diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp | 297 |
1 files changed, 297 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp new file mode 100644 index 0000000000..905c6b41eb --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp @@ -0,0 +1,297 @@ +/* + * 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_COMPUTE_ENABLE_SME) + +template <> +void interleave_block<2, 4, VLType::SME, false>( + int8_t * &out, const int8_t * const *in, + size_t width, size_t height, size_t row_offset, bool +) +{ + __asm__ __volatile__( + ".inst 0xd503477f // SMSTART ZA\n" + "cntb x21\n" + "mov x23, %x[width]\n" + "incb x23\n" + "mov x20, %x[width]\n" + "sub x17, x21, #0x1\n" + "cntw x16\n" + "sub x23, x23, #0x1\n" + "ands x17, x20, x17\n" + "udiv x23, x23, x21\n" // n_passes = ceildiv(width, VL<T>) + "csel x17, x17, x21, NE\n" + "lsl x22, %x[height], #0x1\n" // height * 2 + "lsl x21, x16, #0x1\n" + "sub x20, x23, #0x1\n" + "add x17, x17, #0x3\n" + "sub x15, x16, #0x2\n" + "whilelt p9.b, XZR, x22\n" + "whilelt p8.b, x21, x22\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 x20, x20, #0x1\n" // n_loops = (n_passes - 1) / 2 + "and x25, x23, #0x1\n" // odd_tail = bool(n_passes & 0x1) + "ldr x24, [x11, #0x8]\n" + "lsr x17, x17, #0x2\n" + "ptrue p11.s\n" + "ldr x23, [x10, #0x8]\n" + "zip1 p10.b, p9.b, p8.b\n" + "mov x22, %x[row_offset]\n" + "mov x21, %x[out]\n" + "whilelt p9.b, x14, %x[width]\n" + "whilelt p8.b, 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 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n" + ".inst 0xe0160120 // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n" + ".inst 0x252c6140 // psel p0.b, p8.b/Z, p10.b[w12, #1]\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe0160341 // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n" + ".inst 0x25646141 // psel p1.b, p8.b/Z, p10.b[w12, #4]\n" + ".inst 0x256c6140 // psel p0.b, p8.b/Z, p10.b[w12, #5]\n" + "ldr x26, [x10, #0x0]\n" + ".inst 0xe0160704 // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n" + "ldr x24, [x11, #0x8]\n" + "add x11, x11, #0x10\n" + ".inst 0xe01602e5 // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n" + "add x12, x12, #0x8\n" + "cmp x12, x15, LSL #2\n" + "ldr x23, [x10, #0x8]\n" + "add x10, x10, #0x10\n" + "blt 1b\n" + "2:" // K loop: Charge: End + ".inst 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n" + ".inst 0xe0160120 // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n" + ".inst 0x252c6140 // psel p0.b, p8.b/Z, p10.b[w12, #1]\n" + ".inst 0xe0160341 // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n" + ".inst 0x25646141 // psel p1.b, p8.b/Z, p10.b[w12, #4]\n" + ".inst 0x256c6140 // psel p0.b, p8.b/Z, p10.b[w12, #5]\n" + ".inst 0xe0160704 // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n" + "mov x11, %x[in]\n" + "add x10, %x[in], x16, LSL #3\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe01602e5 // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n" + "ldr x26, [x10, #0x0]\n" + "incb x22\n" + "incb x14\n" + "ldr x24, [x11, #0x8]\n" + "add x11, x11, #0x10\n" + "ldr x23, [x10, #0x8]\n" + "add x10, x10, #0x10\n" + "cbz x20, 8f\n" + "mov x20, x20\n" + "3:" // K loop: Main loop + "whilelt p8.b, x14, %x[width]\n" + "mov x13, #0x0\n" + "mov x12, #0x0\n" + "cbz x15, 5f\n" + "4:" // K loop: Main loop: First: Loop + ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" + ".inst 0xe0162122 // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n" + ".inst 0x253d6140 // psel p0.b, p8.b/Z, p10.b[w13, #3]\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe0162343 // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n" + ".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n" + ".inst 0x257d6142 // psel p2.b, p8.b/Z, p10.b[w13, #7]\n" + "ldr x26, [x10, #0x0]\n" + ".inst 0xe0162306 // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" + "ldr x24, [x11, #0x8]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0162ae7 // ld1b { za0h.b[x13, #7] }, p2/Z, [x23, x22]\n" + "ldr x23, [x10, #0x8]\n" + ".inst 0xe0bf86a0 // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25706d21 // psel p1.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0xe0b082a4 // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n" + ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.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, #0x8\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 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" + ".inst 0xe0162122 // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n" + ".inst 0x253d6140 // psel p0.b, p8.b/Z, p10.b[w13, #3]\n" + ".inst 0xe0162343 // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n" + ".