Lines Matching refs:vaccOPQRSTUV
167 …__m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
183 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
198 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
213 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod2xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
228 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod3xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
243 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
258 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod5xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
273 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod6xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
288 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod7xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
303 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod8xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
318 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod9xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
333 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod10xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
348 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod11xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
363 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
378 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
393 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod14xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
408 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
423 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
438 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod17xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
453 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod18xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
468 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod19xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
483 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod20xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
498 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod21xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
513 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod22xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
528 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod23xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
543 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod24xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
550 __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
567 vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
571 …GHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()