Rewrite dynamic fusion
The new version introduces the following major changes:
* Change public interface to simplify and standardize the user experience
- Use the term "Workload" uniformly
- Simplify operator interface to be a set of static methods:
validate_op(), create_op()
* Separate the kernel writing into its own component (template_writer).
This is to allow the co-development of GpuKernelWriter, and to allow
easy replacement once GpuKernelWriter is mature.
* Optimize the core fusion algorithm used by the component graph. The
details can be found in GpuKernelComponentGraph::fuse()
* Use Gpu instead of Cl prefixes for most of the Workload interfaces
(except for runtime and kernel components, which have to be language specific)
This allows the potential extension to other Gpu langauges in the
future.
* Refactor runtime memory interface so that auxiliary tensor handling
is separate from the user tensor passing. This is because the former
is less stable and may require extension in the future.
* Hide source code object from the user as it is not required at the
moment
* Deprecate the old prototype entirely by disabling it in SCons build
Resolves COMPMID-5510, COMPMID-5512, COMPMID-5513
Change-Id: If69d2362856f2de4503546b7b6cf48a525cf3079
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8406
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp
new file mode 100644
index 0000000..870de64
--- /dev/null
+++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp
@@ -0,0 +1,400 @@
+/*
+ * Copyright (c) 2022 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.
+ */
+#include "ClTemplateDirectConv2d.h"
+
+#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
+#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h"
+
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include "support/StringSupport.h"
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+ClTemplateDirectConv2d::ClTemplateDirectConv2d(ComponentId id,
+ const ArgumentPack<ITensorInfo> &tensors,
+ const Attributes &attributes,
+ const Settings &settings)
+ : IGpuTemplateComponentWriter{ id, tensors },
+ _src{},
+ _weight{},
+ _bias{},
+ _dst{},
+ _attributes{ attributes },
+ _settings{ settings }
+{
+ _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
+ _weight = this->tensors().get_const_tensor(TensorType::ACL_SRC_1);
+ if(this->tensors().get_const_tensor(TensorType::ACL_SRC_2))
+ {
+ _bias = this->tensors().get_const_tensor(TensorType::ACL_SRC_2);
+ }
+ _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
+ ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _weight, _dst);
+}
+
+std::string ClTemplateDirectConv2d::get_name() const
+{
+ return "direct_conv2d";
+}
+
+std::string ClTemplateDirectConv2d::get_component_code(const ComponentGroup &comp_group) const
+{
+ 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 bool leftover_loop = (_src->dimension(channel_idx) % k0) != 0;
+
+ std::string code = R"_(
+//------------------ START KERNEL {{meta_kernel_id}} ---------------------
+// IN_0(src) {{src}}
+// IN_1(wei) {{weight}}
+)_";
+ if(_bias && _bias->has_valid_id())
+ {
+ code += R"_(
+// IN_1(bia) {{bias}}
+)_";
+ }
+ code += R"_(
+// OUT(dst, accum) {{dst}}
+
+// Initialize the accumulators
+TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}});
+{
+ // All the tensor dimensions are passed at compile time.
+ // In case of dynamic tensor support, the following dimensions should be passed as function argument.
