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>

1 files changed, 10 insertions, 12 deletions

diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index e3c2012d05..be588c5b52 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,7 +29,6 @@
 
 #include "src/core/helpers/MemoryHelpers.h"
 #include "src/core/helpers/SoftmaxHelpers.h"
-#include "src/cpu/kernels/CpuSoftmaxKernel.h"
 #include "src/cpu/operators/CpuSoftmax.h"
 
 namespace arm_compute
@@ -37,13 +36,12 @@ namespace arm_compute
 template <bool IS_LOG>
 struct NESoftmaxLayerGeneric<IS_LOG>::Impl
 {
-    const ITensor                                  *src{nullptr};
-    ITensor                                        *dst{nullptr};
-    Tensor                                          max{nullptr};
-    std::unique_ptr<cpu::CpuSoftmaxGeneric<IS_LOG>> op{nullptr};
-    MemoryGroup                                     memory_group{};
-    ITensorPack                                     run_pack{};
-    WorkspaceData<Tensor>                           workspace_tensors{};
+    const ITensor                          *src{nullptr};
+    ITensor                                *dst{nullptr};
+    std::unique_ptr<cpu::CpuSoftmaxGeneric> op{nullptr};
+    MemoryGroup                             memory_group{};
+    ITensorPack                             run_pack{};
+    WorkspaceData<Tensor>                   workspace_tensors{};
 };
 
 template <bool IS_LOG>
@@ -67,8 +65,8 @@ void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, f
 
     _impl->src = input;
     _impl->dst = output;
-    _impl->op  = std::make_unique<cpu::CpuSoftmaxGeneric<IS_LOG>>();
-    _impl->op->configure(input->info(), output->info(), beta, axis);
+    _impl->op  = std::make_unique<cpu::CpuSoftmaxGeneric>();
+    _impl->op->configure(input->info(), output->info(), beta, axis, IS_LOG);
 
     _impl->run_pack          = {{TensorType::ACL_SRC, _impl->src}, {TensorType::ACL_DST, _impl->dst}};
     _impl->workspace_tensors = manage_workspace<Tensor>(_impl->op->workspace(), _impl->memory_group, _impl->run_pack);
@@ -79,7 +77,7 @@ Status
 NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, int32_t axis)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuSoftmaxGeneric<IS_LOG>::validate(input, output, beta, axis));
+    ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuSoftmaxGeneric::validate(input, output, beta, axis, IS_LOG));
     return Status{};
 }