| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-avx-acc2.c | 290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
418 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
|
| D | up8x25-minmax-fma3-acc2.c | 290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
418 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-fma3.c | 290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
416 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-avx.c | 290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
416 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
|
| D | up16x25-minmax-fma3.c | 351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
531 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
657 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
|
| D | up16x25-minmax-fma3-acc2.c | 351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
534 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
662 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
|
| D | up16x25-minmax-avx-acc2.c | 351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
534 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
662 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
|
| D | up16x25-minmax-avx.c | 351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
531 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
657 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up8x25-minmax-neonfp16arith.c | 250 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
360 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
|
| D | up8x25-minmax-neonfp16arith-acc2.c | 250 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
362 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
|
| D | up16x25-minmax-neonfp16arith-acc2.c | 311 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
442 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
554 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
|
| D | up16x25-minmax-neonfp16arith.c | 311 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
439 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
549 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
|