| /dports/graphics/vapoursynth-waifu2x-ncnn-vulkan/vapoursynth-waifu2x-ncnn-vulkan-r4/deps/ncnn/src/layer/arm/ |
| H A D | convolution_3x3_pack8to4_fp16s.h | 256 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
258 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
1119 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
1122 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
|
| H A D | convolution_3x3_pack8_fp16s.h | 238 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
240 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1117 … float16x8_t _tmp4m = vfmaq_n_f16(vfmaq_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
1120 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1389 … float16x8_t _tmp4m = vfmaq_n_f16(vsubq_f16(_r04, _r02), vsubq_f16(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack8_fp16sa_neon() local
1396 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack8_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_fp16s.h | 356 float16x4_t _tmp4m = vsub_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
358 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
1063 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
1066 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_bf16s.h | 134 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_bf16s_neon() local
136 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1727 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_bf16s_neon() local
1730 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1898 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_bf16s_neon() local
1905 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_bf16s_neon()
|
| H A D | convolution_3x3_pack4.h | 365 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_neon() local
367 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
1958 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_neon() local
1961 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
2358 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_neon() local
2365 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_neon()
|
| /dports/graphics/waifu2x-ncnn-vulkan/waifu2x-ncnn-vulkan-20210521/src/ncnn/src/layer/arm/ |
| H A D | convolution_3x3_pack8to4_fp16s.h | 256 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
258 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
1119 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
1122 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
|
| H A D | convolution_3x3_pack8_fp16s.h | 238 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
240 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1117 … float16x8_t _tmp4m = vfmaq_n_f16(vfmaq_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
1120 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1389 … float16x8_t _tmp4m = vfmaq_n_f16(vsubq_f16(_r04, _r02), vsubq_f16(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack8_fp16sa_neon() local
1396 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack8_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_fp16s.h | 356 float16x4_t _tmp4m = vsub_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
358 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
1063 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
1066 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_bf16s.h | 134 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_bf16s_neon() local
136 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1727 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_bf16s_neon() local
1730 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1898 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_bf16s_neon() local
1905 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_bf16s_neon()
|
| H A D | convolution_3x3_pack4.h | 365 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_neon() local
367 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
1958 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_neon() local
1961 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
2358 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_neon() local
2365 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_neon()
|
| /dports/benchmarks/vkpeak/vkpeak-20210430/ncnn/src/layer/arm/ |
| H A D | convolution_3x3_pack8to4_fp16s.h | 256 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
258 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
1119 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
1122 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
|
| H A D | convolution_3x3_pack8_fp16s.h | 238 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
240 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1117 … float16x8_t _tmp4m = vfmaq_n_f16(vfmaq_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
1120 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1389 … float16x8_t _tmp4m = vfmaq_n_f16(vsubq_f16(_r04, _r02), vsubq_f16(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack8_fp16sa_neon() local
1396 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack8_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_fp16s.h | 356 float16x4_t _tmp4m = vsub_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
358 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
1063 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
1066 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_bf16s.h | 134 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_bf16s_neon() local
136 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1727 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_bf16s_neon() local
1730 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1898 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_bf16s_neon() local
