/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 191 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local 322 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
D | 3x3p1-minmax-neon-6x4.c | 188 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local 321 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
|
D | 3x3p1-minmax-neonfma-6x4.c | 188 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local 321 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 191 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local 322 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
D | 5x5p2-minmax-neon-4x4.c | 224 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 442 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 647 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 227 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 448 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 655 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4.c | 227 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 444 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 647 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4.c | 227 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 444 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 647 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c | 227 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 448 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 655 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4.c | 224 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 442 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 647 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 207 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local 331 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 207 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local 331 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
|
D | 3x3p1-minmax-ssse3-6x4.c | 200 …const __m128 vi7x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local 326 … const __m128 vi7x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi7x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 255 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 509 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 747 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c | 253 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 470 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 673 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c | 253 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 470 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 673 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
|
D | 5x5p2-minmax-neon-4x4-acc2.c | 224 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 446 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 655 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4-acc2.c | 224 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 446 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 655 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 255 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 509 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 747 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 253 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 474 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 681 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-5x4.c | 252 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 507 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 747 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
|
D | 5x5p2-minmax-neon-5x4.c | 252 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 507 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 747 const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 253 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local 474 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local 681 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 281 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 535 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 773 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 281 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 535 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 773 const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
|