| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-sse41-c16-acc2.c | 128 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
135 …const __m128i vabsout89AB = _mm_blend_epi16(vabsout8A, _mm_shuffle_epi32(vabsout9B, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 144 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
154 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 151 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
162 …const __m128i vabsout89AB = _mm_blend_epi16(vabsout8A, _mm_shuffle_epi32(vabsout9B, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 149 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
159 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 257 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
264 …const __m128i vabsout89AB = _mm_blend_epi16(vabsout8A, _mm_shuffle_epi32(vabsout9B, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-sse2-c24-acc2.c | 182 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
196 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
| D | 7x-minmax-ssse3-c24-acc2.c | 175 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
189 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 305 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
315 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 310 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
320 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 390 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
401 …const __m128i vabsout89AB = _mm_blend_epi16(vabsout8A, _mm_shuffle_epi32(vabsout9B, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
| D | 7p7x-minmax-sse2-c24-acc2.c | 485 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
499 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 478 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
492 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout8A), _mm_castsi128_ps(vabsout9B), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|