| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
271 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
287 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
303 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
335 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
270 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
286 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
302 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
334 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
269 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
301 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
333 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
269 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
301 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
333 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
146 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
272 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
288 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
304 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
336 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
126 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
252 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
268 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
284 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
316 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
126 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
253 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
269 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
285 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
317 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
146 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
272 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
288 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
304 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
336 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
250 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
266 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
282 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
314 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
126 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
253 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
269 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
285 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
317 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
250 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
266 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
282 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
314 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
251 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
267 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
283 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
315 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
249 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
265 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
281 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
313 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
146 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
273 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
289 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
305 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
337 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
126 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
252 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
268 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
284 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
316 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
270 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
286 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
302 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
334 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
251 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
267 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
283 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
315 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
271 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
287 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
303 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
335 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c | 111 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
146 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
147 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
249 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
265 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
281 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
313 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c | 131 v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local
146 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
166 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
167 vi4x0246 = vi4x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
273 vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
289 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
305 const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
337 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
|
| D | 5x5s2p2-minmax-sse-1x4-acc2.c | 127 __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
136 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
142 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
266 vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
273 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
279 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
322 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
|
| D | 5x5s2p2-minmax-sse-1x4-acc5.c | 127 __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
136 __m128 vo0p4 = _mm_mul_ps(vi4x8ACE, vk42); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
142 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
269 vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
276 __m128 vo0p4 = _mm_mul_ps(vi4x8ACE, vk42); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
282 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
325 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
|
| D | 5x5s2p2-minmax-sse-1x4.c | 127 __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
136 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
142 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
265 vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
272 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
278 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
321 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
|
| D | 5x5s2p2-minmax-sse-1x4-acc4.c | 127 __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
136 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
142 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
268 vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
275 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
281 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
324 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
|
| D | 5x5s2p2-minmax-sse-1x4-acc3.c | 127 __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
136 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
142 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
267 vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
274 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
280 const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
323 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
|