/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-neon-3x4.c | 273 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE9BDF.val[0], vi8xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local 446 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4.c | 273 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE9BDF.val[0], vi8xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local 446 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 273 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE9BDF.val[0], vi8xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local 449 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
|
D | 5x5s2p2-minmax-neon-3x4-acc2.c | 273 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE9BDF.val[0], vi8xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local 449 const float32x4_t vi8xACEG = vextq_f32(vi8x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 358 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local 530 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 358 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local 530 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 358 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local 533 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 358 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local 533 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 378 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local 550 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 378 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local 550 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 378 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local 553 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 378 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local 553 const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-sse-3x4.c | 386 const __m128 vi8xACEG = _mm_shuffle_ps(vi8xGACE, vi8xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local 566 const __m128 vi8xACEG = _mm_shuffle_ps(vi8xGACE, vi8xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 386 const __m128 vi8xACEG = _mm_shuffle_ps(vi8xGACE, vi8xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local 569 const __m128 vi8xACEG = _mm_shuffle_ps(vi8xGACE, vi8xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
|