| /dports/graphics/vapoursynth-waifu2x-ncnn-vulkan/vapoursynth-waifu2x-ncnn-vulkan-r4/deps/ncnn/src/layer/x86/ |
| H A D | convolution_2x2_pack8.h | 76 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
82 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
103 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
119 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
125 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
143 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
269 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
275 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
302 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
308 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
|
| H A D | convolution_2x2_pack8_fp16.h | 257 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
263 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
284 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
300 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
306 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
324 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
450 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
456 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
483 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
489 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
|
| H A D | convolution_3x3_pack8.h | 71 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
82 _sum00 = _mm256_fmadd_ps(_r006, _k06, _sum00); in conv3x3s1_pack8_avx()
100 _sum10 = _mm256_fmadd_ps(_r016, _k06, _sum10); in conv3x3s1_pack8_avx()
456 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
467 _sum0 = _mm256_fmadd_ps(_r006, _k06, _sum0); in conv3x3s1_pack8_avx()
|
| /dports/graphics/waifu2x-ncnn-vulkan/waifu2x-ncnn-vulkan-20210521/src/ncnn/src/layer/x86/ |
| H A D | convolution_2x2_pack8.h | 76 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
82 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
103 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
119 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
125 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
143 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
269 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
275 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
302 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
308 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
|
| H A D | convolution_2x2_pack8_fp16.h | 257 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
263 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
284 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
300 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
306 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
324 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
450 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
456 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
483 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
489 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
|
| H A D | convolution_3x3_pack8.h | 71 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
82 _sum00 = _mm256_fmadd_ps(_r006, _k06, _sum00); in conv3x3s1_pack8_avx()
100 _sum10 = _mm256_fmadd_ps(_r016, _k06, _sum10); in conv3x3s1_pack8_avx()
456 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
467 _sum0 = _mm256_fmadd_ps(_r006, _k06, _sum0); in conv3x3s1_pack8_avx()
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/x86/ |
| H A D | convolution_2x2_pack8.h | 76 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
82 _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
103 _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
119 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
125 _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
143 _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
269 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
275 _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
302 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
308 _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
|
| H A D | convolution_2x2_pack8_fp16.h | 257 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
263 _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
284 _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
300 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
306 _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
324 _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
450 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
456 _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
483 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
489 _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
|
