diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp16_24x8.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp16_24x8.hpp | 2074 |
1 files changed, 2074 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp16_24x8.hpp b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp16_24x8.hpp new file mode 100644 index 0000000000..7bfab412ca --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp16_24x8.hpp @@ -0,0 +1,2074 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#pragma once + +#if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)) + +template<> +void MergeResults<24, 8, false>(__fp16 *out, const __fp16 *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const __fp16 *bias, Activation act, bool append) +{ + const __fp16 *inptr = in; + __fp16 nullbias[24] = { 0 }; + __fp16 minval = - static_cast<__fp16>(std::numeric_limits<float>::infinity()); + __fp16 maxval = static_cast<__fp16>(std::numeric_limits<float>::infinity()); + + switch(act.type) + { + default: + case Activation::Type::None: + break; + case Activation::Type::BoundedReLU: + maxval = static_cast<__fp16>(act.param1); + /* fall through */ + case Activation::Type::ReLU: + minval = 0.0f; + break; + } + + if (!append && !bias) + { + memset(nullbias, 0, (24 * sizeof(__fp16))); + } + + for (int y=y0; y<ymax; y+=8) + { + __fp16 *outptr0 = out + (y * ldout) + x0; + __fp16 *outptr1 = outptr0 + ldout; + __fp16 *outptr2 = outptr1 + ldout; + __fp16 *outptr3 = outptr2 + ldout; + __fp16 *outptr4 = outptr3 + ldout; + __fp16 *outptr5 = outptr4 + ldout; + __fp16 *outptr6 = outptr5 + ldout; + __fp16 *outptr7 = outptr6 + ldout; + + const int height = ymax - y; + + for (int i=x0; i<xmax; i+=24) + { + if (append) + { + switch(height) + { + case 1: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( +#ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + ".arch armv8.2-a+fp16\n" +#endif + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "add %[inptr], %[inptr], #0x180\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr0]]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "str q12, [%[outptr0], #0x20]\n" + "add %[outptr0], %[outptr0], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 2: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "str q13, [%[outptr1]]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "str q15, [%[outptr1], #0x20]\n" + "add %[outptr1], %[outptr1], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 3: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "add %[outptr2], %[outptr2], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 4: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + *outptr3)), maxval); + outptr3++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr3]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x90]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr3], #0x10]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0xa0]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr3], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0xb0]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q11, [%[outptr3]]\n" + "add %[inptr], %[inptr], #0x180\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr3], #0x10]\n" + "str q13, [%[outptr3], #0x20]\n" + "add %[outptr3], %[outptr3], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 5: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + *outptr3)), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + *outptr4)), maxval); + outptr4++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr3]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x90]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr3], #0x10]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0xa0]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr3], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0xb0]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr4]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0xc0]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q7, [%[outptr4], #0x10]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q15, [%[inptr], #0xd0]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q8, [%[outptr4], #0x20]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "ldr q16, [%[inptr], #0xe0]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q11, [%[outptr3]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "str q12, [%[outptr3], #0x10]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr3], #0x20]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr4]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "str q15, [%[outptr4], #0x10]\n" + "add %[inptr], %[inptr], #0x180\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "str q16, [%[outptr4], #0x20]\n" + "add %[outptr4], %[outptr4], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 6: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + *outptr3)), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + *outptr4)), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + *outptr5)), maxval); + outptr5++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr3]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x90]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr3], #0x10]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0xa0]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr3], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0xb0]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr4]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0xc0]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q7, [%[outptr4], #0x10]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q15, [%[inptr], #0xd0]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q8, [%[outptr4], #0x20]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "ldr