/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neonfma-2x4.c | 104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local 172 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
|
D | 3x3s2p1-minmax-neon-2x4.c | 104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local 172 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
|
D | 3x3s2p1-minmax-neon-2x4-acc2.c | 104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() local 174 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() local
|
D | 3x3s2p1-minmax-neonfma-2x4-acc2.c | 104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local 174 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c | 126 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local 198 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 126 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local 200 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
|
D | 3x3s2p1-minmax-neonfma-3x4.c | 122 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local 213 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c | 126 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local 198 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
|
D | 3x3s2p1-minmax-neon-3x4.c | 122 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local 213 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 126 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local 200 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c | 136 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local 208 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c | 136 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local 208 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 136 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local 210 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 136 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local 210 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
|
D | 3x3s2p1-minmax-sse-2x4.c | 137 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local 213 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
|
D | 3x3s2p1-minmax-sse-2x4-acc2.c | 137 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local 215 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
|
D | 3x3s2p1-minmax-neonfma-4x4.c | 140 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local 254 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-neon-4x4.c | 140 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local 254 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c | 152 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local 249 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c | 152 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local 249 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c | 162 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local 259 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c | 162 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local 259 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
|
D | 3x3s2p1-minmax-sse-3x4.c | 167 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local 270 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c | 178 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local 300 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c | 178 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local 300 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
|