| H A D | convolution_3x3_pack8.h | 71 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
82 _sum00 = _mm256_comp_fmadd_ps(_r006, _k06, _sum00); in conv3x3s1_pack8_avx()
100 _sum10 = _mm256_comp_fmadd_ps(_r016, _k06, _sum10); in conv3x3s1_pack8_avx()
456 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
467 _sum0 = _mm256_comp_fmadd_ps(_r006, _k06, _sum0); in conv3x3s1_pack8_avx()
|
| /dports/benchmarks/vkpeak/vkpeak-20210430/ncnn/src/layer/x86/ |
| H A D | convolution_2x2_pack8.h | 76 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
82 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
103 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
119 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
125 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
143 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
269 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
275 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
302 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
308 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
|
| H A D | convolution_2x2_pack8_fp16.h | 257 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
263 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
284 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
300 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
306 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
324 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
450 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
456 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
483 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
489 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
|
| H A D | convolution_3x3_pack8.h | 71 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
82 _sum00 = _mm256_fmadd_ps(_r006, _k06, _sum00); in conv3x3s1_pack8_avx()
100 _sum10 = _mm256_fmadd_ps(_r016, _k06, _sum10); in conv3x3s1_pack8_avx()
456 __m256 _k06 = _mm256_loadu_ps(kptr + 48); in conv3x3s1_pack8_avx() local
467 _sum0 = _mm256_fmadd_ps(_r006, _k06, _sum0); in conv3x3s1_pack8_avx()
|
| /dports/graphics/realsr-ncnn-vulkan/realsr-ncnn-vulkan-20210210/src/ncnn/src/layer/x86/ |
| H A D | convolution_2x2_pack8.h | 76 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
82 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
103 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
119 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
125 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_pack8_avx()
143 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_pack8_avx()
269 __m256 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx() local
275 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
302 _k06 = _mm256_loadu_ps(kptr + 16); in conv2x2s1_pack8_avx()
308 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_pack8_avx()
|
| H A D | convolution_2x2_pack8_fp16.h | 257 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
263 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
284 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
300 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
306 _sum0 = _mm256_fmadd_ps(_k06, _r06, _sum0); in conv2x2s1_fp16_pack8_avx()
324 _sum1 = _mm256_fmadd_ps(_k06, _r06, _sum1); in conv2x2s1_fp16_pack8_avx()
450 __m256 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx() local
456 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
483 _k06 = loadfp16(kptr + 16); in conv2x2s1_fp16_pack8_avx()
489 _sum = _mm256_fmadd_ps(_k06, _r06, _sum); in conv2x2s1_fp16_pack8_avx()
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/riscv/ |
| H A D | convolution_7x7_pack1ton.h | 129 _sum0 = vfmacc_vf_f32m1(_sum0, r0[6], _k06, vl); in conv7x7s2_pack1ton_rvv()
130 _sum1 = vfmacc_vf_f32m1(_sum1, r0[8], _k06, vl); in conv7x7s2_pack1ton_rvv()
131 _sum2 = vfmacc_vf_f32m1(_sum2, r0[10], _k06, vl); in conv7x7s2_pack1ton_rvv()
132 _sum3 = vfmacc_vf_f32m1(_sum3, r0[12], _k06, vl); in conv7x7s2_pack1ton_rvv()
133 _sum4 = vfmacc_vf_f32m1(_sum4, r0[14], _k06, vl); in conv7x7s2_pack1ton_rvv()
134 _sum5 = vfmacc_vf_f32m1(_sum5, r0[16], _k06, vl); in conv7x7s2_pack1ton_rvv()
135 _sum6 = vfmacc_vf_f32m1(_sum6, r0[18], _k06, vl); in conv7x7s2_pack1ton_rvv()
136 _sum7 = vfmacc_vf_f32m1(_sum7, r0[20], _k06, vl); in conv7x7s2_pack1ton_rvv()
600 _sum0 = vfmacc_vf_f32m1(_sum0, r0[6], _k06, vl); in conv7x7s2_pack1ton_rvv()
601 _sum1 = vfmacc_vf_f32m1(_sum1, r0[8], _k06, vl); in conv7x7s2_pack1ton_rvv()
[all …]
|
| H A D | convolution_7x7_pack1ton_fp16s.h | 129 _sum0 = vfmacc_vf_f16m1(_sum0, r0[6], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
130 _sum1 = vfmacc_vf_f16m1(_sum1, r0[8], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
