/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x4c8-xw-minmax-ssse3.c | 131 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3() local 151 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
|
D | 3x4c8-minmax-sse2-ld128.c | 133 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local 153 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-xw-minmax-sse2.c | 131 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2() local 151 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
|
D | 3x4c8-minmax-sse2-ld64.c | 135 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local 155 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 3x4c8-minmax-ssse3-ld128.c | 133 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128() local 153 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-ssse3-ld64.c | 135 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64() local 155 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 4x4c2-minmax-ssse3-ld128.c | 204 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128() local 229 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-sse2-ld64.c | 204 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() local 229 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
|
D | 4x4c2-minmax-ssse3-ld64.c | 204 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64() local 229 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
D | 4x4c2-xw-minmax-ssse3.c | 197 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3() local 222 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
|
D | 4x4c2-minmax-sse2-ld128.c | 204 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() local 229 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-xw-minmax-sse2.c | 197 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() local 222 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x4c8-minmax-ssse3-ld64.c | 152 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64() local 172 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 3x4c8-minmax-ssse3-ld128.c | 150 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128() local 170 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-sse2-ld128.c | 150 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128() local 170 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-minmax-sse2-ld64.c | 152 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64() local 172 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 4x4c2-minmax-sse2-ld128.c | 223 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() local 248 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
|
D | 4x4c2-minmax-ssse3-ld128.c | 223 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128() local 248 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
D | 4x4c2-minmax-ssse3-ld64.c | 223 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64() local 248 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
D | 4x4c2-minmax-sse2-ld64.c | 223 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() local 248 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64()
|