/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-neon-2x4.c | 112 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local 177 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-2x4.c | 113 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local 177 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-2x4.c | 113 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local 177 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local
|
D | 3x3p1-minmax-neonfma-2x4.c | 112 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local 177 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
|
D | 5x5p2-minmax-neonfma-1x4.c | 136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 243 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 343 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc2.c | 136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 244 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 345 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
|
D | 5x5p2-minmax-neon-1x4.c | 136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 243 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 343 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 244 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 344 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
|
D | 5x5p2-minmax-neon-1x4-acc2.c | 136 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 244 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 345 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 243 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 342 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 244 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 344 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 243 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 342 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 113 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local 179 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 113 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local 179 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
|
D | 3x3p1-minmax-neonfma-2x4-acc2.c | 112 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local 179 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
|
D | 3x3p1-minmax-neon-2x4-acc2.c | 112 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local 179 const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c | 125 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local 182 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local
|
D | 3x3p1-minmax-ssse3-2x4.c | 120 …const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local 178 … const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c | 125 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local 182 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 245 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 346 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 137 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 246 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 348 const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
/external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neonfp16arith-1x4.c | 134 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local 241 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local 341 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local
|
D | 5x5p2-minmax-neonfp16arith-1x4-acc2.c | 134 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local 242 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local 343 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local
|
D | 5x5p2-minmax-neonfp16arith-1x4-acc3.c | 134 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local 243 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local 345 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local
|
D | 5x5p2-minmax-neonfp16arith-1x4-acc4.c | 134 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local 244 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local 347 const float16x4_t vi3x5678 = vext_f16(vi3x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local
|