/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c | 153 const v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local 278 …const v128_t vi7x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c | 153 const v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local 278 …const v128_t vi7x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 163 const v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local 288 …const v128_t vi7x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c | 163 const v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local 288 …const v128_t vi7x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
|
D | 3x3s2p1-minmax-sse-4x4.c | 154 const __m128 vi7x9BDF = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local 284 …const __m128 vi7x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
|
D | 3x3s2p1-minmax-sse-5x4.c | 172 const __m128 vi7x9BDF = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local 329 …const __m128 vi7x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
|
D | 3x3s2p1-minmax-sse-6x4.c | 190 const __m128 vi7x9BDF = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local 374 …const __m128 vi7x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|
D | 3x3s2p1-minmax-neonfma-4x4.c | 232 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-neon-4x4.c | 232 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
|
D | 5x5s2p2-minmax-neonfma-3x4.c | 334 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
|
D | 5x5s2p2-minmax-neon-3x4.c | 334 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
|
D | 5x5s2p2-minmax-neon-3x4-acc2.c | 337 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
|
D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 337 …const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi7x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 157 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 157 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 177 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 157 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 157 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 177 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 177 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 177 v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 175 __m128 vi7x9BDF = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
|
D | 5x5s2p2-minmax-sse-3x4.c | 175 __m128 vi7x9BDF = _mm_shuffle_ps(vi7x89AB, vi7xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
|