/external/XNNPACK/src/f16-vmulcaddc/gen/ |
D | c16-minmax-neonfp16arith-2x.c | 51 const float16x8_t vscale01234567 = vld1q_f16(w); w += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 83 const float16x8_t vscale01234567 = vld1q_f16(w); in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 104 const float16x8_t vscale01234567 = vld1q_f16(w); in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local
|
D | c8-minmax-neonfp16arith-2x.c | 51 const float16x8_t vscale01234567 = vld1q_f16(w); in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local 72 const float16x8_t vscale01234567 = vld1q_f16(w); in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local
|
D | c16-minmax-fma3-2x.c | 51 const __m256 vscale01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w)); in xnn_f16_vmulcaddc_minmax_ukernel_c16__fma3_2x() local
|
/external/XNNPACK/src/f16-vmulcaddc/ |
D | neonfp16arith.c.in | 87 const float16x8_t vscale01234567 = vld1q_f16(w); local
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up8x9-minmax-fp32-avx2-mul32.c | 144 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 216 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
D | up16x9-minmax-fp32-avx2-mul32.c | 173 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32() local 260 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32() local
|
D | up16x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 176 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local 297 …const __m256 vscale01234567 = _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 | up24x9-minmax-fp32-avx2-mul32.c | 202 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local 297 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local
|
D | up16x9-minmax-fp32-avx2-mul16-vpmovsx.c | 173 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local 279 …const __m256 vscale01234567 = _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 | 188 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local 309 …const __m256 vscale01234567 = _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 | 226 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local 370 …const __m256 vscale01234567 = _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 | up32x9-minmax-fp32-avx2-mul32.c | 231 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local 330 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local
|
D | up8x25-minmax-fp32-avx2-mul32.c | 320 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local 472 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
|
D | up32x9-minmax-fp32-avx2-mul16-vpmovsx.c | 231 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx() local 360 …const __m256 vscale01234567 = _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 | 250 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck() local 394 …const __m256 vscale01234567 = _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 | up16x25-minmax-fp32-avx2-mul32.c | 397 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local 580 …const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t)… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 1x8c8-xw-minmax-fp32-avx2.c | 92 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2() local
|
D | 1x8c8-minmax-fp32-avx2.c | 96 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__avx2() local
|
D | 2x8c8-xw-minmax-fp32-avx2.c | 114 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2() local
|
D | 2x8c8-minmax-fp32-avx2.c | 118 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__avx2() local
|
D | 3x8c8-xw-minmax-fp32-avx2.c | 136 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2() local
|
D | 3x8c8-minmax-fp32-avx2.c | 140 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_gemm_minmax_fp32_ukernel_3x8c8__avx2() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 1x8c8-minmax-fp32-avx2.c | 109 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__avx2() local
|
D | 2x8c8-minmax-fp32-avx2.c | 133 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__avx2() local
|
D | 3x8c8-minmax-fp32-avx2.c | 157 const __m256 vscale01234567 = _mm256_load_ps(w); in xnn_qc8_igemm_minmax_fp32_ukernel_3x8c8__avx2() local
|