/* * Copyright (c) 2021, 2023-2024 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(ARM_COMPUTE_ENABLE_SVE) namespace { void sve_transpose_interleave_8VL_1x4(uint8_t *out, const uint8_t *in, size_t width, size_t in_stride, size_t height) { uint8_t *pad_row = reinterpret_cast(alloca(width * sizeof(uint8_t))); if (height % 4) { memset(pad_row, 0, width * sizeof(uint8_t)); } size_t out_stride = 8 * roundup(height, 4) * get_vector_length(); __asm__ __volatile__( "ptrue p2.b\n" "1:" // Main row loop: Head "mov x26, %x[in]\n" "mov x25, %x[width]\n" "cntb x20, ALL, MUL #8\n" "cmp %x[height], #0x3\n" "mov x24, %x[out]\n" "add x23, x26, %x[in_stride]\n" "add x22, x23, %x[in_stride]\n" "add x21, x22, %x[in_stride]\n" "add %x[in], x21, %x[in_stride]\n" "csel x21, x21, %x[pad_row], GT\n" "csel x22, x22, %x[pad_row], GE\n" "cmp %x[height], #0x1\n" "csel x23, x23, %x[pad_row], GT\n" "cmp x25, x20\n" "sub %x[height], %x[height], #0x4\n" "blt 3f\n" "2:" // Main row loop: Unroll column loop "ld1b { z22.b }, p2/Z, [x26]\n" "ld1b { z14.b }, p2/Z, [x26, #1, MUL VL]\n" "sub x25, x25, x20\n" "ld1b { z2.b }, p2/Z, [x23]\n" "ld1b { z10.b }, p2/Z, [x23, #1, MUL VL]\n" "cmp x25, x20\n" "ld1b { z27.b }, p2/Z, [x22]\n" "ld1b { z21.b }, p2/Z, [x22, #1, MUL VL]\n" "ld1b { z13.b }, p2/Z, [x21]\n" "ld1b { z28.b }, p2/Z, [x21, #1, MUL VL]\n" "ld1b { z11.b }, p2/Z, [x26, #2, MUL VL]\n" "ld1b { z9.b }, p2/Z, [x26, #3, MUL VL]\n" "ld1b { z23.b }, p2/Z, [x23, #2, MUL VL]\n" "ld1b { z29.b }, p2/Z, [x23, #3, MUL VL]\n" "zip1 z24.b, z22.b, z27.b\n" "zip2 z19.b, z22.b, z27.b\n" "ld1b { z5.b }, p2/Z, [x22, #2, MUL VL]\n" "ld1b { z3.b }, p2/Z, [x22, #3, MUL VL]\n" "zip1 z6.b, z2.b, z13.b\n" "zip2 z20.b, z2.b, z13.b\n" "ld1b { z25.b }, p2/Z, [x21, #2, MUL VL]\n" "ld1b { z30.b }, p2/Z, [x21, #3, MUL VL]\n" "zip1 z18.b, z14.b, z21.b\n" "zip1 z16.b, z10.b, z28.b\n" "ld1b { z26.b }, p2/Z, [x26, #4, MUL VL]\n" "ld1b { z22.b }, p2/Z, [x26, #5, MUL VL]\n" "zip2 z17.b, z14.b, z21.b\n" "zip2 z10.b, z10.b, z28.b\n" "ld1b { z27.b }, p2/Z, [x23, #4, MUL VL]\n" "ld1b { z7.b }, p2/Z, [x23, #5, MUL VL]\n" "zip1 z21.b, z11.b, z5.b\n" "zip2 z12.b, z11.b, z5.b\n" "ld1b { z2.b }, p2/Z, [x22, #4, MUL VL]\n" "ld1b { z5.b }, p2/Z, [x22, #5, MUL VL]\n" "zip1 z13.b, z23.b, z25.b\n" "zip2 z31.b, z23.b, z25.b\n" "ld1b { z25.b }, p2/Z, [x21, #4, MUL VL]\n" "ld1b { z8.b }, p2/Z, [x21, #5, MUL VL]\n" "zip1 z15.b, z9.b, z3.b\n" "zip1 z11.b, z29.b, z30.b\n" "ld1b { z14.b }, p2/Z, [x26, #6, MUL VL]\n" "ld1b { z1.b }, p2/Z, [x26, #7, MUL VL]\n" "zip2 z23.b, z9.b, z3.b\n" "zip2 z0.b, z29.b, z30.b\n" "ld1b { z3.b }, p2/Z, [x23, #6, MUL VL]\n" "ld1b { z29.b }, p2/Z, [x23, #7, MUL VL]\n" "zip1 z9.b, z26.b, z2.b\n" "zip2 