| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-fma3-acc2.c | 314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
434 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-fma3.c | 314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
432 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-avx.c | 314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
432 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
|
| D | up8x25-minmax-avx-acc2.c | 314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
434 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
|
| D | up16x25-minmax-fma3.c | 387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
555 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
673 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
|
| D | up16x25-minmax-fma3-acc2.c | 387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
558 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
678 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
|
| D | up16x25-minmax-avx-acc2.c | 387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
558 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
678 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
|
| D | up16x25-minmax-avx.c | 387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
555 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
673 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up8x25-minmax-neonfp16arith-acc2.c | 266 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
378 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
|
| D | up8x25-minmax-neonfp16arith.c | 266 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
376 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
|
| D | up16x25-minmax-neonfp16arith-acc2.c | 339 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
458 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
570 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
|
| D | up16x25-minmax-neonfp16arith.c | 339 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
455 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
565 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
|