commit 06ea048f062a50404b1b3998a61a45449c2d1f0f
author Anthony Barbier <anthony.barbier@arm.com> Thu Feb 22 15:45:35 2018 +0000
committer Anthony Barbier <anthony.barbier@arm.com> Fri Feb 23 11:49:54 2018 +0000
tree aa0dea3b0c49422538df9a5a02578b2c29e6fa67
parent 292227986edb37b01061afcad6df18ba9d6ccbeb

arm_compute v18.02

Change-Id: I7207aa488e5470f235f39b6c188b4678dc38d1a6

src/runtime/NEON/functions/NESoftmaxLayer.cpp

diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index 8e6773c..4fb8300 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,7 +32,7 @@
 using namespace arm_compute;
 
 NESoftmaxLayer::NESoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _max_kernel(), _shift_exp_sum_kernel(), _norm_kernel(), _fill_border_kernel(), _max(), _sum(), _tmp()
+    : _memory_group(std::move(memory_manager)), _max_kernel(), _softmax_kernel(), _fill_border_kernel(), _max(), _tmp()
 {
 }
 
@@ -40,31 +40,22 @@
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
-    // Create intermediate tensors shapes
-    TensorInfo tensor_info_tmp(input->info()->tensor_shape(), input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position());
-    _tmp.allocator()->init(tensor_info_tmp);
-
-    TensorShape shape = input->info()->tensor_shape();
-    shape.set(0, 1);
-    TensorInfo tensor_info_max_sum(shape, input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position());
-    _max.allocator()->init(tensor_info_max_sum);
-    _sum.allocator()->init(tensor_info_max_sum);
-
-    // Manage intermediate buffers
-    _memory_group.manage(&_tmp);
-    _memory_group.manage(&_max);
-    _memory_group.manage(&_sum);
-
     // Configure Kernels
     _max_kernel.configure(input, &_max);
-    _shift_exp_sum_kernel.configure(input, &_max, &_tmp, &_sum, beta);
-    _norm_kernel.configure(&_tmp, &_sum, output);
     _fill_border_kernel.configure(input, _max_kernel.border_size(), BorderMode::REPLICATE);
+    _softmax_kernel.configure(input, &_max, output, beta, &_tmp);
+
+    // Init intermediate tensors
+    _max.allocator()->init(*_max.info());
+    _tmp.allocator()->init(*_tmp.info());
+
+    // Manage intermediate buffers
+    _memory_group.manage(&_max);
+    _memory_group.manage(&_tmp);
 
     // Allocate intermediate tensors
-    _tmp.allocator()->allocate();
     _max.allocator()->allocate();
-    _sum.allocator()->allocate();
+    _tmp.allocator()->allocate();
 }
 
 Status NESoftmaxLayer::validate(const ITensorInfo *input, const ITensorInfo *output, float beta)
@@ -72,14 +63,12 @@
     // Perform validation step
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
 
-    TensorShape max_sum_shape = input->tensor_shape();
-    max_sum_shape.set(0, 1);
-
-    TensorInfo tensor_info_max_sum(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(max_sum_shape));
+    const TensorShape max_shape           = TensorShape(input->tensor_shape()).set(0, 1);
+    const TensorInfo  tensor_info_max_sum = TensorInfo(*input).set_tensor_shape(max_shape).reset_padding();
+    const TensorInfo  dont_care;
 
     ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DMaxKernel::validate(input, &tensor_info_max_sum));
-    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DShiftExpSumKernel::validate(input, &tensor_info_max_sum, input, &tensor_info_max_sum, beta));
-    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DNormKernel::validate(input, &tensor_info_max_sum, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DSoftmaxKernel::validate(input, &tensor_info_max_sum, output, beta, &dont_care));
 
     return Status{};
 }
@@ -90,8 +79,7 @@
 
     NEScheduler::get().schedule(&_fill_border_kernel, Window::DimY);
     NEScheduler::get().schedule(&_max_kernel, Window::DimY);
-    NEScheduler::get().schedule(&_shift_exp_sum_kernel, Window::DimY);
-    NEScheduler::get().schedule(&_norm_kernel, Window::DimY);
+    NEScheduler::get().schedule(&_softmax_kernel, Window::DimY);
 
     _memory_group.release();
 }