Home
last modified time | relevance | path

Searched refs:vq31prod1x13 (Results 1 – 25 of 63) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D2x4c8-xw-minmax-xop.c118 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__xop() local
121 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__xop()
D2x4c8-minmax-sse41-ld64.c117 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld64() local
120 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld64()
D2x4c8-xw-minmax-sse41.c113 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse41() local
116 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse41()
D2x4c8-minmax-xop-ld128.c120 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld128() local
123 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld128()
D2x4c8-minmax-sse41-ld128.c115 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld128() local
118 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld128()
D2x4c8-minmax-xop-ld64.c122 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64() local
125 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64()
D2x4c8-minmax-ssse3-ld128.c134 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128() local
139 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128()
D2x4c8-xw-minmax-sse2.c132 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2() local
137 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
D2x4c8-minmax-ssse3-ld64.c136 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64() local
141 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64()
D2x4c8-xw-minmax-ssse3.c132 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__ssse3() local
137 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__ssse3()
D3x4c8-minmax-xop-ld128.c143 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128() local
148 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128()
D3x4c8-minmax-sse41-ld128.c138 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128() local
143 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128()
D3x4c8-minmax-xop-ld64.c145 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64() local
150 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64()
D2x4c8-minmax-sse2-ld128.c134 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128() local
139 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
D2x4c8-minmax-sse2-ld64.c136 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64() local
141 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
D3x4c8-xw-minmax-xop.c141 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop() local
146 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop()
/external/XNNPACK/src/qs8-igemm/gen/
D2x4c8-minmax-xop-ld64.c137 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x4c8__xop_ld64() local
140 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_2x4c8__xop_ld64()
D2x4c8-minmax-xop-ld128.c135 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x4c8__xop_ld128() local
138 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_2x4c8__xop_ld128()
D2x4c8-minmax-sse41-ld64.c132 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_ld64() local
135 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_ld64()
D2x4c8-minmax-sse41-ld128.c130 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_ld128() local
133 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_ld128()
D2x4c8-minmax-ssse3-ld128.c149 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld128() local
154 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld128()
D2x4c8-minmax-ssse3-ld64.c151 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld64() local
156 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld64()
D3x4c8-minmax-xop-ld64.c162 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64() local
167 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64()
D2x4c8-minmax-sse2-ld128.c149 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128() local
154 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
/external/XNNPACK/src/qu8-gemm/
D2x4c8-minmax-sse2.c145 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2() local
150 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()

123