/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neonfma-1x4-acc2.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 257 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 358 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
|
D | 5x5p2-minmax-neon-1x4.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 256 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 356 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 258 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 356 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 259 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 358 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neonfma-1x4.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 256 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 356 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
|
D | 5x5p2-minmax-neon-1x4-acc2.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 257 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 358 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 258 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 356 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 259 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 358 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neonfma-1x4-acc3.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 258 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 360 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 261 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 362 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc4.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 259 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 362 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 261 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 362 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 262 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 364 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc3.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 258 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 360 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 260 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 360 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local
|
D | 5x5p2-minmax-neon-1x4-acc4.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 259 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 362 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 260 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 360 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 152 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 262 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 364 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neonfma-1x4-acc5.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 260 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 364 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc5.c | 149 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 260 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 364 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 178 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 284 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 382 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4.c | 186 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 329 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 462 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local
|
D | 5x5p2-minmax-neonfma-2x4.c | 183 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 327 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 462 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
|
D | 5x5p2-minmax-neon-2x4.c | 183 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 327 const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 462 const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-2x4.c | 186 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 329 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 462 const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local
|