diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp | 200 |
1 files changed, 200 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp new file mode 100644 index 0000000000..2756327815 --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp @@ -0,0 +1,200 @@ +/* + * 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. + */ + +#pragma once + +#if defined(ARM_COMPUTE_ENABLE_SME) + +namespace { + +void sme_transpose_interleave_16VL_2x2_fp32bf16(bfloat16 *out, const float *in, size_t width, size_t in_stride, size_t height) +{ + float *pad_row = reinterpret_cast<float *>(alloca(width * sizeof(float))); + + if (height % 2) { + memset(pad_row, 0, width * sizeof(float)); + } + + size_t out_stride = 16 * roundup<size_t>(height, 2) * sme::get_vector_length<uint16_t>(); + + __asm__ __volatile__( + ".inst 0xd503477f // SMSTART ZA\n" + "ptrue p7.b\n" + "1:" // Main row loop: Head + "mov x25, %x[in]\n" + "add x24, x25, %x[in_stride]\n" + "cmp %x[height], #0x1\n" + "add %x[in], x24, %x[in_stride]\n" + "mov x23, %x[out]\n" + "csel x24, x24, %x[pad_row], GT\n" + "sub %x[height], %x[height], #0x2\n" + "mov x22, %x[width]\n" + "2:" // Main row loop: Column loop + "mov x21, x22\n" + "whilelt p1.s, XZR, x21\n" + "ld1w { z16.s }, p1/Z, [x25]\n" + ".inst 0x658abe00 // bfcvt z0.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p0.s, XZR, x21\n" + "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n" + ".inst 0x658abe1f // bfcvt z31.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p6.s, XZR, x21\n" + "ld1w { z16.s }, p6/Z, [x25, #2, MUL VL]\n" + ".inst 0x658abe1e // bfcvt z30.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p5.s, XZR, x21\n" + "ld1w { z16.s }, p5/Z, [x25, #3, MUL VL]\n" + ".inst 0x658abe1d // bfcvt z29.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p4.s, XZR, x21\n" + "ld1w { z16.s }, p4/Z, [x25, #4, MUL VL]\n" + ".inst 0x658abe1c // bfcvt z28.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p3.s, XZR, x21\n" + "ld1w { z16.s }, p3/Z, [x25, #5, MUL VL]\n" + ".inst 0x658abe1b // bfcvt z27.h, p7/M, z16.s\n" + "decw x21\n" + "whilelt p2.s, XZR, x21\n" + "ld1w { z16.s }, p2/Z, [x25, #6, MUL VL]\n" + ".inst 0x658abe1a // bfcvt z26.h, p7/M, z16.s\n" + "decw x21\n" + "ld1w { z16.s }, p1/Z, [x24]\n" + "whilelt p1.s, XZR, x21\n" + ".inst 0x648abe00 // bfcvtnt z0.h, p7/M, z16.s\n" + "decw x21\n" + "ld1w { z16.s }, p1/Z, [x25, #7, MUL VL]\n" + "addvl x25, x25, #16\n" + ".inst 0x658abe19 // bfcvt z25.h, p7/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n" + "whilelt p0.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1f // bfcvtnt z31.h, p7/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x25, #-8, MUL VL]\n" + ".inst 0x658abe18 // bfcvt z24.h, p7/M, z16.s\n" + "mov x20, x23\n" + "decw x22, ALL, MUL #16\n" + "ld1w { z16.s }, p6/Z, [x24, #2, MUL VL]\n" + "whilelt p6.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1e // bfcvtnt z30.h, p7/M, z16.s\n" + "ld1w { z16.s }, p6/Z, [x25, #-7, MUL VL]\n" + ".inst 0x658abe17 // bfcvt z23.h, p7/M, z16.s\n" + "add x23, x23, %x[out_stride]\n" + "ld1w { z16.s }, p5/Z, [x24, #3, MUL VL]\n" + "whilelt p5.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1d // bfcvtnt z29.h, p7/M, z16.s\n" + "ld1w { z16.s }, p5/Z, [x25, #-6, MUL VL]\n" + ".inst 0x658abe16 // bfcvt z22.h, p7/M, z16.s\n" + "ld1w { z16.s }, p4/Z, [x24, #4, MUL VL]\n" + "whilelt p4.