/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c | 186 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local 207 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() 208 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c | 206 const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local 227 const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() 228 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4-acc2.c | 200 const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local 222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() 223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
|
D | 5x5s2p2-minmax-sse-1x4-acc5.c | 200 const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local 222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() 223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
|
D | 5x5s2p2-minmax-sse-1x4.c | 200 const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local 222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() 223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
|
D | 5x5s2p2-minmax-sse-1x4-acc4.c | 200 const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local 222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() 223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
|
D | 5x5s2p2-minmax-sse-1x4-acc3.c | 200 const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local 222 const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() 223 vi0x8ACE = vi0xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
|