q16, [%[inptr], #0xe0]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q9, [%[outptr5]]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xf0]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr5], #0x10]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr3]]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x100]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr5], #0x20]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr3], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x110]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr3], #0x20]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr4]]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "str q15, [%[outptr4], #0x10]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "str q16, [%[outptr4], #0x20]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "str q17, [%[outptr5]]\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "str q10, [%[outptr5], #0x10]\n" + "add %[outptr4], %[outptr4], #0x30\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "str q11, [%[outptr5], #0x20]\n" + "add %[outptr5], %[outptr5], #0x30\n" + "add %[inptr], %[inptr], #0x180\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 7: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + *outptr3)), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + *outptr4)), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + *outptr5)), maxval); + outptr5++; + *outptr6 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 144] + *outptr6)), maxval); + outptr6++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr3]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x90]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr3], #0x10]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0xa0]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr3], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0xb0]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr4]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0xc0]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q7, [%[outptr4], #0x10]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q15, [%[inptr], #0xd0]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q8, [%[outptr4], #0x20]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "ldr q16, [%[inptr], #0xe0]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q9, [%[outptr5]]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xf0]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr5], #0x10]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr3]]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x100]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr5], #0x20]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr3], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x110]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr6]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr3], #0x20]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0x120]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr6], #0x10]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr4]]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0x130]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr6], #0x20]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr4], #0x10]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0x140]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr4], #0x20]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "str q17, [%[outptr5]]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr5], #0x10]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q11, [%[outptr5], #0x20]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr6]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "str q13, [%[outptr6], #0x10]\n" + "add %[outptr4], %[outptr4], #0x30\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add %[outptr5], %[outptr5], #0x30\n" + "str q14, [%[outptr6], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add %[outptr6], %[outptr6], #0x30\n" + "add %[inptr], %[inptr], #0x180\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + default: + case 8: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + *outptr0)), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + *outptr1)), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + *outptr2)), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + *outptr3)), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + *outptr4)), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + *outptr5)), maxval); + outptr5++; + *outptr6 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 144] + *outptr6)), maxval); + outptr6++; + *outptr7 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 168] + *outptr7)), maxval); + outptr7++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[outptr0]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q10, [%[inptr]]\n" + "ldr q3, [%[outptr0], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q11, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q4, [%[outptr0], #0x20]\n" + "ldr q12, [%[inptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q5, [%[outptr1]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x30]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q6, [%[outptr1], #0x10]\n" + "ldr q14, [%[inptr], #0x40]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q7, [%[outptr1], #0x20]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q15, [%[inptr], #0x50]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "ldr q8, [%[outptr2]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "str q10, [%[outptr0]]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q16, [%[inptr], #0x60]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "ldr q9, [%[outptr2], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x70]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr2], #0x20]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr0], #0x10]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x80]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr3]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr0], #0x20]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x90]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr3], #0x10]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr1]]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0xa0]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr3], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr1], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0xb0]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr4]]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr1], #0x20]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0xc0]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q7, [%[outptr4], #0x10]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr2]]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q15, [%[inptr], #0xd0]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q8, [%[outptr4], #0x20]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "str