/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neon-3x4.c | 128 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local 216 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
|
D | 3x3s2p1-minmax-neonfma-3x4.c | 128 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local 216 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
|
D | 3x3s2p1-minmax-neon-4x4.c | 146 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local 257 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
|
D | 3x3s2p1-minmax-neonfma-4x4.c | 146 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local 257 const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c | 158 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local 252 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c | 158 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local 252 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c | 168 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local 262 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c | 168 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local 262 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
|
D | 3x3s2p1-minmax-sse-3x4.c | 170 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local 273 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c | 184 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local 303 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c | 184 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local 303 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c | 194 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local 313 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 194 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local 313 const v128_t vi6x7BDF = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-sse-4x4.c | 200 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local 330 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
|
D | 3x3s2p1-minmax-sse-5x4.c | 230 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local 387 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
|
D | 3x3s2p1-minmax-sse-6x4.c | 260 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local 444 const __m128 vi6x7BDF = _mm_move_ss(vi6xF9BD, vi6x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|
/external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neonfp16arith-3x4.c | 128 const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_3x4() local 216 const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x9BDF, 3); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_3x4() local
|
D | 3x3s2p1-minmax-neonfp16arith-4x4.c | 146 const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x8ACE9BDF.val[1], 3); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_4x4() local 257 const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x9BDF, 3); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_4x4() local
|