Home
last modified time | relevance | path

Searched refs:_k05 (Results 1 – 17 of 17) sorted by relevance

/dports/graphics/vapoursynth-waifu2x-ncnn-vulkan/vapoursynth-waifu2x-ncnn-vulkan-r4/deps/ncnn/src/layer/x86/
H A Dconvolution_2x2_pack8.h75 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
81 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
102 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
118 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
124 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
142 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
268 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
274 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
301 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
307 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
H A Dconvolution_2x2_pack8_fp16.h256 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
262 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
283 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
299 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
305 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
323 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
449 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
455 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
482 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
488 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
H A Dconvolution_3x3_pack8.h70 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
81 _sum01 = _mm256_fmadd_ps(_r005, _k05, _sum01); in conv3x3s1_pack8_avx()
99 _sum11 = _mm256_fmadd_ps(_r015, _k05, _sum11); in conv3x3s1_pack8_avx()
455 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
466 _sum1 = _mm256_fmadd_ps(_r005, _k05, _sum1); in conv3x3s1_pack8_avx()
/dports/graphics/waifu2x-ncnn-vulkan/waifu2x-ncnn-vulkan-20210521/src/ncnn/src/layer/x86/
H A Dconvolution_2x2_pack8.h75 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
81 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
102 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
118 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
124 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
142 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
268 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
274 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
301 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
307 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
H A Dconvolution_2x2_pack8_fp16.h256 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
262 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
283 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
299 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
305 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
323 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
449 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
455 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
482 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
488 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
H A Dconvolution_3x3_pack8.h70 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
81 _sum01 = _mm256_fmadd_ps(_r005, _k05, _sum01); in conv3x3s1_pack8_avx()
99 _sum11 = _mm256_fmadd_ps(_r015, _k05, _sum11); in conv3x3s1_pack8_avx()
455 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
466 _sum1 = _mm256_fmadd_ps(_r005, _k05, _sum1); in conv3x3s1_pack8_avx()
/dports/misc/ncnn/ncnn-20211208/src/layer/x86/
H A Dconvolution_2x2_pack8.h75 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
81 _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
102 _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
118 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
124 _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
142 _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
268 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
274 _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
301 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
307 _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
H A Dconvolution_2x2_pack8_fp16.h256 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
262 _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
283 _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
299 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
305 _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
323 _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
449 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
455 _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
482 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
488 _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
H A Dconvolution_3x3_pack8.h70 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
81 _sum01 = _mm256_comp_fmadd_ps(_r005, _k05, _sum01); in conv3x3s1_pack8_avx()
99 _sum11 = _mm256_comp_fmadd_ps(_r015, _k05, _sum11); in conv3x3s1_pack8_avx()
455 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
466 _sum1 = _mm256_comp_fmadd_ps(_r005, _k05, _sum1); in conv3x3s1_pack8_avx()
/dports/benchmarks/vkpeak/vkpeak-20210430/ncnn/src/layer/x86/
H A Dconvolution_2x2_pack8.h75 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
81 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
102 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
