Add multi-sketch support for dynamic fusion
* Tensors are owned by workload context instead of workload sketch
so that they can be used by multiple sketches.
* Add an integration test for multi-sketch case.
Resolves: COMPMID-6148
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I37d0de5ac103fb2a85020aa1c26e49eb304f47b7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9706
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Add.cpp b/tests/validation/dynamic_fusion/gpu/cl/Add.cpp
index 0034b0f..d9a3d95 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Add.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Add.cpp
@@ -87,12 +87,12 @@
{
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Validate Elementwise Add
- auto lhs_info = sketch.create_tensor_info(input1_info);
- auto rhs_info = sketch.create_tensor_info(input2_info);
+ auto lhs_info = context.create_tensor_info(input1_info);
+ auto rhs_info = context.create_tensor_info(input2_info);
bool res = bool(GpuAdd::validate_op(sketch, &lhs_info, &rhs_info));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp
index 177c02c..dc46dd5 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp
@@ -69,11 +69,11 @@
{
// Create a new workload sketch
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Fuse Clamp
- const TensorInfo src_info = sketch.create_tensor_info(input_info);
+ const TensorInfo src_info = context.create_tensor_info(input_info);
ClampAttributes attributes {};
attributes.min_val(min_val)
diff --git a/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp
index b6331d7..7ab7c8a 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp
@@ -242,12 +242,12 @@
input_info, weights_info, biases_info, padding, stride, depth_multiplier, dilation, expected)
{
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
- const TensorInfo sketch_weights_info = sketch.create_tensor_info(weights_info);
- const TensorInfo sketch_biases_info = sketch.create_tensor_info(biases_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
+ const TensorInfo sketch_weights_info = context.create_tensor_info(weights_info);
+ const TensorInfo sketch_biases_info = context.create_tensor_info(biases_info);
DepthwiseConv2dAttributes attributes {};
attributes.pad(padding)
diff --git a/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp
index cccad18..f27a179 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp
@@ -157,12 +157,12 @@
input_info, weights_info, biases_info, conv2d_attrs, expected)
{
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
- const TensorInfo sketch_weights_info = sketch.create_tensor_info(weights_info);
- const TensorInfo sketch_biases_info = sketch.create_tensor_info(biases_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
+ const TensorInfo sketch_weights_info = context.create_tensor_info(weights_info);
+ const TensorInfo sketch_biases_info = context.create_tensor_info(biases_info);
bool is_valid = bool(GpuConv2d::validate_op(sketch, &sketch_input_info, &sketch_weights_info, &sketch_biases_info, conv2d_attrs));
ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
}
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp b/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp
index a9e8f9c..2da2b9e 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp
@@ -102,12 +102,12 @@
{
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Validate Elementwise Mul
- auto lhs_info = sketch.create_tensor_info(input1_info);
- auto rhs_info = sketch.create_tensor_info(input2_info);
+ auto lhs_info = context.create_tensor_info(input1_info);
+ auto rhs_info = context.create_tensor_info(input2_info);
bool res = bool(GpuMul::validate_op(sketch, &lhs_info, &rhs_info));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp
index a7772ae..f4478db 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp
@@ -101,15 +101,15 @@
{
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Declare GpuPool2d settings
const GpuPool2dSettings &settings = GpuPool2dSettings().mixed_precision(false);
// Validate Pool2d Configuration
- auto src_info = sketch.create_tensor_info(input_info);
- auto dst_info = sketch.create_tensor_info(output_info);
+ auto src_info = context.create_tensor_info(input_info);
+ auto dst_info = context.create_tensor_info(output_info);
bool res = bool(GpuPool2d::validate_op(sketch, &src_info, &dst_info, pool2d_attr, settings));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
}
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp b/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp
index 6d88be4..bdaa1be 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp
@@ -53,13 +53,13 @@
{
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Create sketch tensors
TensorShape input_shape = input_info.tensor_shape();
ARM_COMPUTE_UNUSED(input_shape);
- TensorInfo src_info = sketch.create_tensor_info(input_info);
+ TensorInfo src_info = context.create_tensor_info(input_info);
ReshapeAttributes attributes;
attributes.shape(output_shape);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp
index 696be54..5f99cd6 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp
@@ -95,10 +95,10 @@
const TensorInfo output_info = TensorInfo{ default_output_shape, 1, default_data_type, default_data_layout };
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
// nullptr is given as input
Status status = GpuResize::validate_op(sketch, nullptr, ResizeAttributes());
@@ -135,10 +135,10 @@
const TensorInfo input_info = TensorInfo{ default_input_shape, 1, kv.first, default_data_layout };
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
ResizeAttributes attributes;
attributes.output_width(default_output_shape[0]); // shape is not important unless it's empty
@@ -157,10 +157,10 @@
const TensorInfo output_info = TensorInfo{ default_output_shape, 1, non_default_data_type, default_data_layout };
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
Status status = GpuResize::validate_op(sketch, &sketch_input_info, ResizeAttributes());
ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
@@ -177,10 +177,10 @@
const TensorInfo output_info = TensorInfo{ default_output_shape, 1, default_data_type, default_data_layout };
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
ResizeAttributes attributes{};
attributes.interpolation_policy(interpolation_policy)
@@ -198,10 +198,10 @@
constexpr auto interpolation_policy = InterpolationPolicy::AREA;
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
ResizeAttributes attributes{};
attributes.interpolation_policy(interpolation_policy);
@@ -217,10 +217,10 @@
constexpr auto interpolation_policy = InterpolationPolicy::BILINEAR;
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
- const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info);
+ const TensorInfo sketch_input_info = context.create_tensor_info(input_info);
ResizeAttributes attributes{};
attributes.interpolation_policy(interpolation_policy);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp
index aace23e..5fd1180 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp
@@ -61,11 +61,11 @@
{
// Create a new workload sketch
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Fuse sigmoid
- const TensorInfo src_info = sketch.create_tensor_info(input_info);
+ const TensorInfo src_info = context.create_tensor_info(input_info);
const bool res = static_cast<bool>(GpuSigmoid::validate_op(sketch, &src_info));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp b/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp
index d09454e..e8314d7 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp
@@ -104,13 +104,13 @@
{
// 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 };
+ GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
SoftmaxAttributes softmax_attr{};
softmax_attr.axis(axis).beta(beta).is_log_softmax(false);
- TensorInfo src_info = sketch.create_tensor_info(input_info);
- TensorInfo dst_info = sketch.create_tensor_info(output_info);
+ TensorInfo src_info = context.create_tensor_info(input_info);
+ TensorInfo dst_info = context.create_tensor_info(output_info);
const bool res = static_cast<bool>(GpuSoftmax::validate_op(sketch, &src_info, &dst_info, softmax_attr));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
}
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp b/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp
index 977e011..0bb05c2 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp
@@ -89,12 +89,12 @@
{
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ auto context = GpuWorkloadContext{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Validate Elementwise Sub
- auto lhs_info = sketch.create_tensor_info(input1_info);
- auto rhs_info = sketch.create_tensor_info(input2_info);
+ auto lhs_info = context.create_tensor_info(input1_info);
+ auto rhs_info = context.create_tensor_info(input2_info);
bool res = bool(GpuSub::validate_op(sketch, &lhs_info, &rhs_info));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp
index 183cd07..00c92fb 100644
--- a/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp
+++ b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp
@@ -61,11 +61,11 @@
{
// Create a new workload sketch
CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
- GpuWorkloadContext gpu_ctx{ &cl_compile_ctx };
- GpuWorkloadSketch sketch{ &gpu_ctx };
+ GpuWorkloadContext context{ &cl_compile_ctx };
+ GpuWorkloadSketch sketch{ &context };
// Fuse tanh
- const TensorInfo src_info = sketch.create_tensor_info(input_info);
+ const TensorInfo src_info = context.create_tensor_info(input_info);
const bool res = static_cast<bool>(GpuTanh::validate_op(sketch, &src_info));
ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);