/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neonfma-1x4-acc2.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 220 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 325 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
|
D | 5x5p2-minmax-neon-1x4.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 219 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 323 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 221 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 324 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 222 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 326 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neonfma-1x4.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 219 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 323 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
|
D | 5x5p2-minmax-neon-1x4-acc2.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 220 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 325 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 221 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 324 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 222 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 326 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neonfma-1x4-acc3.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 221 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 327 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 224 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 330 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc4.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 222 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 329 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 224 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 330 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 225 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 332 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc3.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 221 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 327 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 223 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 328 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local
|
D | 5x5p2-minmax-neon-1x4-acc4.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 222 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 329 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 223 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 328 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 115 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 225 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 332 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neonfma-1x4-acc5.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 223 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 331 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc5.c | 112 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 223 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 331 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 141 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 247 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 350 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4.c | 136 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 279 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 418 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local
|
D | 5x5p2-minmax-neonfma-2x4.c | 133 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 277 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 417 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
|
D | 5x5p2-minmax-neon-2x4.c | 133 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 277 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 417 const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-2x4.c | 136 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 279 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 418 const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local
|