/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 229 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 328 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
|
D | 5x5p2-minmax-neon-1x4.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 227 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 327 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc2.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 228 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 329 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 230 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 330 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 230 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 330 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neon-1x4-acc2.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 228 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 329 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 229 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 328 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
|
D | 5x5p2-minmax-neonfma-1x4.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 227 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 327 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 232 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 334 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 233 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 336 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neonfma-1x4-acc3.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 229 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 331 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 231 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 332 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 232 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 334 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 231 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 332 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local
|
D | 5x5p2-minmax-neonfma-1x4-acc4.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 230 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 333 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
|
D | 5x5p2-minmax-neon-1x4-acc4.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 230 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 333 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 123 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 233 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 336 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc3.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 229 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 331 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
|
D | 5x5p2-minmax-neonfma-1x4-acc5.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 231 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 335 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
|
D | 5x5p2-minmax-neon-1x4-acc5.c | 120 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 231 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 335 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 149 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 255 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local 354 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 149 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local 255 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local 354 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4() local
|
D | 5x5p2-minmax-neon-2x4.c | 141 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 285 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 421 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
|
D | 5x5p2-minmax-neonfma-2x4.c | 141 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 285 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 421 const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4.c | 144 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 287 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 422 const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local
|