diff options
author | Renato Arantes <renato.arantes@arm.com> | 2023-11-23 11:12:51 +0000 |
---|---|---|
committer | Jakub Sujak <jakub.sujak@arm.com> | 2024-01-12 10:17:30 +0000 |
commit | 0eb9cfbe6e78f80008164cb0ee18afa09a7fe4eb (patch) | |
tree | 2dcfa82fdc905c91b9263e02815dc3394063ca18 /src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp | |
parent | c5df0c6c5d41a1c4c42ed9b9106d4a2c87689b38 (diff) | |
download | ComputeLibrary-0eb9cfbe6e78f80008164cb0ee18afa09a7fe4eb.tar.gz |
[ONCPUML-1387] Add ACL based reorder for f32 to bf16 data type conversion.
The reorders supported at the moment are:
ab->BA4b4a
ab->BA8b4a
Co-Authored-By: David Mansell <David.Mansell@arm.com>
Change-Id: Ic466465629ce3bcdcee0089e251485b79b60e1f3
Signed-off-by: Renato Arantes <renato.arantes@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10775
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp | 149 |
1 files changed, 149 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp new file mode 100644 index 0000000000..f66fcdc994 --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_2VL_2x4_fp32bf16.hpp @@ -0,0 +1,149 @@ +/* + * Copyright (c) 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_2VL_2x4_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 % 4) { + memset(pad_row, 0, width * sizeof(float)); + } + + size_t out_stride = 2 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>(); + + __asm__ __volatile__( + "ptrue p1.b\n" + "1:" // Main row loop: Head + "mov x26, %x[in]\n" + "mov x25, %x[width]\n" + "cnth x24\n" + "cmp %x[height], #0x3\n" + "mov x23, %x[out]\n" + "add x22, x26, %x[in_stride]\n" + "add x21, x22, %x[in_stride]\n" + "add x20, x21, %x[in_stride]\n" + "add %x[in], x20, %x[in_stride]\n" + "csel x20, x20, %x[pad_row], GT\n" + "csel x21, x21, %x[pad_row], GE\n" + "cmp %x[height], #0x1\n" + "csel x22, x22, %x[pad_row], GT\n" + "cmp x25, x24\n" + "sub %x[height], %x[height], #0x4\n" + "blt 3f\n" + "2:" // Main row loop: Unroll column loop + "ld1w { z18.s }, p1/Z, [x26]\n" + "ld1w { z17.s }, p1/Z, [x21]\n" + "sub x25, x25, x24\n" + "ld1w { z21.s }, p1/Z, [x26, #1, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x21, #1, MUL VL]\n" + "cmp x25, x24\n" + "addvl x26, x26, #2\n" + "ld1w { z26.s }, p1/Z, [x22]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" + "addvl x21, x21, #2\n" + "zip1 z19.s, z18.s, z17.s\n" + "zip2 z18.s, z18.s, z17.s\n" + "ld1w { z25.s }, p1/Z, [x22, #1, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x20, #1, MUL VL]\n" + "addvl x22, x22, #2\n" + "zip1 z17.s, z21.s, z16.s\n" + "zip2 z16.s, z21.s, z16.s\n" + "addvl x20, x20, #2\n" + ".inst 0x658aa677 // bfcvt z23.h, p1/M, z19.s\n" + "zip1 z22.s, z26.s, z20.s\n" + ".inst 0x658aa655 // bfcvt z21.h, p1/M, z18.s\n" + "zip2 z20.s, z26.s, z20.s\n" + ".inst 0x658aa633 // bfcvt z19.h, p1/M, z17.s\n" + "zip1 z18.s, z25.s, z24.s\n" + ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" + "zip2 z16.s, z25.s, z24.s\n" + ".inst 0x648aa6d7 // bfcvtnt z23.h, p1/M, z22.s\n" + ".inst 0x648aa695 // bfcvtnt z21.h, p1/M, z20.s\n" + ".inst 0x648aa653 // bfcvtnt z19.h, p1/M, z18.s\n" + ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" + "st1h { z23.h }, p1, [x23]\n" + "st1h { z21.h }, p1, [x23, #1, MUL VL]\n" + "add x23, x23, %x[out_stride]\n" + "st1h { z19.h }, p1, [x23]\n" + "st1h { z17.h }, p1, [x23, #1, MUL VL]\n" + "add x23, x23, %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 + "whilelt p0.s, XZR, x25\n" + "decd x25, ALL, MUL #2\n" + "ld1w { z19.s }, p0/Z, [x26]\n" + "addvl x26, x26, #1\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "addvl x21, x21, #1\n" + "ld1w { z20.s }, p0/Z, [x22]\n" + "addvl x22, x22, #1\n" + "ld1w { z18.s }, p0/Z, [x20]\n" + "addvl x20, x20, #1\n" + "cmp x25, #0x0\n" + "zip1 z17.s, z19.s, z16.s\n" + "zip2 z16.s, z19.s, z16.s\n" + "zip1 z19.s, z20.s, z18.s\n" + "zip2 z18.s, z20.s, z18.s\n" + ".inst 0x658aa631 // bfcvt z17.h, p1/M, z17.s\n" + ".inst 0x658aa610 // bfcvt z16.h, p1/M, z16.s\n" + ".inst 0x648aa671 // bfcvtnt z17.h, p1/M, z19.s\n" + ".inst 0x648aa650 // bfcvtnt z16.h, p1/M, z18.s\n" + "st1h { z17.h }, p1, [x23]\n" + "st1h { z16.h }, p1, [x23, #1, MUL VL]\n" + "add x23, x23, %x[out_stride]\n" + "bgt 4b\n" + "5:" // Main row loop: Column loop skip + "cmp %x[height], #0x1\n" + "addvl %x[out], %x[out], #2\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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26" + ); +} + +} // anonymous namespace +template<> +void Transform<2, 4, true, VLType::SVE>( + bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax) +{ + sve_transpose_interleave_2VL_2x4_fp32bf16( + out, + in + k0 * stride + x0, + (xmax-x0), + stride * sizeof(float), + (kmax-k0) + ); +} + + +#endif // defined(ARM_COMPUTE_ENABLE_SVE) |