/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-sse41-c8-acc2.c | 103 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 107 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() 170 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 174 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c8-acc2.c | 111 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 117 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() 191 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 197 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
D | 7x-minmax-sse2-c8-acc2.c | 114 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 120 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 197 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 203 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 126 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 134 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() 203 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 207 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7p7x-minmax-sse41-c8-acc2.c | 198 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 202 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() 266 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 270 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 149 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 161 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() 238 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 242 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 142 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 152 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() 236 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 242 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 147 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 157 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 244 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 250 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse2-c8-acc2.c | 225 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 231 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 309 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 315 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 222 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 228 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() 303 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 309 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 173 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 187 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() 283 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 289 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 180 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 194 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 293 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 299 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 255 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 263 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 333 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 337 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 303 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 313 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 398 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 404 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 308 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 318 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 406 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 412 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 388 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 400 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 478 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 482 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 483 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 497 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 597 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 603 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 476 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 490 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 587 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 593 …_mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|