1905 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_bf16s_neon()
|
| H A D | convolution_3x3_pack4.h | 365 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_neon() local
367 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
1958 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_neon() local
1961 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_neon()
2358 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_neon() local
2365 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_neon()
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/arm/ |
| H A D | convolution_3x3_pack8to4_fp16s.h | 256 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
258 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
1119 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
1122 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
|
| H A D | convolution_3x3_pack8_fp16s.h | 238 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
240 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1117 … float16x8_t _tmp4m = vfmaq_n_f16(vfmaq_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
1120 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1389 … float16x8_t _tmp4m = vfmaq_n_f16(vsubq_f16(_r04, _r02), vsubq_f16(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack8_fp16sa_neon() local
1396 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack8_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_fp16s.h | 356 float16x4_t _tmp4m = vsub_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
358 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
1063 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
1066 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_bf16s.h | 134 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_bf16s_neon() local
136 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1727 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_bf16s_neon() local
1730 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1898 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_bf16s_neon() local
1905 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_bf16s_neon()
|
| /dports/graphics/realsr-ncnn-vulkan/realsr-ncnn-vulkan-20210210/src/ncnn/src/layer/arm/ |
| H A D | convolution_3x3_pack8to4_fp16s.h | 256 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
258 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
1119 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8to4_fp16sa_neon() local
1122 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8to4_fp16sa_neon()
|
| H A D | convolution_3x3_pack8_fp16s.h | 238 float16x8_t _tmp4m = vsubq_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
240 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1117 … float16x8_t _tmp4m = vfmaq_n_f16(vfmaq_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack8_fp16sa_neon() local
1120 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack8_fp16sa_neon()
1389 … float16x8_t _tmp4m = vfmaq_n_f16(vsubq_f16(_r04, _r02), vsubq_f16(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack8_fp16sa_neon() local
1396 vst1q_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack8_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_fp16s.h | 356 float16x4_t _tmp4m = vsub_f16(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
358 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
1063 … float16x4_t _tmp4m = vfma_n_f16(vfma_n_f16(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_fp16sa_neon() local
1066 vst1_f16(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_fp16sa_neon()
|
| H A D | convolution_3x3_pack4_bf16s.h | 134 float32x4_t _tmp4m = vsubq_f32(_tmp34a, _tmp34b); in conv3x3s1_winograd64_pack4_bf16s_neon() local
136 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1727 … float32x4_t _tmp4m = vmlaq_n_f32(vmlaq_n_f32(_tmp024a, _tmp024b, 16.f), _tmp024c, 2.f); in conv3x3s1_winograd64_pack4_bf16s_neon() local
1730 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd64_pack4_bf16s_neon()
1898 … float32x4_t _tmp4m = vmlaq_n_f32(vsubq_f32(_r04, _r02), vsubq_f32(_r01, _r03), 2.f); in conv3x3s1_winograd42_pack4_bf16s_neon() local
1905 vst1q_f32(tmp[4][m], _tmp4m); in conv3x3s1_winograd42_pack4_bf16s_neon()
|
| /dports/misc/ncnn/ncnn-20211208/src/layer/riscv/ |
| H A D | convolution_3x3_packn.h | 209 vfloat32m1_t _tmp4m = vfsub_vv_f32m1(_tmp34a, _tmp34b, vl); in conv3x3s1_winograd64_packn_rvv() local
211 vse32_v_f32m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd64_packn_rvv()
678 …vfloat32m1_t _tmp4m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c… in conv3x3s1_winograd64_packn_rvv() local
681 vse32_v_f32m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd64_packn_rvv()
908 …vfloat32m1_t _tmp4m = vfmacc_vf_f32m1(vfsub_vv_f32m1(_r04, _r02, vl), 2.f, vfsub_vv_f32m1(_r01, _r… in conv3x3s1_winograd42_packn_rvv() local
915 vse32_v_f32m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd42_packn_rvv()
|
| H A D | convolution_3x3_packn_fp16s.h | 209 vfloat16m1_t _tmp4m = vfsub_vv_f16m1(_tmp34a, _tmp34b, vl); in conv3x3s1_winograd64_packn_fp16sa_rvv() local
211 vse16_v_f16m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd64_packn_fp16sa_rvv()
678 …vfloat16m1_t _tmp4m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c… in conv3x3s1_winograd64_packn_fp16sa_rvv() local
681 vse16_v_f16m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd64_packn_fp16sa_rvv()
908 …vfloat16m1_t _tmp4m = vfmacc_vf_f16m1(vfsub_vv_f16m1(_r04, _r02, vl), 2.f, vfsub_vv_f16m1(_r01, _r… in conv3x3s1_winograd42_packn_fp16sa_rvv() local
915 vse16_v_f16m1(tmp[4][m], _tmp4m, vl); in conv3x3s1_winograd42_packn_fp16sa_rvv()
|