q17, [%[outptr2], #0x10]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "ldr q16, [%[inptr], #0xe0]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q9, [%[outptr5]]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr2], #0x20]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xf0]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q2, [%[outptr5], #0x10]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr3]]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "ldr q10, [%[inptr], #0x100]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "ldr q3, [%[outptr5], #0x20]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr3], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q11, [%[inptr], #0x110]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q4, [%[outptr6]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr3], #0x20]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "ldr q12, [%[inptr], #0x120]\n" + "fadd v10.8h, v10.8h, v2.8h\n" + "ldr q5, [%[outptr6], #0x10]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr4]]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q13, [%[inptr], #0x130]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q6, [%[outptr6], #0x20]\n" + "fmin v10.8h, v10.8h, v0.8h\n" + "str q15, [%[outptr4], #0x10]\n" + "fadd v11.8h, v11.8h, v3.8h\n" + "ldr q14, [%[inptr], #0x140]\n" + "fadd v12.8h, v12.8h, v4.8h\n" + "ldr q7, [%[outptr7]]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr4], #0x20]\n" + "fmax v10.8h, v10.8h, v1.8h\n" + "ldr q15, [%[inptr], #0x150]\n" + "fmin v11.8h, v11.8h, v0.8h\n" + "ldr q8, [%[outptr7], #0x10]\n" + "fmin v12.8h, v12.8h, v0.8h\n" + "str q17, [%[outptr5]]\n" + "fadd v13.8h, v13.8h, v5.8h\n" + "ldr q16, [%[inptr], #0x160]\n" + "fadd v14.8h, v14.8h, v6.8h\n" + "ldr q9, [%[outptr7], #0x20]\n" + "fmax v11.8h, v11.8h, v1.8h\n" + "str q10, [%[outptr5], #0x10]\n" + "fmax v12.8h, v12.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x170]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q11, [%[outptr5], #0x20]\n" + "fadd v15.8h, v15.8h, v7.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q12, [%[outptr6]]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr6], #0x10]\n" + "fadd v16.8h, v16.8h, v8.8h\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "fadd v17.8h, v17.8h, v9.8h\n" + "str q14, [%[outptr6], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "str q15, [%[outptr7]]\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "str q16, [%[outptr7], #0x10]\n" + "add %[outptr4], %[outptr4], #0x30\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "str q17, [%[outptr7], #0x20]\n" + "add %[outptr5], %[outptr5], #0x30\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add %[outptr6], %[outptr6], #0x30\n" + "prfm PLDL1KEEP, [%[outptr7], #0x60]\n" + "add %[outptr7], %[outptr7], #0x30\n" + "add %[inptr], %[inptr], #0x180\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + + } + } + else + { + const __fp16 *biasptr = nullbias; + if (bias) + { + biasptr = bias + i; + } + + switch(height) + { + case 1: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "ldr q15, [%[inptr], #0x20]\n" + "add %[inptr], %[inptr], #0x180\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q13, [%[outptr0]]\n" + "str q14, [%[outptr0], #0x10]\n" + "str q15, [%[outptr0], #0x20]\n" + "add %[outptr0], %[outptr0], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 2: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "str q13, [%[outptr0]]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr1]]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "str q18, [%[outptr1], #0x20]\n" + "add %[outptr1], %[outptr1], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 3: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr1]]\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "str q19, [%[outptr2]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "str q13, [%[outptr2], #0x20]\n" + "add %[outptr2], %[outptr2], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 4: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + biasptr[xi])), maxval); + outptr3++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x90]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "ldr q15, [%[inptr], #0xa0]\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr1]]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "ldr q16, [%[inptr], #0xb0]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "fadd v14.8h, v14.8h, v2.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q19, [%[outptr2]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "fadd v15.8h, v15.8h, v3.8h\n" + "fadd v16.8h, v16.8h, v4.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr2], #0x20]\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr3]]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "str q15, [%[outptr3], #0x10]\n" + "str q16, [%[outptr3], #0x20]\n" + "add %[outptr3], %[outptr3], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 5: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + biasptr[xi])), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + biasptr[xi])), maxval); + outptr4++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x90]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "ldr q15, [%[inptr], #0xa0]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q16, [%[outptr1]]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "ldr q16, [%[inptr], #0xb0]\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xc0]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "fadd v14.8h, v14.8h, v2.8h\n" + "ldr q18, [%[inptr], #0xd0]\n" + "fadd v15.8h, v15.8h, v3.