/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x4c8-minmax-xop-ld128.c | 144 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128() local 149 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128()
|
D | 3x4c8-minmax-sse41-ld128.c | 139 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128() local 144 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128()
|
D | 3x4c8-minmax-xop-ld64.c | 146 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64() local 151 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64()
|
D | 3x4c8-xw-minmax-xop.c | 142 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop() local 147 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop()
|
D | 3x4c8-minmax-sse41-ld64.c | 141 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64() local 146 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64()
|
D | 3x4c8-xw-minmax-sse41.c | 137 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41() local 142 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41()
|
D | 3x4c8-xw-minmax-ssse3.c | 147 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3() local 170 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
|
D | 3x4c8-minmax-sse2-ld64.c | 151 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local 174 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 3x4c8-minmax-sse2-ld128.c | 149 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local 172 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-minmax-ssse3-ld128.c | 149 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128() local 172 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-ssse3-ld64.c | 151 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64() local 174 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 3x4c8-xw-minmax-sse2.c | 147 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2() local 170 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
|
D | 4x4c2-minmax-xop-ld128.c | 216 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128() local 223 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128()
|
D | 4x4c2-minmax-xop-ld64.c | 216 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64() local 223 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64()
|
D | 4x4c2-xw-minmax-xop.c | 209 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop() local 216 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop()
|
D | 4x4c2-xw-minmax-sse41.c | 204 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse41() local 211 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse41()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x4c8-minmax-xop-ld64.c | 163 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64() local 168 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64()
|
D | 3x4c8-minmax-xop-ld128.c | 161 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128() local 166 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128()
|
D | 3x4c8-minmax-sse41-ld64.c | 158 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64() local 163 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64()
|
D | 3x4c8-minmax-sse41-ld128.c | 156 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128() local 161 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128()
|
D | 3x4c8-minmax-ssse3-ld64.c | 168 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64() local 191 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
D | 3x4c8-minmax-sse2-ld128.c | 166 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128() local 189 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
|
D | 3x4c8-minmax-ssse3-ld128.c | 166 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128() local 189 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
D | 3x4c8-minmax-sse2-ld64.c | 168 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64() local 191 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
|
D | 4x4c2-minmax-sse41-ld128.c | 230 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld128() local 237 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse41_ld128()
|