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1c // bfcvtnt z28.h, p7/M, z16.s\n" + "ld1w { z16.s }, p4/Z, [x25, #-5, MUL VL]\n" + ".inst 0x658abe15 // bfcvt z21.h, p7/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x24, #5, MUL VL]\n" + "whilelt p3.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1b // bfcvtnt z27.h, p7/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x25, #-4, MUL VL]\n" + ".inst 0x658abe14 // bfcvt z20.h, p7/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x24, #6, MUL VL]\n" + "whilelt p2.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe1a // bfcvtnt z26.h, p7/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #-3, MUL VL]\n" + ".inst 0x658abe13 // bfcvt z19.h, p7/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x24, #7, MUL VL]\n" + "whilelt p1.s, XZR, x21\n" + "decw x21\n" + ".inst 0x648abe19 // bfcvtnt z25.h, p7/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x25, #-2, MUL VL]\n" + "addvl x24, x24, #16\n" + ".inst 0x658abe12 // bfcvt z18.h, p7/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x24, #-8, MUL VL]\n" + "whilelt p0.s, XZR, x21\n" + "cmp x22, #0x0\n" + ".inst 0x648abe18 // bfcvtnt z24.h, p7/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x25, #-1, MUL VL]\n" + ".inst 0x658abe11 // bfcvt z17.h, p7/M, z16.s\n" + "ld1w { z16.s }, p6/Z, [x24, #-7, MUL VL]\n" + ".inst 0x648abe17 // bfcvtnt z23.h, p7/M, z16.s\n" + "ld1w { z16.s }, p5/Z, [x24, #-6, MUL VL]\n" + ".inst 0x648abe16 // bfcvtnt z22.h, p7/M, z16.s\n" + "ld1w { z16.s }, p4/Z, [x24, #-5, MUL VL]\n" + ".inst 0x648abe15 // bfcvtnt z21.h, p7/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x24, #-4, MUL VL]\n" + ".inst 0x648abe14 // bfcvtnt z20.h, p7/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x24, #-3, MUL VL]\n" + ".inst 0x648abe13 // bfcvtnt z19.h, p7/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x24, #-2, MUL VL]\n" + ".inst 0x648abe12 // bfcvtnt z18.h, p7/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x24, #-1, MUL VL]\n" + "st1h { z0.h }, p7, [x20]\n" + ".inst 0x648abe11 // bfcvtnt z17.h, p7/M, z16.s\n" + "st1h { z31.h }, p7, [x20, #1, MUL VL]\n" + "st1h { z30.h }, p7, [x20, #2, MUL VL]\n" + "st1h { z29.h }, p7, [x20, #3, MUL VL]\n" + "st1h { z28.h }, p7, [x20, #4, MUL VL]\n" + "st1h { z27.h }, p7, [x20, #5, MUL VL]\n" + "st1h { z26.h }, p7, [x20, #6, MUL VL]\n" + "st1h { z25.h }, p7, [x20, #7, MUL VL]\n" + "addvl x20, x20, #16\n" + "st1h { z24.h }, p7, [x20, #-8, MUL VL]\n" + "st1h { z23.h }, p7, [x20, #-7, MUL VL]\n" + "st1h { z22.h }, p7, [x20, #-6, MUL VL]\n" + "st1h { z21.h }, p7, [x20, #-5, MUL VL]\n" + "st1h { z20.h }, p7, [x20, #-4, MUL VL]\n" + "st1h { z19.h }, p7, [x20, #-3, MUL VL]\n" + "st1h { z18.h }, p7, [x20, #-2, MUL VL]\n" + "st1h { z17.h }, p7, [x20, #-1, MUL VL]\n" + "bgt 2b\n" + "3:" // Main row loop: Column loop skip + "cmp %x[height], #0x1\n" + "addvl %x[out], %x[out], #16\n" + "bge 1b\n" + ".inst 0xd503467f // SMSTOP\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", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "x25", "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<16, 2, true, VLType::SME>( + bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax) +{ + sme_transpose_interleave_16VL_2x2_fp32bf16( + out, + in + k0 * stride + x0, + (xmax-x0), + stride * sizeof(float), + (kmax-k0) + ); +} + + +#endif // defined(ARM_COMPUTE_ENABLE_SME) |