/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-wasmsimd-x86-splat-2x4.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 260 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local 399 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local
|
D | 5x5p2-minmax-neon-2x4.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 260 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local 400 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 260 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local 399 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local
|
D | 5x5p2-minmax-neonfma-2x4.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 260 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local 400 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local 262 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local 403 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local
|
D | 5x5p2-minmax-neonfma-2x4-acc2.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local 262 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local 404 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local 262 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local 403 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local
|
D | 5x5p2-minmax-neon-2x4-acc2.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local 262 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local 404 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local 264 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local 407 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c | 117 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local 264 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local 407 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-4x4.c | 122 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local 226 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-4x4.c | 122 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local 226 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3p1-minmax-ssse3-4x4.c | 133 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local 231 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local
|
D | 3x3p1-minmax-neonfma-4x4.c | 121 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local 226 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local
|
D | 3x3p1-minmax-neon-4x4.c | 121 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local 226 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 143 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local 286 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local 425 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4() local
|
D | 5x5p2-minmax-neon-2x4-acc3.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local 264 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local 408 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
|
D | 5x5p2-minmax-neon-3x4.c | 131 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 312 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 488 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
|
D | 5x5p2-minmax-neonfma-3x4.c | 131 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 312 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local 488 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-3x4.c | 132 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 312 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local 487 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-3x4.c | 132 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 312 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local 487 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c | 143 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local 286 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local 425 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_2x4() local
|
D | 5x5p2-minmax-neonfma-2x4-acc3.c | 116 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local 264 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local 408 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 135 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local 257 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 138 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4() local 235 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4() local
|