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n" + "mov x11, %x[in]\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe0162306 // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n" + "add x10, %x[in], x16, LSL #3\n" + ".inst 0x257d6141 // psel p1.b, p8.b/Z, p10.b[w13, #7]\n" + "ldr x26, [x10, #0x0]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe01626e7 // ld1b { za0h.b[x13, #7] }, p1/Z, [x23, x22]\n" + "ldr x24, [x11, #0x8]\n" + ".inst 0x25306d22 // psel p2.s, p11.s/Z, p9.s[w12]\n" + "ldr x23, [x10, #0x8]\n" + ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25706d21 // psel p1.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0xe0b08aa4 // st1w { za1v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n" + "whilelt p9.b, x14, %x[width]\n" + "incb 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" + "incb x22\n" + "whilelt p8.b, x14, %x[width]\n" + "mov x13, #0x0\n" + "mov x12, #0x0\n" + "cbz x15, 7f\n" + "6:" // K loop: Main loop: Second: Loop + ".inst 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n" + ".inst 0xe0162120 // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n" + ".inst 0x252d6140 // psel p0.b, p8.b/Z, p10.b[w13, #1]\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe0162341 // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n" + ".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n" + ".inst 0x256d6142 // psel p2.b, p8.b/Z, p10.b[w13, #5]\n" + "ldr x26, [x10, #0x0]\n" + ".inst 0xe0162304 // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" + "ldr x24, [x11, #0x8]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0162ae5 // ld1b { za0h.b[x13, #5] }, p2/Z, [x23, x22]\n" + "ldr x23, [x10, #0x8]\n" + ".inst 0xe0bf86a8 // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25706d21 // psel p1.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0xe0b082ac // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n" + ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.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, #0x8\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 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n" + ".inst 0xe0162120 // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n" + ".inst 0x252d6140 // psel p0.b, p8.b/Z, p10.b[w13, #1]\n" + ".inst 0xe0162341 // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n" + ".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n" + "mov x11, %x[in]\n" + "ldr x9, [x11, #0x0]\n" + ".inst 0xe0162304 // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n" + "add x10, %x[in], x16, LSL #3\n" + ".inst 0x256d6141 // psel p1.b, p8.b/Z, p10.b[w13, #5]\n" + "ldr x26, [x10, #0x0]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe01626e5 // ld1b { za0h.b[x13, #5] }, p1/Z, [x23, x22]\n" + "ldr x24, [x11, #0x8]\n" + ".inst 0x25306d22 // psel p2.s, p11.s/Z, p9.s[w12]\n" + "ldr x23, [x10, #0x8]\n" + ".inst 0xe0bf82a8 // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25706d21 // psel p1.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0xe0b08aac // st1w { za3v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n" + "whilelt p9.b, 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" + "incb x14\n" + "incb x22\n" + "bgt 3b\n" + "8:" // K loop: Tails + "cbnz x25, 11f\n" + "mov x11, %x[in]\n" + "whilelt p8.b, x14, %x[width]\n" + "mov x13, #0x0\n" + "mov x12, #0x0\n" + "9:" // K loop: Tails: Even: First + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0b082a4 // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n" + "ldr x20, [x11, #0x0]\n" + ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" + ".inst 0xe0162282 // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n" + "ldr x20, [x11, x16, LSL #0x3]\n" + "add x12, x12, #0x1\n" + ".inst 0x253d6140 // psel p0.b, p8.b/Z, p10.b[w13, #3]\n" + ".inst 0xe0162283 // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x22]\n" + "cmp x12, x16\n" + "add x11, x11, #0x8\n" + "addvl x21, x21, #2\n" + "add x13, x13, #0x4\n" + "blt 9b\n" + "whilelt p9.b, x14, %x[width]\n" + "whilelt p8.b, x14, %x[width]\n" + "mov x20, #0x0\n" + "mov x12, #0x0\n" + "10:" // K loop: Tails: Even: Second + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0bf82a8 // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.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, #0x4\n" + "blt 10b\n" + "whilelt p8.b, x14, %x[width]\n" + "b 13f\n" + "11:" // K loop: Tails: Odd + "mov x12, #0x0\n" + "12:" // K loop: Tails: Odd: Loop + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.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_COMPUTE_ENABLE_SME) |