131 _sum2 = vfmacc_vf_f16m1(_sum2, r0[10], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
132 _sum3 = vfmacc_vf_f16m1(_sum3, r0[12], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
133 _sum4 = vfmacc_vf_f16m1(_sum4, r0[14], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
134 _sum5 = vfmacc_vf_f16m1(_sum5, r0[16], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
135 _sum6 = vfmacc_vf_f16m1(_sum6, r0[18], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
136 _sum7 = vfmacc_vf_f16m1(_sum7, r0[20], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
600 _sum0 = vfmacc_vf_f16m1(_sum0, r0[6], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
601 _sum1 = vfmacc_vf_f16m1(_sum1, r0[8], _k06, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
[all …]
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/mips/ |
| H A D | convolution_7x7_pack1to4.h | 70 v4f32 _k06 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0); in conv7x7s2_pack1to4_msa() local
116 _sum0 = __msa_fmadd_w(_sum0, _r06, _k06); in conv7x7s2_pack1to4_msa()
117 _sum1 = __msa_fmadd_w(_sum1, _r08, _k06); in conv7x7s2_pack1to4_msa()
118 _sum2 = __msa_fmadd_w(_sum2, _r0a, _k06); in conv7x7s2_pack1to4_msa()
119 _sum3 = __msa_fmadd_w(_sum3, _r0c, _k06); in conv7x7s2_pack1to4_msa()
488 v4f32 _k06 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0); in conv7x7s2_pack1to4_msa() local
501 _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0n, 2), _k06); in conv7x7s2_pack1to4_msa()
|
| /dports/graphics/waifu2x-ncnn-vulkan/waifu2x-ncnn-vulkan-20210521/src/ncnn/src/layer/arm/ |
| H A D | convolution_3x3.h | 64 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s1_neon() local
220 "w"(_k06), // %20 in conv3x3s1_neon()
364 "w"(_k06), // %20 in conv3x3s1_neon()
384 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
391 _sum0n = vmlaq_f32(_sum0n, _r30, _k06); in conv3x3s1_neon()
582 "w"(_k06), // %14 in conv3x3s1_neon()
672 "w"(_k06), // %14 in conv3x3s1_neon()
691 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
7999 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s2_neon() local
8111 "w"(_k06), // %14 in conv3x3s2_neon()
[all …]
|
| /dports/benchmarks/vkpeak/vkpeak-20210430/ncnn/src/layer/arm/ |
| H A D | convolution_3x3.h | 64 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s1_neon() local
220 "w"(_k06), // %20 in conv3x3s1_neon()
364 "w"(_k06), // %20 in conv3x3s1_neon()
384 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
391 _sum0n = vmlaq_f32(_sum0n, _r30, _k06); in conv3x3s1_neon()
582 "w"(_k06), // %14 in conv3x3s1_neon()
672 "w"(_k06), // %14 in conv3x3s1_neon()
691 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
7999 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s2_neon() local
8111 "w"(_k06), // %14 in conv3x3s2_neon()
[all …]
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/arm/ |
| H A D | convolution_3x3.h | 64 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s1_neon() local
220 "w"(_k06), // %20 in conv3x3s1_neon()
364 "w"(_k06), // %20 in conv3x3s1_neon()
384 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
391 _sum0n = vmlaq_f32(_sum0n, _r30, _k06); in conv3x3s1_neon()
582 "w"(_k06), // %14 in conv3x3s1_neon()
672 "w"(_k06), // %14 in conv3x3s1_neon()
691 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
7999 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s2_neon() local
8111 "w"(_k06), // %14 in conv3x3s2_neon()
[all …]
|
| /dports/graphics/vapoursynth-waifu2x-ncnn-vulkan/vapoursynth-waifu2x-ncnn-vulkan-r4/deps/ncnn/src/layer/arm/ |
| H A D | convolution_3x3.h | 64 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s1_neon() local
220 "w"(_k06), // %20 in conv3x3s1_neon()
364 "w"(_k06), // %20 in conv3x3s1_neon()
384 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
391 _sum0n = vmlaq_f32(_sum0n, _r30, _k06); in conv3x3s1_neon()
582 "w"(_k06), // %14 in conv3x3s1_neon()
672 "w"(_k06), // %14 in conv3x3s1_neon()
691 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
7999 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s2_neon() local
8111 "w"(_k06), // %14 in conv3x3s2_neon()
[all …]
|
| /dports/graphics/realsr-ncnn-vulkan/realsr-ncnn-vulkan-20210210/src/ncnn/src/layer/arm/ |
| H A D | convolution_3x3.h | 64 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s1_neon() local
220 "w"(_k06), // %20 in conv3x3s1_neon()
364 "w"(_k06), // %20 in conv3x3s1_neon()
384 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
391 _sum0n = vmlaq_f32(_sum0n, _r30, _k06); in conv3x3s1_neon()
582 "w"(_k06), // %14 in conv3x3s1_neon()
672 "w"(_k06), // %14 in conv3x3s1_neon()
691 _sum0 = vmlaq_f32(_sum0, _r20, _k06); in conv3x3s1_neon()
7999 float32x4_t _k06 = vld1q_f32(k0 + 6); in conv3x3s2_neon() local
8111 "w"(_k06), // %14 in conv3x3s2_neon()
[all …]
|