/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neonfma-3x4.c | 229 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 410 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 574 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-3x4.c | 232 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 412 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 574 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5p2-minmax-neon-3x4.c | 229 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 410 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 574 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-3x4.c | 232 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 412 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 574 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 232 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local 415 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local 580 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 229 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local 413 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local 580 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 232 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local 415 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local 580 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 229 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local 413 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local 580 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
|
D | 5x5p2-minmax-neon-4x4.c | 263 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 481 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 680 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 258 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local 438 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local 600 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 266 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 487 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 688 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 258 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local 441 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local 606 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 266 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 483 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 680 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 258 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local 438 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local 600 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 258 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local 441 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local 606 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4.c | 266 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 483 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 680 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 266 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 487 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 688 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4.c | 263 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 481 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 680 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 300 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 554 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 786 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 292 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 509 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 706 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 292 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 509 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 706 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
|
D | 5x5p2-minmax-neon-4x4-acc2.c | 263 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 485 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 688 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4-acc2.c | 263 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 485 const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 688 const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 300 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 554 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 786 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, 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 | 292 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 513 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 714 const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
|