| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
164 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
261 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
272 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
360 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
368 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
164 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
262 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
273 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
362 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
370 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
164 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
260 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
271 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
358 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
366 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
164 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
260 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
271 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
358 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
366 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
164 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
262 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
273 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
362 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
370 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4-acc2.c | 153 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
164 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
261 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
272 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
360 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
368 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
|
| D | 5x5p2-minmax-neon-2x4-acc2.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
200 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
202 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
333 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
346 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
348 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
468 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
477 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
479 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-2x4.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
200 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
202 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
331 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
344 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
346 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
464 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
473 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
475 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
|
| D | 5x5p2-minmax-neon-2x4.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
200 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
202 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
331 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
344 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
346 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
464 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
473 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
475 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
|
| D | 5x5p2-minmax-neonfma-2x4-acc2.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
200 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
202 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
333 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
346 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
348 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
468 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
477 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
479 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-3x4.c | 221 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
234 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
237 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
240 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
402 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
415 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
418 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
421 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
570 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
578 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 250 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
263 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
266 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
269 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
430 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
443 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
446 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
449 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
596 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
604 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 250 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
263 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
266 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
269 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
433 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
446 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
449 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
452 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
602 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
610 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 250 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
263 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
266 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
269 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
430 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
443 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
446 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
449 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
596 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
604 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-neon-3x4-acc2.c | 221 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
234 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
237 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
240 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
405 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
418 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
421 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
424 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
576 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
584 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-3x4.c | 221 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
234 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
237 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
240 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
402 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
415 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
418 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
421 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
570 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
578 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4-acc2.c | 221 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
234 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
237 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
240 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
405 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
418 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
421 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
424 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
576 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
584 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-2x4-acc3.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
200 vo1p2 = vfmaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
202 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
335 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
348 vo1p2 = vfmaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
350 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
472 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
481 vo1p2 = vfmaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
483 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c | 216 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
229 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
359 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
372 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
374 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
490 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
499 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
501 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c | 216 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
229 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
363 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
376 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
378 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
498 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3() local
507 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
509 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 216 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
229 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
361 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
374 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
376 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
494 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2() local
503 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
505 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4_acc2()
|
| D | 5x5p2-minmax-neon-2x4-acc3.c | 187 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
200 vo1p2 = vmlaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
202 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
335 const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
348 vo1p2 = vmlaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
350 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
472 const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
481 vo1p2 = vmlaq_lane_f32(vo1p2, vi2x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
483 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 216 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
229 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
359 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
372 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
374 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
490 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
499 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
501 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 216 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
229 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
231 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
361 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
374 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
376 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
494 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2() local
503 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
505 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 182 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
193 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
288 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
299 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
384 const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
392 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
|