Home
last modified time | relevance | path

Searched refs:vi4x1357 (Results 1 – 25 of 84) sorted by relevance

1234

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-neonfma-2x4.c76 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
107 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
173 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3s2p1-minmax-neon-2x4-acc2.c76 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() local
106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
107 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
175 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c76 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
107 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
173 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4-acc2.c76 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
107 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
175 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c79 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
131 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
203 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c89 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
141 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
211 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c79 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
131 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
201 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c89 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
141 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
213 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c89 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
141 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
211 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c89 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
141 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
213 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-neon-1x4.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
248 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4-acc3.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
250 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4-acc2.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
249 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c79 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
131 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
203 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
D5x5s2p2-minmax-neon-1x4-acc4.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
251 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
250 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
248 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c79 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
131 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
201 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
D5x5s2p2-minmax-neonfma-1x4-acc4.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
251 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
249 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neon-1x4-acc5.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
252 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c82 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
144 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
252 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D3x3s2p1-minmax-neon-3x4.c86 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
124 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
125 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
214 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
D3x3s2p1-minmax-neonfma-3x4.c86 float32x4_t vi4x1357 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
124 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
125 vi4x1357 = vi4x8ACE9BDF.val[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
214 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c105 v128_t vi4x1357 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
318 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()

1234