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "str q19, [%[outptr2]]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "ldr q19, [%[inptr], #0xe0]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "fadd v16.8h, v16.8h, v4.8h\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fadd v17.8h, v17.8h, v2.8h\n" + "str q13, [%[outptr2], #0x20]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q14, [%[outptr3]]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "fadd v18.8h, v18.8h, v3.8h\n" + "fadd v19.8h, v19.8h, v4.8h\n" + "str q15, [%[outptr3], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "str q16, [%[outptr3], #0x20]\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr4]]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "str q18, [%[outptr4], #0x10]\n" + "str q19, [%[outptr4], #0x20]\n" + "add %[outptr4], %[outptr4], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 6: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + biasptr[xi])), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + biasptr[xi])), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + biasptr[xi])), maxval); + outptr5++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x90]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "ldr q15, [%[inptr], #0xa0]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q16, [%[outptr1]]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "ldr q16, [%[inptr], #0xb0]\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xc0]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "fadd v14.8h, v14.8h, v2.8h\n" + "ldr q18, [%[inptr], #0xd0]\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q19, [%[outptr2]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "ldr q19, [%[inptr], #0xe0]\n" + "fadd v15.8h, v15.8h, v3.8h\n" + "fadd v16.8h, v16.8h, v4.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "ldr q20, [%[inptr], #0xf0]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q13, [%[outptr2], #0x20]\n" + "fadd v17.8h, v17.8h, v2.8h\n" + "ldr q13, [%[inptr], #0x100]\n" + "fadd v18.8h, v18.8h, v3.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr3]]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x110]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q15, [%[outptr3], #0x10]\n" + "fadd v19.8h, v19.8h, v4.8h\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "fadd v20.8h, v20.8h, v2.8h\n" + "str q16, [%[outptr3], #0x20]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "str q17, [%[outptr4]]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "fadd v13.8h, v13.8h, v3.8h\n" + "fadd v14.8h, v14.8h, v4.8h\n" + "str q18, [%[outptr4], #0x10]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q19, [%[outptr4], #0x20]\n" + "add %[outptr4], %[outptr4], #0x30\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q20, [%[outptr5]]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "str q13, [%[outptr5], #0x10]\n" + "str q14, [%[outptr5], #0x20]\n" + "add %[outptr5], %[outptr5], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + case 7: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + biasptr[xi])), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + biasptr[xi])), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + biasptr[xi])), maxval); + outptr5++; + *outptr6 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 144] + biasptr[xi])), maxval); + outptr6++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x90]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "ldr q15, [%[inptr], #0xa0]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q16, [%[outptr1]]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "ldr q16, [%[inptr], #0xb0]\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xc0]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "fadd v14.8h, v14.8h, v2.8h\n" + "ldr q18, [%[inptr], #0xd0]\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q19, [%[outptr2]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "ldr q19, [%[inptr], #0xe0]\n" + "fadd v15.8h, v15.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q20, [%[inptr], #0xf0]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "fadd v16.8h, v16.8h, v4.8h\n" + "str q13, [%[outptr2], #0x20]\n" + "fadd v17.8h, v17.8h, v2.8h\n" + "ldr q13, [%[inptr], #0x100]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q14, [%[outptr3]]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q14, [%[inptr], #0x110]\n" + "fadd v18.8h, v18.8h, v3.8h\n" + "fadd v19.8h, v19.8h, v4.8h\n" + "str q15, [%[outptr3], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q15, [%[inptr], #0x120]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "str q16, [%[outptr3], #0x20]\n" + "fadd v20.8h, v20.8h, v2.8h\n" + "ldr q16, [%[inptr], #0x130]\n" + "fadd v13.8h, v13.8h, v3.8h\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr4]]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "ldr q17, [%[inptr], #0x140]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q18, [%[outptr4], #0x10]\n" + "fadd v14.8h, v14.8h, v4.8h\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "fadd v15.8h, v15.8h, v2.8h\n" + "str q19, [%[outptr4], #0x20]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "add %[outptr4], %[outptr4], #0x30\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q20, [%[outptr5]]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "fadd v16.8h, v16.8h, v3.8h\n" + "fadd v17.8h, v17.8h, v4.8h\n" + "str q13, [%[outptr5], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "str q14, [%[outptr5], #0x20]\n" + "add %[outptr5], %[outptr5], #0x30\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "str q15, [%[outptr6]]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "str q16, [%[outptr6], #0x10]\n" + "str q17, [%[outptr6], #0x20]\n" + "add %[outptr6], %[outptr6], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + default: + case 8: + { + if ((i+23) >= xmax) + { + for (int xi=0; xi<23; xi++) + { + if ((i+xi) < xmax) + { + *outptr0 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi] + biasptr[xi])), maxval); + outptr0++; + *outptr1 