Home
last modified time | relevance | path

Searched refs:vprod2x13 (Results 1 – 25 of 54) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D3x4c8-minmax-xop-ld128.c138 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128() local
145 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128()
D3x4c8-minmax-sse41-ld128.c133 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128() local
140 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128()
D3x4c8-minmax-xop-ld64.c140 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64() local
147 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64()
D3x4c8-xw-minmax-xop.c136 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop() local
143 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop()
D3x4c8-minmax-sse41-ld64.c135 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64() local
142 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64()
D3x4c8-xw-minmax-sse41.c131 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41() local
138 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41()
D4x4c2-minmax-xop-ld128.c209 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128() local
217 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128()
D4x4c2-minmax-xop-ld64.c209 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64() local
217 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64()
D4x4c2-xw-minmax-xop.c202 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop() local
210 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop()
D4x4c2-xw-minmax-sse41.c197 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse41() local
205 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse41()
D4x4c2-minmax-sse41-ld128.c204 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld128() local
212 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld128()
D4x4c2-minmax-sse41-ld64.c204 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld64() local
212 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld64()
D3x4c8-xw-minmax-ssse3.c159 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3() local
163 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
D3x4c8-minmax-sse2-ld64.c163 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local
167 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
D3x4c8-minmax-sse2-ld128.c161 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local
165 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
D3x4c8-minmax-ssse3-ld128.c161 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128() local
165 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
/external/XNNPACK/src/qs8-igemm/gen/
D3x4c8-minmax-xop-ld64.c157 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64() local
164 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64()
D3x4c8-minmax-xop-ld128.c155 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128() local
162 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128()
D3x4c8-minmax-sse41-ld64.c152 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64() local
159 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64()
D3x4c8-minmax-sse41-ld128.c150 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128() local
157 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128()
D4x4c2-minmax-sse41-ld128.c223 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld128() local
231 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld128()
D4x4c2-minmax-sse41-ld64.c223 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld64() local
231 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld64()
D4x4c2-minmax-xop-ld128.c228 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld128() local
236 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld128()
D4x4c2-minmax-xop-ld64.c228 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding); in xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld64() local
236 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13); in xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld64()
D3x4c8-minmax-ssse3-ld64.c180 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64() local
184 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()

123