/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x4c8-xw-minmax-ssse3.c | 139 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3() local 143 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
|
D | 3x4c8-minmax-sse2-ld128.c | 141 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local 145 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-xw-minmax-sse2.c | 139 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2() local 143 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
|
D | 3x4c8-minmax-sse2-ld64.c | 143 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local 147 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 3x4c8-minmax-ssse3-ld128.c | 141 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128() local 145 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-ssse3-ld64.c | 143 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64() local 147 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 4x4c2-minmax-ssse3-ld128.c | 214 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128() local 219 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-sse2-ld64.c | 214 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() local 219 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
|
D | 4x4c2-minmax-ssse3-ld64.c | 214 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64() local 219 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
D | 4x4c2-xw-minmax-ssse3.c | 207 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3() local 212 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
|
D | 4x4c2-minmax-sse2-ld128.c | 214 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() local 219 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-xw-minmax-sse2.c | 207 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() local 212 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x4c8-minmax-ssse3-ld64.c | 160 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64() local 164 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 3x4c8-minmax-ssse3-ld128.c | 158 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128() local 162 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-sse2-ld128.c | 158 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128() local 162 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-minmax-sse2-ld64.c | 160 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64() local 164 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 4x4c2-minmax-sse2-ld128.c | 233 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() local 238 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-minmax-ssse3-ld128.c | 233 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128() local 238 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-ssse3-ld64.c | 233 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64() local 238 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
D | 4x4c2-minmax-sse2-ld64.c | 233 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() local 238 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64()
|
/external/XNNPACK/src/qu8-gemm/ |
D | 4x4c2-minmax-sse2.c | 219 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2() local 224 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
|
/external/XNNPACK/src/qu8-igemm/ |
D | 4x4c2-minmax-sse2.c | 203 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0)); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2() local 208 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
|