| /external/XNNPACK/src/qc8-dwconv/gen/ |
| D | up16x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 178 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local 299 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local
|
| D | up16x9-minmax-fp32-avx2-mul16-vpmovsx.c | 175 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local 281 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local
|
| D | up16x9-minmax-fp32-avx2-mul16-vpunpck.c | 190 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local 311 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local
|
| D | up32x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 228 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local 372 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local
|
| D | up16x3-minmax-fp32-avx2-mul32.c | 91 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx2_mul32() local
|
| D | up32x9-minmax-fp32-avx2-mul16-vpmovsx.c | 233 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx() local 362 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx() local
|
| D | up32x9-minmax-fp32-avx2-mul16-vpunpck.c | 252 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck() local 396 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck() local
|
| D | up16x9-minmax-fp32-avx2-mul32.c | 175 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32() local
|
| D | up16x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 394 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local 659 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local
|
| D | up16x25-minmax-fp32-avx2-mul16-vpmovsx.c | 399 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local 633 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local
|
| D | up24x9-minmax-fp32-avx2-mul32.c | 204 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local
|
| D | up16x25-minmax-fp32-avx2-mul16-vpunpck.c | 430 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local 695 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local
|
| D | up32x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 516 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local 820 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local
|
| D | up32x9-minmax-fp32-avx2-mul32.c | 233 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local
|
| D | up32x25-minmax-fp32-avx2-mul16-vpmovsx.c | 553 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local 826 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
|
| D | up32x25-minmax-fp32-avx2-mul16-vpunpck.c | 588 … const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local 892 …const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local
|
| D | up16x25-minmax-fp32-avx2-mul32.c | 399 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local
|
| D | up24x25-minmax-fp32-avx2-mul32.c | 476 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx2_mul32() local
|
| D | up32x25-minmax-fp32-avx2-mul32.c | 553 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
|
| /external/XNNPACK/src/f16-vmulcaddc/gen/ |
| D | c16-minmax-fma3-2x.c | 53 const __m256 vscale89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8))); in xnn_f16_vmulcaddc_minmax_ukernel_c16__fma3_2x() local
|
| D | c16-minmax-neonfp16arith-2x.c | 52 const float16x8_t vscale89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local
|
| /external/XNNPACK/src/amalgam/ |
| D | avx2.c | 2658 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local 2956 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx2_mul32() local 3206 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32() local
|