/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-neon-3x4.c | 185 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local 385 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4.c | 185 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local 385 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 185 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local 388 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
|
D | 5x5s2p2-minmax-neon-3x4-acc2.c | 185 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local 388 const float32x4_t vi8x68AC = vextq_f32(vi8x0246, vi8x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 225 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local 470 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 225 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local 470 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 225 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local 473 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 225 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local 473 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 245 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local 490 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 245 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local 490 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 245 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local 493 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 245 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local 493 const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-sse-3x4.c | 239 const __m128 vi8x68AC = _mm_move_ss(vi8xE8AC, vi8x6024); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local 493 const __m128 vi8x68AC = _mm_move_ss(vi8xE8AC, vi8x6024); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 239 const __m128 vi8x68AC = _mm_move_ss(vi8xE8AC, vi8x6024); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local 496 const __m128 vi8x68AC = _mm_move_ss(vi8xE8AC, vi8x6024); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
|