| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
187 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
193 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
246 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
266 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
267 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
185 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
191 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
244 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
264 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
265 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
186 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
192 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
245 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
265 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
266 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neon-1x4-acc2.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
186 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
192 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
245 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
265 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
266 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
187 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
193 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
246 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
266 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
267 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
185 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
191 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
244 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
264 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
265 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
190 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
196 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
249 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
269 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
270 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
187 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
193 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
246 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
266 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
267 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
189 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
195 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
248 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
268 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
269 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
215 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
221 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
274 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
294 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
295 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
214 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2() local
220 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
273 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
293 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
294 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
216 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4() local
222 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
275 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
295 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
296 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
213 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
219 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
272 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
292 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
293 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4()
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
188 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
194 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
247 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
267 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
268 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
214 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2() local
220 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
273 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
293 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
294 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc5.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
189 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
195 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
248 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
268 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
269 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
215 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3() local
221 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
274 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
294 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
295 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5p2-minmax-neonfma-1x4-acc4.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
188 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
194 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
247 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
267 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
268 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4()
|
| D | 5x5p2-minmax-neon-1x4-acc5.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
189 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
195 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
248 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
268 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
269 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
213 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local
219 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
272 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
292 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
293 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4()
|
| D | 5x5p2-minmax-neon-1x4-acc4.c | 84 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
158 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
188 float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
194 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
247 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
267 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
268 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
188 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
194 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
247 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
267 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
268 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
216 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4() local
222 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
275 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
295 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
296 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 113 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5() local
166 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
186 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
187 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
217 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5() local
223 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
276 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
296 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
297 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 87 const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
140 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
160 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
161 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
187 v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
193 vi4x89AB = wasm_v128_and(vmask, vi4x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
246 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
266 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
267 vi4x4567 = vi4x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
|