| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5s2p2-minmax-neon-1x4-acc4.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
129 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
233 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
241 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc3.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
129 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
232 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
240 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
129 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
233 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
241 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
129 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
231 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
239 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
129 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
232 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
240 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
| D | 5x5s2p2-minmax-neon-1x4.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
129 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
230 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
238 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
| D | 5x5s2p2-minmax-neonfma-1x4.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
129 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
230 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
238 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc2.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
129 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
231 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
239 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4-acc2.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
158 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
163 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
307 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
314 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
319 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc2.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
158 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
163 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
307 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
314 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
319 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
158 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
163 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
305 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
312 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
317 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
|
| D | 5x5s2p2-minmax-neon-2x4-acc3.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
158 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
163 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
309 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
316 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
321 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc3.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
158 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
163 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
309 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
316 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
321 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4.c | 146 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
158 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
163 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
305 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
312 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
317 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc5.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
129 vo0p4 = vmlaq_lane_f32(vo0p4, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
234 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
242 vo0p4 = vmlaq_lane_f32(vo0p4, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc5.c | 118 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
129 vo0p4 = vfmaq_lane_f32(vo0p4, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
234 const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
242 vo0p4 = vfmaq_lane_f32(vo0p4, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
215 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
392 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
399 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
404 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
215 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
394 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
401 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
406 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
215 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
394 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
401 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
406 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
215 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
392 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
399 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
404 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
215 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
396 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
403 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
408 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c | 198 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
215 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
396 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
403 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
408 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 162 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
173 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
301 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
309 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 162 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
173 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
300 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
308 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 162 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
173 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
299 const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
307 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|