| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
267 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
279 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
299 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
331 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
266 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
278 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
298 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
330 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
138 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
265 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
297 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
329 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
138 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
265 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
297 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
329 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
268 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
280 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
300 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
332 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
248 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
260 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
280 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
312 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
249 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
261 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
281 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
313 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
268 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
280 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
300 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
332 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
246 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
258 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
278 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
310 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
249 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
261 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
281 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
313 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
246 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
258 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
278 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
310 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
247 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
259 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
279 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
311 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
118 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
245 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
257 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
277 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
309 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
269 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
281 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
301 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
333 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
248 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
260 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
280 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
312 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
266 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
278 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
298 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
330 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local
118 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
247 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
259 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
279 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
311 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
267 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
279 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
299 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
331 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 103 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local
118 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
138 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
139 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
245 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
257 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
277 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
309 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, 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 | 123 v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
158 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
159 vi0x0246 = vi0x8ACE; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
269 vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
281 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
301 const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
333 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
|
| D | 5x5s2p2-minmax-sse-1x4-acc2.c | 119 __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
132 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
138 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
258 vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
269 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
275 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
318 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
|
| D | 5x5s2p2-minmax-sse-1x4-acc5.c | 119 __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
132 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
138 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
261 vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
272 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
278 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
321 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
|
| D | 5x5s2p2-minmax-sse-1x4.c | 119 __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
132 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
138 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
257 vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
268 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
274 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
317 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
|
| D | 5x5s2p2-minmax-sse-1x4-acc4.c | 119 __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
132 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
138 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
260 vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
271 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
277 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
320 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
|
| D | 5x5s2p2-minmax-sse-1x4-acc3.c | 119 __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
132 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
138 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
259 vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
270 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
276 const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
319 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
|