118 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
124 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
142 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
268 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
274 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
301 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
307 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
H A Dconvolution_2x2_pack8_fp16.h256 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
262 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
283 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
299 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
305 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
323 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
449 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
455 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
482 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
488 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
H A Dconvolution_3x3_pack8.h70 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
81 _sum01 = _mm256_fmadd_ps(_r005, _k05, _sum01); in conv3x3s1_pack8_avx()
99 _sum11 = _mm256_fmadd_ps(_r015, _k05, _sum11); in conv3x3s1_pack8_avx()
455 __m256 _k05 = _mm256_loadu_ps(kptr + 40); in conv3x3s1_pack8_avx() local
466 _sum1 = _mm256_fmadd_ps(_r005, _k05, _sum1); in conv3x3s1_pack8_avx()
/dports/graphics/realsr-ncnn-vulkan/realsr-ncnn-vulkan-20210210/src/ncnn/src/layer/x86/
H A Dconvolution_2x2_pack8.h75 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
81 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
102 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
118 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
124 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_pack8_avx()
142 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_pack8_avx()
268 __m256 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx() local
274 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
301 _k05 = _mm256_loadu_ps(kptr + 8); in conv2x2s1_pack8_avx()
307 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_pack8_avx()
H A Dconvolution_2x2_pack8_fp16.h256 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
262 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
283 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
299 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
305 _sum0 = _mm256_fmadd_ps(_k05, _r05, _sum0); in conv2x2s1_fp16_pack8_avx()
323 _sum1 = _mm256_fmadd_ps(_k05, _r05, _sum1); in conv2x2s1_fp16_pack8_avx()
449 __m256 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx() local
455 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
482 _k05 = loadfp16(kptr + 8); in conv2x2s1_fp16_pack8_avx()
488 _sum = _mm256_fmadd_ps(_k05, _r05, _sum); in conv2x2s1_fp16_pack8_avx()
/dports/misc/ncnn/ncnn-20211208/src/layer/riscv/
H A Dconvolution_7x7_pack1ton.h121 _sum0 = vfmacc_vf_f32m1(_sum0, r0[5], _k05, vl); in conv7x7s2_pack1ton_rvv()
122 _sum1 = vfmacc_vf_f32m1(_sum1, r0[7], _k05, vl); in conv7x7s2_pack1ton_rvv()
123 _sum2 = vfmacc_vf_f32m1(_sum2, r0[9], _k05, vl); in conv7x7s2_pack1ton_rvv()
124 _sum3 = vfmacc_vf_f32m1(_sum3, r0[11], _k05, vl); in conv7x7s2_pack1ton_rvv()
125 _sum4 = vfmacc_vf_f32m1(_sum4, r0[13], _k05, vl); in conv7x7s2_pack1ton_rvv()
126 _sum5 = vfmacc_vf_f32m1(_sum5, r0[15], _k05, vl); in conv7x7s2_pack1ton_rvv()
127 _sum6 = vfmacc_vf_f32m1(_sum6, r0[17], _k05, vl); in conv7x7s2_pack1ton_rvv()
128 _sum7 = vfmacc_vf_f32m1(_sum7, r0[19], _k05, vl); in conv7x7s2_pack1ton_rvv()
596 _sum0 = vfmacc_vf_f32m1(_sum0, r0[5], _k05, vl); in conv7x7s2_pack1ton_rvv()
597 _sum1 = vfmacc_vf_f32m1(_sum1, r0[7], _k05, vl); in conv7x7s2_pack1ton_rvv()
[all …]
H A Dconvolution_7x7_pack1ton_fp16s.h121 _sum0 = vfmacc_vf_f16m1(_sum0, r0[5], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
122 _sum1 = vfmacc_vf_f16m1(_sum1, r0[7], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
123 _sum2 = vfmacc_vf_f16m1(_sum2, r0[9], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
124 _sum3 = vfmacc_vf_f16m1(_sum3, r0[11], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
125 _sum4 = vfmacc_vf_f16m1(_sum4, r0[13], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
126 _sum5 = vfmacc_vf_f16m1(_sum5, r0[15], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
127 _sum6 = vfmacc_vf_f16m1(_sum6, r0[17], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
128 _sum7 = vfmacc_vf_f16m1(_sum7, r0[19], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
596 _sum0 = vfmacc_vf_f16m1(_sum0, r0[5], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
597 _sum1 = vfmacc_vf_f16m1(_sum1, r0[7], _k05, vl); in conv7x7s2_pack1ton_fp16sa_rvv()
[all …]
/dports/misc/ncnn/ncnn-20211208/src/layer/mips/
H A Dconvolution_7x7_pack1to4.h69 v4f32 _k05 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0); in conv7x7s2_pack1to4_msa() local
112 _sum0 = __msa_fmadd_w(_sum0, _r05, _k05); in conv7x7s2_pack1to4_msa()
113 _sum1 = __msa_fmadd_w(_sum1, _r07, _k05); in conv7x7s2_pack1to4_msa()
114 _sum2 = __msa_fmadd_w(_sum2, _r09, _k05); in conv7x7s2_pack1to4_msa()
115 _sum3 = __msa_fmadd_w(_sum3, _r0b, _k05); in conv7x7s2_pack1to4_msa()
487 v4f32 _k05 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0); in conv7x7s2_pack1to4_msa() local
500 _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0n, 1), _k05); in conv7x7s2_pack1to4_msa()