/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neonfma-3x4.c | 195 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 376 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 546 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-3x4.c | 198 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 378 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 546 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5p2-minmax-neon-3x4.c | 195 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 376 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 546 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-3x4.c | 198 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 378 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 546 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 172 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local 286 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-5x4.c | 172 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local 286 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 198 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local 381 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local 552 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local
|
D | 5x5p2-minmax-neon-3x4-acc2.c | 195 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local 379 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local 552 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 198 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local 381 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local 552 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local
|
D | 5x5p2-minmax-neonfma-3x4-acc2.c | 195 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local 379 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local 552 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 187 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local 294 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
|
D | 3x3p1-minmax-ssse3-5x4.c | 180 …const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local 289 … const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi6x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 187 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local 294 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
|
D | 3x3p1-minmax-neon-5x4.c | 169 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local 285 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
|
D | 3x3p1-minmax-neonfma-5x4.c | 169 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local 285 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 190 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local 321 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
D | 3x3p1-minmax-neon-6x4.c | 187 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local 320 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
|
D | 3x3p1-minmax-neonfma-6x4.c | 187 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local 320 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 190 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local 321 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
D | 5x5p2-minmax-neon-4x4.c | 223 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 441 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 646 const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 224 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local 404 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local 572 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 226 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 447 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 654 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 224 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local 407 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local 578 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4.c | 226 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 443 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 646 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 224 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local 404 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local 572 const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
|