src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_fp16_fp16.hpp

/*
 * Copyright (c) 2022-2023 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.
 */

#if defined(__ARM_FEATURE_SVE)

template <>
void interleave_block<2, 2, VLType::SME, false>(
  __fp16 * &out, const __fp16 * const *in,
  size_t width, size_t height, size_t row_offset, bool first
)
{
  ARM_COMPUTE_UNUSED(first);

  __asm__ __volatile__(
      ".inst 0xd503477f  // SMSTART ZA\n"
      "cnth x21\n"
      "mov x22, %x[width]\n"
      "inch x22\n"
      "mov x20, %x[width]\n"
      "sub x17, x21, #0x1\n"
      "sub x22, x22, #0x1\n"
      "ands x17, x20, x17\n"
      "cntw x16\n"
      "udiv x22, x22, x21\n"  // n_passes = ceildiv(width, VL<T>)
      "csel x17, x17, x21, NE\n"
      "sub x13, x22, #0x1\n"
      "add x17, x17, #0x1\n"
      "sub x15, x16, #0x2\n"
      "lsl x21, %x[height], #0x1\n"  // height * 2
      "lsl x20, x16, #0x1\n"
      "mov x14, #0x0\n"
      "mov x11, %x[in]\n"
      "add x10, %x[in], x16, LSL #3\n"
      "ldr x9, [x11, #0x0]\n"
      "cntw x28, ALL, MUL #2\n"
      "cntw x27, ALL, MUL #3\n"
      "ldr x26, [x10, #0x0]\n"
      "lsr x13, x13, #0x1\n"  // n_loops = (n_passes - 1) / 2
      "and x25, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
      "ldr x24, [x11, #0x8]\n"
      "lsr x17, x17, #0x1\n"
      "ptrue p13.s\n"
      "ldr x23, [x10, #0x8]\n"
      "whilelt p12.h, XZR, x21\n"
      "whilelt p11.h, x20, x21\n"
      "mov x22, %x[row_offset]\n"
      "mov x21, %x[out]\n"
      "whilelt p10.h, x14, %x[width]\n"
      "whilelt p9.h, x14, %x[width]\n"
      "whilelt p8.h, x14, %x[width]\n"
      "add x11, x11, #0x10\n"
      "add x10, x10, #0x10\n"
      "mov x12, #0x0\n"
      "cbz x15, 2f\n"
      "1:"  // K loop: Charge: Loop
      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
      ".inst 0xe0560520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x22, LSL #1]\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0xe0560348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x22, LSL #1]\n"
      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
      "ldr x26, [x10, #0x0]\n"
      ".inst 0xe0560702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x22, LSL #1]\n"
      "ldr x24, [x11, #0x8]\n"
      "add x11, x11, #0x10\n"
      ".inst 0xe05602ea  // ld1h { za1h.h[x12, #2] }, p0/Z, [x23, x22, LSL #1]\n"
      "add x12, x12, #0x4\n"
      "cmp x12, x15, LSL #1\n"
      "ldr x23, [x10, #0x8]\n"
      "add x10, x10, #0x10\n"
      "blt 1b\n"
      "2:"  // K loop: Charge: End
      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
      ".inst 0xe0560520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x22, LSL #1]\n"
      ".inst 0xe0560348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x22, LSL #1]\n"
      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
      "mov x11, %x[in]\n"
      "add x10, %x[in], x16, LSL #3\n"
      ".inst 0xe0560702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x22, LSL #1]\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0xe05602ea  // ld1h { za1h.h[x12, #2] }, p0/Z, [x23, x22, LSL #1]\n"
      "ldr x26, [x10, #0x0]\n"
      "inch x22\n"
      "inch x14\n"
      "ldr x24, [x11, #0x8]\n"
      "add x11, x11, #0x10\n"
      "ldr x23, [x10, #0x8]\n"
      "add x10, x10, #0x10\n"
      "cbz x13, 8f\n"
      "mov x20, x13\n"
      "3:"  // K loop: Main loop
      "whilelt p9.h, x14, %x[width]\n"
      "whilelt p8.h, x14, %x[width]\n"
      "mov x13, #0x0\n"
      "mov x12, #0x0\n"
      "cbz x15, 5f\n"
      "4:"  // K loop: Main loop: First: Loop
      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
      ".inst 0xe0562521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x22, LSL #1]\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0xe0562349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x22, LSL #1]\n"
      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
      ".inst 0x25796162  // psel p2.h, p8.h/Z, p11.h[w13, #3]\n"
      "ldr x26, [x10, #0x0]\n"
      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0562303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x22, LSL #1]\n"
      "ldr x24, [x11, #0x8]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0562aeb  // ld1h { za1h.h[x13, #3] }, p2/Z, [x23, x22, LSL #1]\n"
      "ldr x23, [x10, #0x8]\n"
      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
      "add x11, x11, #0x10\n"
      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
      "add x10, x10, #0x10\n"
      "add x13, x13, #0x4\n"
      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
      "add x12, x12, #0x2\n"
      "cmp x12, x15\n"
      "addvl x21, x21, #4\n"
      "blt 4b\n"
      "5:"  // K loop: Main loop: First: Tail
      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
      ".inst 0xe0562521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x22, LSL #1]\n"
      ".inst 0xe0562349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x22, LSL #1]\n"
      "mov x11, %x[in]\n"
      "add x10, %x[in], x16, LSL #3\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
      ".inst 0x25796161  // psel p1.h, p8.h/Z, p11.h[w13, #3]\n"
