| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
186 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
192 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
245 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
264 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
265 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
184 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
190 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
243 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
262 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
263 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
185 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
191 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
244 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
263 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
264 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neon-1x4-acc2.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
185 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
191 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
244 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
263 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
264 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
186 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
192 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
245 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
264 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
265 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
184 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
190 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
243 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
262 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
263 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
189 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
195 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
248 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
267 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
268 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
186 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
192 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
245 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
264 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
265 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
188 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
194 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
247 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
266 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
267 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
214 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
220 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
273 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
292 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
293 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
213 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2() local
219 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
272 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
291 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
292 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
215 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4() local
221 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
274 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
293 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
294 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
212 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
218 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
271 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
290 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
291 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
187 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
193 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
246 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
265 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
266 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
213 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2() local
219 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
272 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
291 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
292 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc5.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
188 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
194 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
247 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
266 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
267 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
214 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3() local
220 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
273 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
292 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
293 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5p2-minmax-neonfma-1x4-acc4.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
187 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
193 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
246 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
265 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
266 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
|
| D | 5x5p2-minmax-neon-1x4-acc5.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
188 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
194 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
247 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
266 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
267 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
212 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local
218 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
271 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
290 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
291 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
|
| D | 5x5p2-minmax-neon-1x4-acc4.c | 83 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
155 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
156 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
187 float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
193 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
246 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
265 const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
266 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
187 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
193 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
246 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
265 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
266 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
215 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4() local
221 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
274 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
293 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
294 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 112 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5() local
165 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
184 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
185 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
216 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5() local
222 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
275 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
294 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
295 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 86 const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
139 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
158 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
159 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
186 v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
192 vi3x89AB = wasm_v128_and(vmask, vi3x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
245 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
264 const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
265 vi3x4567 = vi3x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
|