Home
last modified time | relevance | path

Searched refs:vi6x1357 (Results 1 – 25 of 42) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-neonfma-3x4.c88 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
128 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
129 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
216 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
D3x3s2p1-minmax-neon-3x4.c88 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
128 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
129 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
216 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c101 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
170 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
171 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
264 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c101 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
170 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
171 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
264 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c91 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
160 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
161 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
254 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c91 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
160 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
161 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
254 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
D3x3s2p1-minmax-neonfma-4x4.c98 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
146 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
147 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
257 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
D3x3s2p1-minmax-neon-4x4.c98 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
146 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
147 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
257 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c111 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
196 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
197 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
315 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
D5x5s2p2-minmax-neon-2x4-acc2.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
334 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c111 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
196 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
197 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
315 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c101 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
186 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
187 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
305 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c101 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
186 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4()
187 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4()
305 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4()
D5x5s2p2-minmax-neonfma-2x4-acc2.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
334 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
332 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
336 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
336 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c97 float32x4_t vi6x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
184 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
185 vi6x1357 = vi6x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
332 const float32x4_t vi6x79BD = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
419 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
421 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
421 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
419 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
423 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c120 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
236 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
237 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
423 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c100 v128_t vi6x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2() local
216 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2()
217 vi6x1357 = vi6x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2()
401 const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_2x4_acc2()

12