| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
127 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
224 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
235 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
327 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
335 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
127 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
225 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
236 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
329 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
337 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
223 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
234 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
325 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
333 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
127 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
223 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
234 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
325 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
333 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
127 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
225 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
236 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
329 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
337 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4-acc2.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
224 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
235 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
327 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
335 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
|
| D | 5x5p2-minmax-neon-2x4-acc2.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
150 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
283 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
296 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
298 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
423 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
432 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
434 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-2x4.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
150 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
152 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
281 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
294 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
296 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
419 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
428 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
430 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
|
| D | 5x5p2-minmax-neon-2x4.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
150 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
281 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
294 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
296 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
419 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
428 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
430 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
|
| D | 5x5p2-minmax-neonfma-2x4-acc2.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
150 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
152 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
283 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
296 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
298 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
423 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
432 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
434 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-3x4.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
171 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
174 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
339 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
352 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
355 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
358 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
513 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
521 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 187 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
200 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
203 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
206 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
367 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
380 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
383 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
386 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
540 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
548 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 187 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
200 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
203 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
206 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
370 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
383 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
386 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
389 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
546 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
554 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 187 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
200 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
203 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
206 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
367 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
380 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
383 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
386 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
540 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
548 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-neon-3x4-acc2.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
171 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
174 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
342 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
355 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
358 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
361 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
519 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
527 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-3x4.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
171 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
174 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
339 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
352 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
355 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
358 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
513 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
521 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4-acc2.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
171 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
174 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
342 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
355 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
358 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
361 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
519 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
527 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-2x4-acc3.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
150 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
152 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
285 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
298 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
300 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
427 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
436 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
438 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c | 166 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
179 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
181 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
309 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
322 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
324 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
446 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
455 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
457 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c | 166 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
179 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
181 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
313 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
326 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
328 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
454 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
463 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
465 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 166 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
179 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
181 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
311 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
324 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
326 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
450 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
459 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
461 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
|
| D | 5x5p2-minmax-neon-2x4-acc3.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
150 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
152 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
285 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
298 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
300 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
427 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
436 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
438 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 166 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
179 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
181 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
309 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
322 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
324 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
446 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
455 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
457 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 166 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
179 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
181 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
311 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
324 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
326 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
450 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
459 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
461 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 145 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
156 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
251 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
262 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
352 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
360 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
|