/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 309 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local 329 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() 330 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 309 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local 329 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() 330 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 309 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local 329 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() 330 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 309 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local 329 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() 330 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 289 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local 309 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() 310 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
|
D | 5x5s2p2-minmax-sse-3x4.c | 290 const __m128 vi7xKLMN = _mm_loadu_ps(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local 336 const __m128 vi7xGIKM = _mm_shuffle_ps(vi7xGHIJ, vi7xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() 337 const __m128 vi7xHJLN = _mm_shuffle_ps(vi7xGHIJ, vi7xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 289 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local 309 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() 310 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 289 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local 309 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() 310 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 289 const v128_t vi7xKLMN = wasm_v128_load(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local 309 const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() 310 const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
|
D | 5x5s2p2-minmax-sse-3x4-acc2.c | 290 const __m128 vi7xKLMN = _mm_loadu_ps(i7 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local 336 const __m128 vi7xGIKM = _mm_shuffle_ps(vi7xGHIJ, vi7xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() 337 const __m128 vi7xHJLN = _mm_shuffle_ps(vi7xGHIJ, vi7xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2()
|