z26.b, z26.b, z2.b\n" "ld1b { z2.b }, p2/Z, [x22, #6, MUL VL]\n" "ld1b { z4.b }, p2/Z, [x22, #7, MUL VL]\n" "zip1 z30.b, z27.b, z25.b\n" "zip2 z28.b, z27.b, z25.b\n" "ld1b { z27.b }, p2/Z, [x21, #6, MUL VL]\n" "zip1 z25.b, z22.b, z5.b\n" "zip2 z5.b, z22.b, z5.b\n" "addvl x26, x26, #8\n" "zip1 z22.b, z7.b, z8.b\n" "zip2 z7.b, z7.b, z8.b\n" "addvl x23, x23, #8\n" "addvl x22, x22, #8\n" "zip1 z8.b, z24.b, z6.b\n" "zip2 z6.b, z24.b, z6.b\n" "ld1b { z24.b }, p2/Z, [x21, #7, MUL VL]\n" "addvl x21, x21, #8\n" "st1b { z8.b }, p2, [x24]\n" "zip1 z8.b, z19.b, z20.b\n" "zip2 z20.b, z19.b, z20.b\n" "zip1 z19.b, z18.b, z16.b\n" "zip2 z18.b, z18.b, z16.b\n" "zip1 z16.b, z17.b, z10.b\n" "zip2 z17.b, z17.b, z10.b\n" "st1b { z6.b }, p2, [x24, #1, MUL VL]\n" "st1b { z8.b }, p2, [x24, #2, MUL VL]\n" "zip1 z8.b, z14.b, z2.b\n" "zip1 z6.b, z3.b, z27.b\n" "st1b { z20.b }, p2, [x24, #3, MUL VL]\n" "zip2 z10.b, z14.b, z2.b\n" "zip2 z14.b, z3.b, z27.b\n" "st1b { z19.b }, p2, [x24, #4, MUL VL]\n" "zip1 z2.b, z1.b, z4.b\n" "zip1 z3.b, z29.b, z24.b\n" "st1b { z18.b }, p2, [x24, #5, MUL VL]\n" "zip2 z1.b, z1.b, z4.b\n" "zip2 z4.b, z29.b, z24.b\n" "st1b { z16.b }, p2, [x24, #6, MUL VL]\n" "zip1 z24.b, z21.b, z13.b\n" "zip2 z16.b, z21.b, z13.b\n" "st1b { z17.b }, p2, [x24, #7, MUL VL]\n" "add x24, x24, %x[out_stride]\n" "zip1 z29.b, z12.b, z31.b\n" "zip2 z21.b, z12.b, z31.b\n" "zip1 z20.b, z15.b, z11.b\n" "zip2 z19.b, z15.b, z11.b\n" "zip1 z18.b, z23.b, z0.b\n" "zip2 z17.b, z23.b, z0.b\n" "st1b { z24.b }, p2, [x24]\n" "st1b { z16.b }, p2, [x24, #1, MUL VL]\n" "zip1 z16.b, z9.b, z30.b\n" "zip2 z30.b, z9.b, z30.b\n" "st1b { z29.b }, p2, [x24, #2, MUL VL]\n" "zip1 z29.b, z26.b, z28.b\n" "zip2 z28.b, z26.b, z28.b\n" "st1b { z21.b }, p2, [x24, #3, MUL VL]\n" "zip1 z27.b, z25.b, z22.b\n" "zip2 z26.b, z25.b, z22.b\n" "st1b { z20.b }, p2, [x24, #4, MUL VL]\n" "zip1 z25.b, z5.b, z7.b\n" "zip2 z24.b, z5.b, z7.b\n" "st1b { z19.b }, p2, [x24, #5, MUL VL]\n" "zip1 z23.b, z8.b, z6.b\n" "zip2 z22.b, z8.b, z6.b\n" "st1b { z18.b }, p2, [x24, #6, MUL VL]\n" "zip1 z21.b, z10.b, z14.b\n" "zip2 z20.b, z10.b, z14.b\n" "st1b { z17.b }, p2, [x24, #7, MUL VL]\n" "add x24, x24, %x[out_stride]\n" "zip1 z19.b, z2.b, z3.b\n" "zip2 z18.b, z2.b, z3.b\n" "st1b { z16.b }, p2, [x24]\n" "zip1 z17.b, z1.b, z4.b\n" "zip2 z16.b, z1.b, z4.b\n" "st1b { z30.b }, p2, [x24, #1, MUL VL]\n" "st1b { z29.b }, p2, [x24, #2, MUL VL]\n" "st1b { z28.b }, p2, [x24, #3, MUL VL]\n" "st1b { z27.b }, p2, [x24, #4, MUL VL]\n" "st1b { z26.b }, p2, [x24, #5, MUL VL]\n" "st1b { z25.b }, p2, [x24, #6, MUL VL]\n" "st1b { z24.b }, p2, [x24, #7, MUL VL]\n" "add x24, x24, %x[out_stride]\n" "st1b { z23.b }, p2, [x24]\n" "st1b { z22.b }, p2, [x24, #1, MUL