| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local 131 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local 130 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local 131 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local 130 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local 132 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local 133 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local 133 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c | 80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3() local 132 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local 140 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4() local 140 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2() local 141 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2() local
|
| D | 3x3s2p1-minmax-sse-1x4.c | 84 const __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local 133 …const __m128 vi0x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local 141 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3() local 142 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3() local 142 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3() local
|
| D | 3x3s2p1-minmax-sse-1x4-acc3.c | 84 const __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local 135 …const __m128 vi0x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4() local 143 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4() local
|
| D | 3x3s2p1-minmax-sse-1x4-acc2.c | 84 const __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local 134 …const __m128 vi0x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 90 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4() local 143 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4() local
|
| D | 3x3s2p1-minmax-sse-1x4-acc4.c | 84 const __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local 136 …const __m128 vi0x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c | 100 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local 175 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c | 100 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local 175 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 100 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local 177 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 100 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local 177 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c | 110 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local 185 …const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
|