/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c | 188 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local 210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() 211 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c | 208 const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local 230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() 231 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4-acc2.c | 203 const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local 224 const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() 225 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
|
D | 5x5s2p2-minmax-sse-1x4-acc5.c | 203 const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local 224 const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() 225 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4.c | 203 const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local 224 const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() 225 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
|
D | 5x5s2p2-minmax-sse-1x4-acc4.c | 203 const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local 224 const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() 225 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
|
D | 5x5s2p2-minmax-sse-1x4-acc3.c | 203 const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local 224 const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() 225 vi1x8ACE = vi1xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
|