diff options
author | David Mansell <David.Mansell@arm.com> | 2023-08-23 16:00:13 +0100 |
---|---|---|
committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2023-08-29 10:14:10 +0000 |
commit | b7aefd71d07d56b001e795410700cae71a518eca (patch) | |
tree | d913041260c04ad56872a4fb45d39c28a18a9528 /src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp | |
parent | cea7060684ae6c33fc8e16affc1c7998d17815ae (diff) | |
download | ComputeLibrary-b7aefd71d07d56b001e795410700cae71a518eca.tar.gz |
GEMM: AArch32: Split assembler block in a32_merge_float_8x6.hpp
Inline assembler blocks attempting to bind 8 integer
registers don't compile in certain configurations (notably GCC 13.2 debug
builds with -O0 -g). Fix this by splitting the offending block into two
separate parts (straightforward as there is no flow control in the block).
Fixes: COMPMID-6532
Signed-off-by: David Mansell <David.Mansell@arm.com>
Change-Id: I80e9a10e6a91574176d50e63c45fab055aefa659
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10197
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Emanuele Rocca <ema@linux.it>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp | 10 |
1 files changed, 8 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp b/src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp index bea455ca67..989bb17dfb 100644 --- a/src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp +++ b/src/core/NEON/kernels/arm_gemm/merges/a32_merge_float_8x6.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 Arm Limited. + * Copyright (c) 2017-2018,2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -274,7 +274,13 @@ void MergeResults<8, 6, false>(float *out, const float *in, const int ldout, con "VMIN.f32 q6, q6, %q[maxv]\n" "VMIN.f32 q7, q7, %q[maxv]\n" "VST1.32 {d12-d15}, [%[outptr3]]!\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), + [inptr] "+r" (inptr) + : [minv] "w" (minv), [maxv] "w" (maxv), [biasptr] "r" (biasptr) + : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "memory" + ); + __asm __volatile ( // Rows 4-5 "VLD1.32 {d8-d11}, [%[inptr]]!\n" "VLD1.32 {d12-d15}, [%[inptr]]!\n" @@ -296,7 +302,7 @@ void MergeResults<8, 6, false>(float *out, const float *in, const int ldout, con "VMIN.f32 q6, q6, %q[maxv]\n" "VMIN.f32 q7, q7, %q[maxv]\n" "VST1.32 {d12-d15}, [%[outptr5]]!\n" - : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), + : [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [inptr] "+r" (inptr) : [minv] "w" (minv), [maxv] "w" (maxv), [biasptr] "r" (biasptr) : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "memory" |