/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neonfma-4x4.c | 150 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local 259 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-neon-4x4.c | 150 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local 259 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c | 188 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local 305 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c | 188 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local 305 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 198 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local 315 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c | 198 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local 315 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-sse-4x4.c | 202 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local 332 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
|
D | 3x3s2p1-minmax-sse-5x4.c | 232 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local 389 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
|
D | 3x3s2p1-minmax-sse-6x4.c | 262 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local 446 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|