/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x4c2-minmax-sse2-ld64.c | 195 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() local 200 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() 215 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() 235 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
|
D | 4x4c2-minmax-sse2-ld128.c | 195 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() local 200 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() 215 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() 235 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-xw-minmax-sse2.c | 188 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() local 193 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() 208 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() 228 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
|
D | 4x4c2-minmax-ssse3-ld128.c | 195 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128() local 215 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128() 235 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-ssse3-ld64.c | 195 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64() local 215 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64() 235 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
D | 4x4c2-xw-minmax-ssse3.c | 188 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3() local 208 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3() 228 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
|
/external/XNNPACK/src/qu8-gemm/ |
D | 4x4c2-minmax-sse2.c | 200 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2() local 205 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2() 220 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2() 240 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x4c2-minmax-sse2-ld128.c | 214 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() local 219 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() 234 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() 254 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-minmax-sse2-ld64.c | 214 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() local 219 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() 234 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() 254 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64()
|
D | 4x4c2-minmax-ssse3-ld128.c | 214 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128() local 234 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128() 254 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-ssse3-ld64.c | 214 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64() local 234 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64() 254 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
/external/XNNPACK/src/qu8-igemm/ |
D | 4x4c2-minmax-sse2.c | 184 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2() local 189 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2() 204 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2() 224 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
|