/aosp_15_r20/external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
H A D | 5x5p2-minmax-neonfp16arith-1x4-acc3.c | 135 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local 244 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local 346 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local
|
H A D | 5x5p2-minmax-neonfp16arith-1x4.c | 135 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local 242 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local 342 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local
|
H A D | 5x5p2-minmax-neonfp16arith-1x4-acc2.c | 135 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local 243 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local 344 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local
|
H A D | 5x5p2-minmax-neonfp16arith-1x4-acc4.c | 135 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local 245 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local 348 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local
|
H A D | 5x5p2-minmax-neonfp16arith-1x4-acc5.c | 135 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc5() local 246 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc5() local 350 const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc5() local
|
/aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
H A D | 5x5p2-minmax-neonfma-1x4-acc2.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 245 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local 346 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
|
H A D | 5x5p2-minmax-neon-1x4.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 244 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local 344 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
|
H A D | 5x5p2-minmax-neon-1x4-acc2.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 245 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local 346 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 245 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local 345 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
|
H A D | 5x5p2-minmax-neonfma-1x4.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 244 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local 344 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 244 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local 343 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 245 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local 345 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc2() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 244 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local 343 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 247 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 349 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 246 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local 347 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc3() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 247 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 349 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
H A D | 5x5p2-minmax-neon-1x4-acc3.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 246 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local 348 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 248 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 351 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local
|
H A D | 5x5p2-minmax-neonfma-1x4-acc4.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 247 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 350 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
|
H A D | 5x5p2-minmax-neonfma-1x4-acc3.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 246 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local 348 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 248 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 351 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 138 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 246 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local 347 const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
|
H A D | 5x5p2-minmax-neon-1x4-acc4.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 247 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 350 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
|
H A D | 5x5p2-minmax-neon-1x4-acc5.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 248 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local 352 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc5() local
|
H A D | 5x5p2-minmax-neonfma-1x4-acc5.c | 137 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 248 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 352 const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
|