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 24] + biasptr[xi])), maxval); + outptr1++; + *outptr2 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 48] + biasptr[xi])), maxval); + outptr2++; + *outptr3 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 72] + biasptr[xi])), maxval); + outptr3++; + *outptr4 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 96] + biasptr[xi])), maxval); + outptr4++; + *outptr5 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 120] + biasptr[xi])), maxval); + outptr5++; + *outptr6 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 144] + biasptr[xi])), maxval); + outptr6++; + *outptr7 = std::min(std::max(minval, static_cast<__fp16>(inptr[xi + 168] + biasptr[xi])), maxval); + outptr7++; + } + } + inptr += 192; + } else { + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "dup v0.8h, %[maxval].h[0]\n" + "ldr q2, [%[biasptr]]\n" + "dup v1.8h, %[minval].h[0]\n" + "ldr q3, [%[biasptr], #0x10]\n" + "ldr q4, [%[biasptr], #0x20]\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ldr q13, [%[inptr]]\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ldr q14, [%[inptr], #0x10]\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "fadd v13.8h, v13.8h, v2.8h\n" + "ldr q15, [%[inptr], #0x20]\n" + "ldr q16, [%[inptr], #0x30]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "fadd v14.8h, v14.8h, v3.8h\n" + "ldr q17, [%[inptr], #0x40]\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "ldr q18, [%[inptr], #0x50]\n" + "fadd v15.8h, v15.8h, v4.8h\n" + "ldr q19, [%[inptr], #0x60]\n" + "fadd v16.8h, v16.8h, v2.8h\n" + "ldr q20, [%[inptr], #0x70]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "str q13, [%[outptr0]]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "ldr q13, [%[inptr], #0x80]\n" + "fadd v17.8h, v17.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "str q14, [%[outptr0], #0x10]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "ldr q14, [%[inptr], #0x90]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "fadd v18.8h, v18.8h, v4.8h\n" + "str q15, [%[outptr0], #0x20]\n" + "fadd v19.8h, v19.8h, v2.8h\n" + "ldr q15, [%[inptr], #0xa0]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "add %[outptr0], %[outptr0], #0x30\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "str q16, [%[outptr1]]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "ldr q16, [%[inptr], #0xb0]\n" + "fadd v20.8h, v20.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr1], #0x10]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "ldr q17, [%[inptr], #0xc0]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "fadd v13.8h, v13.8h, v4.8h\n" + "str q18, [%[outptr1], #0x20]\n" + "fadd v14.8h, v14.8h, v2.8h\n" + "ldr q18, [%[inptr], #0xd0]\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "add %[outptr1], %[outptr1], #0x30\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q19, [%[outptr2]]\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "ldr q19, [%[inptr], #0xe0]\n" + "fadd v15.8h, v15.8h, v3.8h\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "str q20, [%[outptr2], #0x10]\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "ldr q20, [%[inptr], #0xf0]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "fadd v16.8h, v16.8h, v4.8h\n" + "str q13, [%[outptr2], #0x20]\n" + "fadd v17.8h, v17.8h, v2.8h\n" + "ldr q13, [%[inptr], #0x100]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "add %[outptr2], %[outptr2], #0x30\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "str q14, [%[outptr3]]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "ldr q14, [%[inptr], #0x110]\n" + "fadd v18.8h, v18.8h, v3.8h\n" + "prfm PSTL1KEEP, [%[outptr7], #0x60]\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "str q15, [%[outptr3], #0x10]\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "ldr q15, [%[inptr], #0x120]\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "fadd v19.8h, v19.8h, v4.8h\n" + "str q16, [%[outptr3], #0x20]\n" + "fadd v20.8h, v20.8h, v2.8h\n" + "ldr q16, [%[inptr], #0x130]\n" + "fadd v13.8h, v13.8h, v3.8h\n" + "add %[outptr3], %[outptr3], #0x30\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "str q17, [%[outptr4]]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "ldr q17, [%[inptr], #0x140]\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "fmin v13.8h, v13.8h, v0.8h\n" + "str q18, [%[outptr4], #0x10]\n" + "fadd v14.8h, v14.8h, v4.8h\n" + "ldr q18, [%[inptr], #0x150]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "fmax v13.8h, v13.8h, v1.8h\n" + "fmin v14.8h, v14.8h, v0.8h\n" + "str q19, [%[outptr4], #0x20]\n" + "fadd v15.8h, v15.8h, v2.8h\n" + "ldr q19, [%[inptr], #0x160]\n" + "fadd v16.8h, v16.8h, v3.8h\n" + "add %[outptr4], %[outptr4], #0x30\n" + "fmax v14.8h, v14.8h, v1.8h\n" + "str q20, [%[outptr5]]\n" + "fmin v15.8h, v15.8h, v0.8h\n" + "ldr q20, [%[inptr], #0x170]\n" + "fmin v16.8h, v16.8h, v0.8h\n" + "add %[inptr], %[inptr], #0x180\n" + "fadd v17.8h, v17.8h, v4.8h\n" + "str q13, [%[outptr5], #0x10]\n" + "fmax v15.8h, v15.8h, v1.8h\n" + "fmax v16.8h, v16.8h, v1.8h\n" + "fadd v18.8h, v18.8h, v2.8h\n" + "str q14, [%[outptr5], #0x20]\n" + "fmin v17.8h, v17.8h, v0.8h\n" + "add %[outptr5], %[outptr5], #0x30\n" + "fadd v19.8h, v19.8h, v3.8h\n" + "str q15, [%[outptr6]]\n" + "fmin v18.8h, v18.8h, v0.8h\n" + "fmax v17.8h, v17.8h, v1.8h\n" + "fadd v20.8h, v20.8h, v4.8h\n" + "str q16, [%[outptr6], #0x10]\n" + "fmin v19.8h, v19.8h, v0.8h\n" + "fmax v18.8h, v18.8h, v1.8h\n" + "fmin v20.8h, v20.8h, v0.8h\n" + "str q17, [%[outptr6], #0x20]\n" + "fmax v19.8h, v19.8h, v1.8h\n" + "add %[outptr6], %[outptr6], #0x30\n" + "fmax v20.8h, v20.8h, v1.8h\n" + "str q18, [%[outptr7]]\n" + "str q19, [%[outptr7], #0x10]\n" + "str q20, [%[outptr7], #0x20]\n" + "add %[outptr7], %[outptr7], #0x30\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [biasptr] "r" (biasptr), [minval] "w" (minval), [maxval] "w" (maxval) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "memory" + ); + } + } + break; + + + } + } + } + } +} + +#endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC) |