeof(int16_t));
     static __m256i idx = _mm256_setr_epi32(0, 2, 4, 6, 0, 0, 0, 0);
     a = _mm256_shufflelo_epi16(a, 0x08);
@@ -50,7 +50,7 @@ void store_overflow<int16_t>(void* ptr, __m256i a, int remain) {
 }
 
 template <>
-void store_overflow<int32_t>(void* ptr, __m256i a, int remain) {
+void inline store_overflow<int32_t>(void* ptr, __m256i a, int remain) {
     __m256i mask = _m256_continue_mask(remain);
     _mm256_maskstore_epi32(reinterpret_cast<int32_t*>(ptr), mask, a);
 }
@@ -866,4 +866,4 @@ static inline void gemm_s8s8s32_avx2_4x16x2_pack_at(
 #endif
 #undef DNN_AVX2_TARGET
 
-// vim: syntax=cpp.doxygen
\ No newline at end of file
+// vim: syntax=cpp.doxygen

dnn/src/x86/matrix_mul/int8/kernel_sse_4x8x2.h

@@ -18,25 +18,25 @@ namespace matmul_sse_4x8x2 {
 
 template <typename CType>
 MEGDNN_ATTRIBUTE_TARGET("sse4.1")
-void store_overflow(void* ptr, __m128i a);
+void inline store_overflow(void* ptr, __m128i a);
 
 template <>
-void store_overflow<int16_t>(void* ptr, __m128i a) {
+void inline store_overflow<int16_t>(void* ptr, __m128i a) {
     a = _mm_shufflelo_epi16(a, 0x08);
     a = _mm_shufflehi_epi16(a, 0x08);
     a = _mm_shuffle_epi32(a, 0x08);
     _mm_storel_epi64((__m128i*)ptr, a);
 }
 template <>
-void store_overflow<int32_t>(void* ptr, __m128i a) {
+void inline store_overflow<int32_t>(void* ptr, __m128i a) {
     _mm_storeu_si128((__m128i*)(ptr), a);
 }
 template <typename CType>
 MEGDNN_ATTRIBUTE_TARGET("sse4.1")
-void store_overflow(void* ptr, __m128i a, int remain);
+void inline store_overflow(void* ptr, __m128i a, int remain);
 
 template <>
-void store_overflow<int16_t>(void* ptr, __m128i a, int remain) {
+void inline store_overflow<int16_t>(void* ptr, __m128i a, int remain) {
     __m128i mask = _mm_continue_mask(remain * sizeof(int16_t));
     a = _mm_shufflelo_epi16(a, 0x08);
     a = _mm_shufflehi_epi16(a, 0x08);
@@ -44,7 +44,7 @@ void store_overflow<int16_t>(void* ptr, __m128i a, int remain) {
     _mm_maskmoveu_si128(a, mask, reinterpret_cast<char*>(ptr));
 }
 template <>
-void store_overflow<int32_t>(void* ptr, __m128i a, int remain) {
+void inline store_overflow<int32_t>(void* ptr, __m128i a, int remain) {
     __m128i mask = _mm_continue_mask(remain * sizeof(int32_t));
     _mm_maskmoveu_si128(a, mask, reinterpret_cast<char*>(ptr));
 }
@@ -661,4 +661,4 @@ static inline void gemm_s8s8s32_sse_4x8x2_pack_at(
 }  // namespace x86
 }  // namespace megdnn
 
-// vim: syntax=cpp.doxygen
\ No newline at end of file
+// vim: syntax=cpp.doxygen