VL]\n" "st1b { z21.b }, p2, [x24, #2, MUL VL]\n" "st1b { z20.b }, p2, [x24, #3, MUL VL]\n" "st1b { z19.b }, p2, [x24, #4, MUL VL]\n" "st1b { z18.b }, p2, [x24, #5, MUL VL]\n" "st1b { z17.b }, p2, [x24, #6, MUL VL]\n" "st1b { z16.b }, p2, [x24, #7, MUL VL]\n" "add x24, x24, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip "cbz x25, 5f\n" "4:" // Main row loop: Column loop "mov x20, x25\n" "decw x25, ALL, MUL #8\n" "whilelt p1.b, XZR, x20\n" "decb x20\n" "whilelt p0.b, XZR, x20\n" "cmp x25, #0x0\n" "ld1b { z20.b }, p1/Z, [x26]\n" "ld1b { z19.b }, p1/Z, [x23]\n" "ld1b { z18.b }, p1/Z, [x22]\n" "ld1b { z17.b }, p1/Z, [x21]\n" "ld1b { z24.b }, p0/Z, [x26, #1, MUL VL]\n" "addvl x26, x26, #2\n" "ld1b { z25.b }, p0/Z, [x23, #1, MUL VL]\n" "addvl x23, x23, #2\n" "ld1b { z23.b }, p0/Z, [x22, #1, MUL VL]\n" "ld1b { z16.b }, p0/Z, [x21, #1, MUL VL]\n" "addvl x22, x22, #2\n" "addvl x21, x21, #2\n" "zip1 z22.b, z20.b, z18.b\n" "zip1 z21.b, z19.b, z17.b\n" "zip2 z20.b, z20.b, z18.b\n" "zip2 z19.b, z19.b, z17.b\n" "zip1 z18.b, z24.b, z23.b\n" "zip1 z17.b, z25.b, z16.b\n" "zip2 z24.b, z24.b, z23.b\n" "zip2 z16.b, z25.b, z16.b\n" "zip1 z23.b, z22.b, z21.b\n" "zip2 z22.b, z22.b, z21.b\n" "zip1 z21.b, z20.b, z19.b\n" "zip2 z20.b, z20.b, z19.b\n" "zip1 z19.b, z18.b, z17.b\n" "zip2 z18.b, z18.b, z17.b\n" "zip1 z17.b, z24.b, z16.b\n" "zip2 z16.b, z24.b, z16.b\n" "st1b { z23.b }, p2, [x24]\n" "st1b { z22.b }, p2, [x24, #1, MUL VL]\n" "st1b { z21.b }, p2, [x24, #2, MUL VL]\n" "st1b { z20.b }, p2, [x24, #3, MUL VL]\n" "st1b { z19.b }, p2, [x24, #4, MUL VL]\n" "st1b { z18.b }, p2, [x24, #5, MUL VL]\n" "st1b { z17.b }, p2, [x24, #6, MUL VL]\n" "st1b { z16.b }, p2, [x24, #7, MUL VL]\n" "add x24, x24, %x[out_stride]\n" "bgt 4b\n" "5:" // Main row loop: Column loop skip "cmp %x[height], #0x1\n" "addvl %x[out], %x[out], #8\n" "bge 1b\n" : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out) : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width) : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } } // anonymous namespace template<> void Transform<8, 4, true, VLType::SVE>( uint8_t *out, const uint8_t *in, int stride, int x0, int xmax, int k0, int kmax) { sve_transpose_interleave_8VL_1x4( reinterpret_cast(out), reinterpret_cast(in + k0 * stride + x0), (xmax-x0) * sizeof(uint8_t) / 1, stride * sizeof(uint8_t), (kmax-k0) ); } template<> void Transform<8, 4, true, VLType::SVE>( int8_t *out, const int8_t *in, int stride, int x0, int xmax, int k0, int kmax) { sve_transpose_interleave_8VL_1x4( reinterpret_cast(out), reinterpret_cast(in + k0 * stride + x0), (xmax-x0) * sizeof(int8_t) / 1, stride * sizeof(int8_t), (kmax-k0) ); } #endif // defined(ARM_COMPUTE_ENABLE_SVE)