/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-neon-1x4-acc4.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 235 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
D | 5x5s2p2-minmax-neon-1x4-acc3.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 234 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 235 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 233 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 234 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
D | 5x5s2p2-minmax-neon-1x4.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 232 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
D | 5x5s2p2-minmax-neonfma-1x4.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 232 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
D | 5x5s2p2-minmax-neon-1x4-acc2.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 233 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
D | 5x5s2p2-minmax-neon-1x4-acc5.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() 236 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc5.c | 76 float32x4_t vi4x0246 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local 122 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() 123 vi4x0246 = vi4x8ACE9BDF.val[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() 236 const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 99 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 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() 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()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 79 v128_t vi4x0246 = vzero; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 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() 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()
|