/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c | 154 const v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local 279 …const v128_t vi7x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c | 154 const v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local 279 …const v128_t vi7x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 164 const v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local 289 …const v128_t vi7x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c | 164 const v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local 289 …const v128_t vi7x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-sse-4x4.c | 153 const __m128 vi7x8ACE = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local 283 …const __m128 vi7x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
|
D | 3x3s2p1-minmax-sse-5x4.c | 171 const __m128 vi7x8ACE = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local 328 …const __m128 vi7x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
|
D | 3x3s2p1-minmax-sse-6x4.c | 189 const __m128 vi7x8ACE = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local 373 …const __m128 vi7x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|
D | 3x3s2p1-minmax-neonfma-4x4.c | 231 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-neon-4x4.c | 231 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
|
D | 5x5s2p2-minmax-neon-3x4.c | 324 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4.c | 324 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 327 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
|
D | 5x5s2p2-minmax-neon-3x4-acc2.c | 327 …const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 158 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 158 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 158 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 158 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 178 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 178 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 178 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 178 v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-sse-3x4.c | 174 __m128 vi7x8ACE = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 174 __m128 vi7x8ACE = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
|