Searched refs:vk3xGHIJKLMN (Results 1 – 8 of 8) sorted by relevance
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up24x9-minmax-neon-mul16.c | 144 …const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local 150 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() 151 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
|
D | up32x9-minmax-neon-mul16.c | 158 …const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local 166 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() 167 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
|
D | up24x9-minmax-avx2-mul32.c | 129 …const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local 134 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
|
D | up32x9-minmax-avx2-mul32.c | 139 …const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local 146 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
|
D | up24x9-minmax-wasmsimd-mul16.c | 153 …const v128_t vk3xGHIJKLMN = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local 158 const v128_t vprod3xGHIJKLMN = wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 183 …const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 191 …st __m128i vxk3xGHIJKLMN = _mm_unpacklo_epi8(vk3xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vk3… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 183 …const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 191 …st __m128i vxk3xGHIJKLMN = _mm_unpacklo_epi8(vk3xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vk3… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|
D | up24x9-minmax-sse41-mul16.c | 188 …const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 189 const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|