      ".inst 0xe0562303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x22, LSL #1]\n"
      "ldr x26, [x10, #0x0]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe05626eb  // ld1h { za1h.h[x13, #3] }, p1/Z, [x23, x22, LSL #1]\n"
      "ldr x24, [x11, #0x8]\n"
      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
      "ldr x23, [x10, #0x8]\n"
      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0xe0b08aa4  // st1w { za1v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
      "whilelt p10.h, x14, %x[width]\n"
      "inch x14\n"
      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
      "add x11, x11, #0x10\n"
      "add x10, x10, #0x10\n"
      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
      "addvl x21, x21, #4\n"
      "inch x22\n"
      "whilelt p9.h, x14, %x[width]\n"
      "whilelt p8.h, x14, %x[width]\n"
      "mov x13, #0x0\n"
      "mov x12, #0x0\n"
      "cbz x15, 7f\n"
      "6:"  // K loop: Main loop: Second: Loop
      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
      ".inst 0xe0562520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x22, LSL #1]\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0xe0562348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x22, LSL #1]\n"
      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
      ".inst 0x25696162  // psel p2.h, p8.h/Z, p11.h[w13, #2]\n"
      "ldr x26, [x10, #0x0]\n"
      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0562302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x22, LSL #1]\n"
      "ldr x24, [x11, #0x8]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0562aea  // ld1h { za1h.h[x13, #2] }, p2/Z, [x23, x22, LSL #1]\n"
      "ldr x23, [x10, #0x8]\n"
      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
      "add x11, x11, #0x10\n"
      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
      "add x10, x10, #0x10\n"
      "add x13, x13, #0x4\n"
      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
      "add x12, x12, #0x2\n"
      "cmp x12, x15\n"
      "addvl x21, x21, #4\n"
      "blt 6b\n"
      "7:"  // K loop: Main loop: Second: Tail
      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
      ".inst 0xe0562520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x22, LSL #1]\n"
      ".inst 0xe0562348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x22, LSL #1]\n"
      "mov x11, %x[in]\n"
      "add x10, %x[in], x16, LSL #3\n"
      "ldr x9, [x11, #0x0]\n"
      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
      ".inst 0xe0562302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x22, LSL #1]\n"
      "ldr x26, [x10, #0x0]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe05626ea  // ld1h { za1h.h[x13, #2] }, p1/Z, [x23, x22, LSL #1]\n"
      "ldr x24, [x11, #0x8]\n"
      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
      "ldr x23, [x10, #0x8]\n"
      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
      ".inst 0xe0b08aac  // st1w { za3v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
      "whilelt p10.h, x14, %x[width]\n"
      "subs x20, x20, #0x1\n"
      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
      "add x11, x11, #0x10\n"
      "add x10, x10, #0x10\n"
      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
      "addvl x21, x21, #4\n"
      "inch x14\n"
      "inch x22\n"
      "bgt 3b\n"
      "8:"  // K loop: Tails
      "cbnz x25, 11f\n"
      "mov x11, %x[in]\n"
      "whilelt p9.h, x14, %x[width]\n"
      "whilelt p8.h, x14, %x[width]\n"
      "mov x13, #0x0\n"
      "mov x12, #0x0\n"
      "9:"  // K loop: Tails: Even: First
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
      "ldr x9, [x11, #0x0]\n"
      "add x12, x12, #0x1\n"
      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
      "ldr x26, [x11, x16, LSL #0x3]\n"
      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
      "cmp x12, x16\n"
      ".inst 0xe0562521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x22, LSL #1]\n"
      ".inst 0xe0562349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x22, LSL #1]\n"
      "add x11, x11, #0x8\n"
      "addvl x21, x21, #2\n"
      "add x13, x13, #0x2\n"
      "blt 9b\n"
      "whilelt p10.h, x14, %x[width]\n"
      "whilelt p9.h, x14, %x[width]\n"
      "whilelt p8.h, x14, %x[width]\n"
      "mov x20, #0x0\n"
      "mov x12, #0x0\n"
      "10:"  // K loop: Tails: Even: Second
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
      "add x12, x12, #0x1\n"
      "cmp x12, x17\n"
      "addvl x21, x21, #2\n"
      "add x20, x20, #0x2\n"
      "blt 10b\n"
      "whilelt p10.h, x14, %x[width]\n"
      "b 13f\n"
      "11:"  // K loop: Tails: Odd
      "mov x12, #0x0\n"
      "12:"  // K loop: Tails: Odd: Loop
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
      "add x12, x12, #0x1\n"
      "cmp x12, x17\n"
      "addvl x21, x21, #2\n"
      "blt 12b\n"
      "13:"  // K loop: End
      "mov %x[out], x21\n"
      ".inst 0xd503467f  // SMSTOP\n"
      : [out] "+&r" (out)
      : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "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"
    );
}

#endif  // defined(__ARM_FEATURE_SVE)