/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c | 179 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local 192 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() 193 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c | 199 const v128_t vi3xGHIJ = wasm_v128_load(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local 212 const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() 213 const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4-acc2.c | 182 const __m128 vi3xGHIJ = _mm_loadu_ps(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local 209 const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() 210 const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
|
D | 5x5s2p2-minmax-sse-1x4-acc5.c | 182 const __m128 vi3xGHIJ = _mm_loadu_ps(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local 209 const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() 210 const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4.c | 182 const __m128 vi3xGHIJ = _mm_loadu_ps(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local 209 const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() 210 const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
|
D | 5x5s2p2-minmax-sse-1x4-acc4.c | 182 const __m128 vi3xGHIJ = _mm_loadu_ps(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local 209 const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() 210 const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
|
D | 5x5s2p2-minmax-sse-1x4-acc3.c | 182 const __m128 vi3xGHIJ = _mm_loadu_ps(i3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local 209 const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() 210 const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
|