Implement dynamic fusion softmax operator
- Return aux tensorInfo by get_aux_tensors() at runtime to init the aux
tensor with the right size.
- Keep softmax unfusable for this commit
- Hence, added Tensor3D to template writer arguments declaration, for sake of
keeping dynamic fusion softmax componenets' kernels matching their cl
counterparts.
Resolves: COMPMID-5523
Change-Id: I667f39545db925f667036ef448302c79a0330373
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/483924
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8986
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>
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h
index 235c860..b15de71 100644
--- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h
@@ -145,10 +145,11 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
- tensor->allocator()->allocate();
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h
index e0aecf5..d9ce4df 100644
--- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h
@@ -133,9 +133,10 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
@@ -273,10 +274,11 @@
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
- tensor->allocator()->allocate();
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
TensorType t_input{};
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h
index e2722a1..faed610 100644
--- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h
@@ -131,10 +131,11 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- TensorType *tensor = data.first;
- AuxMemoryInfo aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
- tensor->allocator()->allocate();
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h
index efb67f8..efb5cf1 100644
--- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h
@@ -109,9 +109,10 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
@@ -142,8 +143,8 @@
return reference::pooling_layer<T>(src, pool_info, QuantizationInfo(), nullptr, DataLayout::NCHW);
}
- TensorType _target{};
- SimpleTensor<T> _reference{};
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
};
template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
diff --git a/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h
index bd99902..cd39ec0 100644
--- a/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h
@@ -132,10 +132,11 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
- tensor->allocator()->allocate();
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h
index 0d3b1f0..e0b62d0 100644
--- a/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h
@@ -90,9 +90,10 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
tensor->allocator()->allocate(); // Use ACL allocated memory
}
diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h
index 7eb820e..581a3e8 100644
--- a/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h
+++ b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h
@@ -158,10 +158,11 @@
// (Important) Allocate auxiliary tensor memory if there are any
for(auto &data : runtime.get_auxiliary_tensors())
{
- auto tensor = data.first;
- const auto aux_mem_req = data.second;
- tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
- tensor->allocator()->allocate();
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
}
// Construct user tensors
diff --git a/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h
new file mode 100644
index 0000000..3817711
--- /dev/null
+++ b/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h
@@ -0,0 +1,161 @@
+/*
+* Copyright (c) 2023 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.
+ */
+#ifndef TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE
+#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE
+
+#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
+
+#include "tests/SimpleTensor.h"
+#include "tests/framework/Fixture.h"
+#include "tests/framework/Macros.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/reference/SoftmaxLayer.h"
+
+using namespace arm_compute::experimental::dynamic_fusion;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionSoftmaxValidationGenericFixture : public framework::Fixture
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, float beta, size_t axis, bool is_log)
+ {
+ _reference = compute_reference(shape, data_type, beta, axis, is_log);
+ _target = compute_target(shape, data_type, beta, axis, is_log);
+ }
+
+protected:
+ template <typename U>
+ void fill(U &&tensor)
+ {
+ if(tensor.data_type() == DataType::F32)
+ {
+ std::uniform_real_distribution<float> distribution(-10.0f, 10.0f);
+ library->fill(tensor, distribution, 0);
+ }
+ else if(tensor.data_type() == DataType::F16)
+ {
+ arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -10.0f, 10.0f };
+ library->fill(tensor, distribution, 0);
+ }
+ else if(!is_data_type_quantized(tensor.data_type()))
+ {
+ std::uniform_int_distribution<> distribution(0, 100);
+ library->fill(tensor, distribution, 0);
+ }
+ else
+ {
+ library->fill_tensor_uniform(tensor, 0);
+ }
+ }
+
+ TensorType compute_target(const TensorShape &shape, DataType data_type, float beta, int32_t axis, bool is_log)
+ {
+ // Create a new workload sketch
+ CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+ GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &gpu_ctx };
+
+ SoftmaxAttributes softmax_attr{};
+ softmax_attr.axis(axis).beta(beta).is_log_softmax(is_log);
+ TensorInfo src_info = sketch.create_tensor_info(shape, 1, data_type);
+ TensorInfo dst_info = sketch.create_tensor_info(shape, 1, data_type);
+ FunctionType::create_op(sketch, &src_info, &dst_info, softmax_attr);
+
+ // Configure runtime
+ ClWorkloadRuntime runtime;
+ runtime.configure(sketch);
+
+ // (Important) Allocate auxiliary tensor memory if there are any
+ // Instead of using ACL allocated memory, the user can choose to import memory into the tensors
+ for(auto &data : runtime.get_auxiliary_tensors())
+ {
+ CLTensor *tensor = std::get<0>(data);
+ TensorInfo info = std::get<1>(data);
+ AuxMemoryInfo aux_mem_req = std::get<2>(data);
+ tensor->allocator()->init(info, aux_mem_req.alignment);
+ tensor->allocator()->allocate(); // Use ACL allocated memory
+ }
+ // Construct user tensors
+ TensorType src{};
+ TensorType dst{};
+
+ // Initialize user tensors
+ src.allocator()->init(src_info);
+ dst.allocator()->init(dst_info);
+
+ // Allocate and fill user tensors
+ src.allocator()->allocate();
+ dst.allocator()->allocate();
+ fill(AccessorType(src));
+
+ // Run runtime
+ runtime.run({ &src, &dst });
+
+ return dst;
+ }
+
+ SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, float beta, int32_t axis, bool is_log)
+ {
+ // Create reference
+ SimpleTensor<T> src{ shape, data_type, 1 };
+
+ // Fill reference
+ fill(src);
+
+ return reference::softmax_layer<T>(src, beta, axis, is_log);
+ }
+
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionSoftmaxValidationFixture : public DynamicFusionSoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, float beta, size_t axis, bool is_log)
+ {
+ DynamicFusionSoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape,
+ data_type,
+ beta,
+ axis,
+ is_log);
+ }
+};
+
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+
+#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE */