Use heuristics for setting dynamic fusion direct conv2d tile sizes
Resolves: COMPMID-5735
Change-Id: I9958413b69c5052cfa205dd0e9457cc4953aaf35
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/474818
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8724
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
index dc05825..1fbcb41 100644
--- a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
+++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
@@ -57,13 +57,24 @@
return _fast_relaxed_math;
}
+ClComponentDirectConv2dSettings &ClComponentDirectConv2dSettings::direct_conv_descriptor(const DirectConvComputeKernelInfo &desc)
+{
+ _desc = desc;
+ return *this;
+}
+
+DirectConvComputeKernelInfo ClComponentDirectConv2dSettings::direct_conv_descriptor() const
+{
+ return _desc;
+}
+
Status ClComponentDirectConv2d::validate(
const Properties &properties,
const ArgumentPack<ITensorInfo> &tensors,
const Attributes &attributes,
const Settings &settings)
{
- ARM_COMPUTE_UNUSED(properties, settings);
+ ARM_COMPUTE_UNUSED(properties);
const auto src = tensors.get_const_tensor(TensorType::ACL_SRC_0);
const auto wei = tensors.get_const_tensor(TensorType::ACL_SRC_1);
const auto bia = tensors.get_const_tensor(TensorType::ACL_SRC_2);
@@ -125,6 +136,11 @@
// Data layout
ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC);
+ const auto desc = settings.direct_conv_descriptor();
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.n0 != 1 && desc.n0 != 2 && desc.n0 != 3 && desc.n0 != 4 && desc.n0 != 8 && desc.n0 != 16,
+ "N0 can only be: 1, 2, 3, 4, 8, and 16");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.k0 != 1 && desc.k0 != 2 && desc.k0 != 3 && desc.k0 != 4 && desc.k0 != 8 && desc.k0 != 16,
+ "K0 can only be: 1, 2, 3, 4, 8, and 16");
return Status{};
}
diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h
index fec22b8..c3a70ef 100644
--- a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h
+++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h
@@ -25,6 +25,7 @@
#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDIRECTCONV2D
#include "arm_compute/core/Error.h"
+#include "arm_compute/core/KernelDescriptors.h"
#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h"
#include <memory>
@@ -56,9 +57,15 @@
/** Get fast_relaxed_math flag */
bool fast_relaxed_math() const;
+ /** Set direct convolution descriptor */
+ ClComponentDirectConv2dSettings &direct_conv_descriptor(const DirectConvComputeKernelInfo &desc);
+ /** Get direct convolution descriptor */
+ DirectConvComputeKernelInfo direct_conv_descriptor() const;
+
private:
- bool _export_to_cl_image{ false };
- bool _fast_relaxed_math{ true };
+ bool _export_to_cl_image{ false };
+ bool _fast_relaxed_math{ true };
+ DirectConvComputeKernelInfo _desc{}; // Direct convolution descriptor
};
/** Forward declaration */
diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp
index 9cb4ee7..048ee01 100644
--- a/src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp
+++ b/src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp
@@ -23,16 +23,19 @@
*/
#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h"
+#include "arm_compute/core/KernelDescriptors.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "src/common/utils/Log.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/dynamic_fusion/sketch/ArgumentPack.h"
#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h"
#include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
-
-#include "src/common/utils/Log.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
namespace arm_compute
{
@@ -85,6 +88,16 @@
return true;
}
+DirectConvComputeKernelInfo config_direct_convolution_nhwc(const ITensorInfo *src, const ITensorInfo *weights, const PadStrideInfo &conv_info)
+{
+ // Get GPU target
+ GPUTarget gpu_target = CLScheduler::get().target();
+
+ std::unique_ptr<arm_compute::cl_direct_conv::IClDirectConvKernelConfig> t = arm_compute::cl_direct_conv::ClDirectConvKernelConfigurationFactory::create(gpu_target);
+
+ return t->configure(src, weights, conv_info);
+}
+
constexpr GpuOperatorType operator_type = GpuOperatorType::Complex;
} // namespace
@@ -112,6 +125,11 @@
attributes.pad().right,
attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR)); // use the default DimensionRoundingType
+ // Checks performed when dst is configured
+ if(dst->total_size() != 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), shape);
+ }
auto_init_if_empty(dst_info_to_validate, src->clone()->set_tensor_shape(shape));
}
@@ -175,6 +193,12 @@
const Conv2dAttributes &attributes)
{
ARM_COMPUTE_LOG_PARAMS(src, wei, bia, dst, attributes);
+ PadStrideInfo conv_info(attributes.stride().x(), attributes.stride().y(), attributes.pad().left,
+ attributes.pad().right,
+ attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR);
+ // Initialize the direct convolution descriptor
+ const DirectConvComputeKernelInfo desc = config_direct_convolution_nhwc(src, wei, conv_info);
+
// Assert validation
ARM_COMPUTE_ERROR_THROW_ON(GpuConv2d::validate_op(sketch, src, wei, bia, dst, attributes));
ARM_COMPUTE_ERROR_ON_NULLPTR(src, wei, dst);
@@ -182,10 +206,7 @@
// Auto initialize dst tensor
{
- auto shape = misc::shape_calculator::compute_deep_convolution_shape(src->tensor_shape(), data_layout, wei->tensor_shape(),
- PadStrideInfo(attributes.stride().x(), attributes.stride().y(), attributes.pad().left,
- attributes.pad().right,
- attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR)); // use the default DimensionRoundingType
+ auto shape = misc::shape_calculator::compute_deep_convolution_shape(src->tensor_shape(), data_layout, wei->tensor_shape(), conv_info); // use the default DimensionRoundingType
auto_init_if_empty(*dst, src->clone()->set_tensor_shape(shape));
}
@@ -221,6 +242,8 @@
arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(wei);
}
+ settings.direct_conv_descriptor(desc);
+
ArgumentPack<ITensorInfo> arguments;
arguments.add_const_tensor(ACL_SRC_0, src);
arguments.add_const_tensor(ACL_SRC_1, wei);
diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp
index 75e812a..6f7bf72 100644
--- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp
+++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp
@@ -69,7 +69,7 @@
ARM_COMPUTE_UNUSED(comp_group);
const auto channel_idx = get_data_layout_dimension_index(_src->data_layout(), DataLayoutDimension::CHANNEL);
- const auto k0 = adjust_vec_size(is_data_type_quantized(_src->data_type()) ? 16u : 8u, _src->dimension(channel_idx));
+ const auto k0 = adjust_vec_size(_settings.direct_conv_descriptor().k0, _src->dimension(channel_idx));
const bool leftover_loop = (_src->dimension(channel_idx) % k0) != 0;
std::string code = R"_(
@@ -303,13 +303,11 @@
CLBuildOptions ClTemplateDirectConv2d::get_build_options(const ComponentGroup &comp_group) const
{
const unsigned int channel_idx = get_data_layout_dimension_index(_src->data_layout(), DataLayoutDimension::CHANNEL);
- const DataType data_type = _src->data_type();
- /// NOTE: For now tile sizes (n0, m0, k0) are set by the execution window. This may change in the future
const auto root_window = comp_group.get_root_component()->template_writer()->get_window();
const unsigned int n0 = root_window.x().step();
const unsigned int m0 = root_window.y().step();
- const unsigned int k0 = adjust_vec_size(is_data_type_quantized(data_type) ? 16u : 8u, _src->dimension(channel_idx));
+ const unsigned int k0 = adjust_vec_size(_settings.direct_conv_descriptor().k0, _src->dimension(channel_idx));
const unsigned int partial_store_n0 = _dst->dimension(0) % n0;
CLBuildOptions build_opts{};
@@ -369,15 +367,16 @@
ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
const auto output_shape = _dst->tensor_shape();
+ const auto desc = _settings.direct_conv_descriptor();
- const unsigned int vec_size = std::min(static_cast<unsigned int>(output_shape[0]), 4u);
- const unsigned int num_rows = (_dst->tensor_shape()[0] > 16) ? ((_src->data_type() == DataType::F32) ? 2U : 4U) : 1U;
+ const unsigned int n0 = adjust_vec_size(desc.n0, output_shape[0]);
+ const unsigned int m0 = adjust_vec_size(desc.m0, output_shape[1] * output_shape[2]);
// Create and configure kernel window
- Window win = calculate_max_window(output_shape, Steps(vec_size, num_rows));
+ Window win = calculate_max_window(output_shape, Steps(n0, m0));
- const size_t dim_y_collapsed = ceil_to_multiple(output_shape[1] * output_shape[2], num_rows);
- win.set(Window::DimY, Window::Dimension(0, dim_y_collapsed, num_rows));
+ const size_t dim_y_collapsed = ceil_to_multiple(output_shape[1] * output_shape[2], m0);
+ win.set(Window::DimY, Window::Dimension(0, dim_y_collapsed, m0));
win.set(Window::DimZ, Window::Dimension(0, output_shape.total_size_upper(3), 1));
return win;
diff --git a/src/gpu/cl/operators/ClDirectConv2d.cpp b/src/gpu/cl/operators/ClDirectConv2d.cpp
index ded275d..0215dba 100644
--- a/src/gpu/cl/operators/ClDirectConv2d.cpp
+++ b/src/gpu/cl/operators/ClDirectConv2d.cpp
@@ -30,10 +30,10 @@
#include "src/core/helpers/AutoConfiguration.h"
#include "src/gpu/cl/kernels/ClActivationKernel.h"
#include "src/gpu/cl/kernels/ClDirectConv2dKernel.h"
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h"
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h"
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvKernelConfig.h"
-#include "src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
#include "src/common/utils/Log.h"
diff --git a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.cpp b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp
similarity index 98%
rename from src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.cpp
rename to src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp
index ba176f8..1bfb812 100644
--- a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.cpp
+++ b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp
@@ -21,7 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
diff --git a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h
similarity index 88%
rename from src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h
rename to src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h
index 1e4cb66..6b60b2c 100644
--- a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h
+++ b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h
@@ -21,10 +21,10 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#ifndef ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_BIFROST_H
-#define ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_BIFROST_H
+#ifndef SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGBIFROST
+#define SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGBIFROST
-#include "src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
namespace arm_compute
{
@@ -52,4 +52,4 @@
};
} // namespace opencl
} // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_BIFROST_H */
+#endif /* SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGBIFROST */
diff --git a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.cpp b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp
similarity index 98%
rename from src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.cpp
rename to src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp
index b693568..8f2fd82 100644
--- a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.cpp
+++ b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp
@@ -21,7 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
diff --git a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h
similarity index 88%
rename from src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h
rename to src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h
index 2c65b88..f9d5c52 100644
--- a/src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h
+++ b/src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h
@@ -21,10 +21,10 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#ifndef ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_VALHALL_H
-#define ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_VALHALL_H
+#ifndef SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGVALHALL
+#define SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGVALHALL
-#include "src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
namespace arm_compute
{
@@ -52,4 +52,4 @@
};
} // namespace opencl
} // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_DIRECT_CONV_DEFAULT_CONFIG_VALHALL_H */
+#endif /* SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVDEFAULTCONFIGVALHALL */
diff --git a/src/gpu/cl/kernels/direct_conv/ClDirectConvKernelConfig.h b/src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h
similarity index 82%
rename from src/gpu/cl/kernels/direct_conv/ClDirectConvKernelConfig.h
rename to src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h
index c1c2e43..232167f 100644
--- a/src/gpu/cl/kernels/direct_conv/ClDirectConvKernelConfig.h
+++ b/src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h
@@ -21,12 +21,12 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#ifndef ARM_COMPUTE_CL_DIRECT_CONV_KERNEL_CONFIGURATION_H
-#define ARM_COMPUTE_CL_DIRECT_CONV_KERNEL_CONFIGURATION_H
+#ifndef SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVKERNELCONFIG
+#define SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVKERNELCONFIG
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigBifrost.h"
-#include "src/gpu/cl/kernels/direct_conv/ClDirectConvDefaultConfigValhall.h"
-#include "src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h"
+#include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h"
+#include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
#include <memory>
@@ -61,4 +61,4 @@
};
} // namespace opencl
} // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_DIRECT_CONV_KERNEL_CONFIGURATION_H */
+#endif /* SRC_RUNTIME_HEURISTICS_DIRECT_CONV_CLDIRECTCONVKERNELCONFIG */
diff --git a/src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h b/src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h
similarity index 94%
rename from src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h
rename to src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h
index 837fa35..6104d73 100644
--- a/src/gpu/cl/kernels/direct_conv/IClDirectConvKernelConfig.h
+++ b/src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h
@@ -21,8 +21,8 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#ifndef ARM_COMPUTE_ICL_DIRECT_CONV_KERNEL_CONFIG_H
-#define ARM_COMPUTE_ICL_DIRECT_CONV_KERNEL_CONFIG_H
+#ifndef SRC_RUNTIME_HEURISTICS_DIRECT_CONV_ICLDIRECTCONVKERNELCONFIG
+#define SRC_RUNTIME_HEURISTICS_DIRECT_CONV_ICLDIRECTCONVKERNELCONFIG
#include "arm_compute/core/GPUTarget.h"
#include "arm_compute/core/KernelDescriptors.h"
@@ -112,4 +112,4 @@
};
} // namespace opencl
} // namespace arm_compute
-#endif /* ARM_COMPUTE_ICL_DIRECT_CONV_KERNEL_CONFIG_H */
+#endif /* SRC_RUNTIME_HEURISTICS_DIRECT_CONV_ICLDIRECTCONVKERNELCONFIG */