Home
last modified time | relevance | path

Searched refs:vi2xKLMN (Results 1 – 25 of 50) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c177 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local
190 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
191 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c197 const v128_t vi2xKLMN = wasm_v128_load(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local
210 const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
211 const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
D5x5s2p2-minmax-sse-1x4-acc2.c180 const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
206 const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
207 const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
D5x5s2p2-minmax-sse-1x4-acc5.c180 const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
206 const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
207 const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
D5x5s2p2-minmax-sse-1x4.c180 const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
206 const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
207 const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
D5x5s2p2-minmax-sse-1x4-acc4.c180 const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
206 const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
207 const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
D5x5s2p2-minmax-sse-1x4-acc3.c180 const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
206 const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
207 const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()

12