diff options
| author | Gunes Bayir <gunes.bayir@arm.com> | 2023-11-07 05:43:07 +0000 |
|---|---|---|
| committer | Gunes Bayir <gunes.bayir@arm.com> | 2023-12-05 13:52:17 +0000 |
| commit | fadc9b1e0bba90d6a91beb65466b2a0895b3a5e4 (patch) | |
| tree | 7d095fefe3634b4ca86dc9088bb2990d64d3a7c8 /src/cpu/kernels/softmax/generic/sve/impl.cpp | |
| parent | 23158b0a69b85c9c6e5a7f2457bfe10be04d6132 (diff) | |
| download | ComputeLibrary-fadc9b1e0bba90d6a91beb65466b2a0895b3a5e4.tar.gz | |
Optimize CpuSoftmaxKernel for axis=0
Implement a single kernel instead of having two consecutive ones. In the previous setup, one kernel was calculating the maximum value in the axis, and this maximum was being subtracted from each data while calculating the softmax, i.e.
softmax(x_i) = exp(x_i - max) / sum_i( exp(x_i - max) )
This patch integrates these two stages into a single kernel for Neon™ for all data types. This will save some memory because we don't need to hold the max values in a separate auxiliary tensor.
It also introduces some other optimizations that will ease memory pressure when the data type is float/half, by using the dst tensor as temporary storage for already exponentiated inputs.
It removes the references to SVE and SVE2 implementations, and most of the associated files; but, it leaves the implementations as these may be used in the future.
Resolves: COMPMID-6500
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: Icff9976d1214c4c6cbe15a62ca60b8a77d3784cc
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10688
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/kernels/softmax/generic/sve/impl.cpp')
| -rw-r--r-- | src/cpu/kernels/softmax/generic/sve/impl.cpp | 25 |
1 files changed, 4 insertions, 21 deletions
diff --git a/src/cpu/kernels/softmax/generic/sve/impl.cpp b/src/cpu/kernels/softmax/generic/sve/impl.cpp index 24f1bb8143..0d4b7f4509 100644 --- a/src/cpu/kernels/softmax/generic/sve/impl.cpp +++ b/src/cpu/kernels/softmax/generic/sve/impl.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -30,6 +30,9 @@ namespace arm_compute { namespace cpu { +/// TODO: (COMPMID-6505) Similar to Neon(TM), this implementation be converted to +/// a single kernel that performs softmax operation. Leaving the SVE code here for +/// future references. Implementation for Neon(TM) is introduced in COMPMID-6500 template <typename ScalarType> void sve_logits_1d_max(const ITensor *in, ITensor *out, const Window &window) { @@ -172,25 +175,5 @@ void sve_softmax_logits_1d_float(const ITensor *in, }, in_it, max_it, out_it); } - -template void sve_logits_1d_max<float>(const ITensor *in, ITensor *out, const Window &window); -template void sve_logits_1d_max<float16_t>(const ITensor *in, ITensor *out, const Window &window); -template void sve_logits_1d_max<qasymm8_t>(const ITensor *in, ITensor *out, const Window &window); -template void sve_logits_1d_max<qasymm8_signed_t>(const ITensor *in, ITensor *out, const Window &window); - -template void sve_softmax_logits_1d_float<float>(const ITensor *in, - const ITensor *max, - void *const tmp, - ITensor *out, - const float beta, - bool is_log, - const Window &window); -template void sve_softmax_logits_1d_float<float16_t>(const ITensor *in, - const ITensor *max, - void *const tmp, - ITensor *out, - const float beta, - bool is_log, - const Window &window); } // namespace cpu } // namespace arm_compute |