+#define _IWEI_WIDTH {{WEI_WIDTH}}
+#define _IWEI_HEIGHT {{WEI_HEIGHT}}
+#define _ISRC_WIDTH {{src}}_w
+#define _ISRC_HEIGHT {{src}}_h
+#define _ISRC_CHANNELS {{src}}_c
+#define _IDST_WIDTH {{arg_dst}}_w
+#define _IDST_HEIGHT {{arg_dst}}_h
+#define _IDST_CHANNELS {{arg_dst}}_c
+#define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT)
+
+ // .v = access the whole vector (OpenCL vector)
+ // .s[x] = access the vector element at position x (scalar access)
+ TILE(int, M0, 1, xi);
+ TILE(int, M0, 1, yi);
+
+ // Convert the linear index to coordinate
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ xi[i].v = ((g_ind_1 + i) % _IDST_WIDTH) * {{STRIDE_X}};
+ yi[i].v = ((g_ind_1 + i) / _IDST_WIDTH) * {{STRIDE_Y}};
+ xi[i].v -= {{PAD_LEFT}};
+ yi[i].v -= {{PAD_TOP}};
+ })
+
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ {{dst}}[i].v = 0;
+ })
+
+ for(int i = 0; i < (_IWEI_WIDTH * _IWEI_HEIGHT); ++i)
+ {
+ int ck = 0;
+ int xk = i % _IWEI_WIDTH;
+ int yk = i / _IWEI_WIDTH;
+
+ int k = 0;
+ for(; k <= (_ISRC_CHANNELS - K0); k += K0)
+ {
+ TILE({{SRC_DATA_TYPE}}, M0, K0, a);
+ TILE({{WEI_DATA_TYPE}}, N0, K0, b);
+
+ // Initialize tiles
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ a[i].v = {{ZERO_VALUE}};
+ })
+
+ LOOP_UNROLLING(int, i, 0, 1, N0,
+ {
+ b[i].v = {{ZERO_VALUE}};
+ })
+
+ // Load tile from the src tensor
+ T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, K0, {{SRC_TENSOR_TYPE}}, {{src}}, g_ind_2, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a);
+
+ // Load tile from the weights tensor
+ T_LOAD({{WEI_DATA_TYPE}}, N0, K0, {{WEI_TENSOR_TYPE}}, {{weight}}, ck, g_ind_0 * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b);
+
+ // Compute the matrix multiplication between two tiles
+ T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, K0, NT, T, a, b, {{dst}});
+
+ ck += K0;
+ }
+
+ // We voluntarily use SRC_CHANNELS rather than _DSRC_CHANNELS
+ // This #if directive should be removed in case of dynamic tensor support
+)_";
+
+ if(leftover_loop)
+ {
+ code += R"_(
+ // Left-over accumulations
+ for(; k < _ISRC_CHANNELS; ++k)
+ {
+ TILE({{SRC_DATA_TYPE}}, M0, 1, a);
+ TILE({{WEI_DATA_TYPE}}, N0, 1, b);
+
+ // Initialize tiles
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ a[i].v = {{ZERO_VALUE}};
+ })
+
+ LOOP_UNROLLING(int, i, 0, 1, N0,
+ {
+ b[i].v = {{ZERO_VALUE}};
+ })
+
+ // Load tile from the src tensor
+ T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, 1, {{SRC_TENSOR_TYPE}}, {{src}}, g_ind_2, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a);
+
+ // Load tile from the weights tensor
+ // The T_LOAD for the left-over elements can only use BUFFER because we load one element per iteration
+ T_LOAD({{WEI_DATA_TYPE}}, N0, 1, BUFFER, {{weight}}, ck, g_ind_0 * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b);
+
+ // Compute the matrix multiplication between two tiles
+ T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, 1, NT, T, a, b, {{dst}});
+
+ ++ck;
+ }
+ )_";
+}
+
+code += R"_(
+#undef _I_WEI_WIDTH
+#undef _I_WEI_HEIGHT
+#undef _ISRC_WIDTH
+#undef _ISRC_HEIGHT
+#undef _ISRC_CHANNELS
+#undef _IDST_WIDTH
+#undef _IDST_HEIGHT
+#undef _IDST_CHANNELS
+#undef _IY_MULTIPLIER
+
+ }
+)_";
+
+ if(_bias && _bias->has_valid_id())
+ {
+ code += R"_(
+ TILE({{BIA_DATA_TYPE}}, 1, N0, bias0);
+
+ T_LOAD({{BIA_DATA_TYPE}}, 1, N0, BUFFER, {{bias}}, g_ind_0, 0, 1, 0, bias0);
+
+ // c = c + bias[broadcasted]
+ T_ELTWISE_BROADCAST_ADD_X({{ACC_DATA_TYPE}}, M0, N0, {{dst}}, bias0, {{dst}});
+ )_";
+}
+
+code += R"_(
+}
+//------------------ END KERNEL {{meta_kernel_id}} ---------------------
+)_";
+ return code;
+}
+
+void ClTemplateDirectConv2d::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
+{
+ vtable.declare_variable(
+ _src,
+ GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
+ comp_group.is_intermediate_tensor(_src),
+ "src");
+
+ const GpuKernelArgumentInfo::Type weight_type = _settings.export_to_cl_image() ? GpuKernelArgumentInfo::Type::Tensor_4D_t_Image : GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer;
+ vtable.declare_variable(
+ _weight,
+ GpuKernelArgumentInfo(weight_type),
+ comp_group.is_intermediate_tensor(_weight),
+ "weight");
+
+ if(_bias && _bias->has_valid_id()) // optional bias
+ {
+ vtable.declare_variable(
+ _bias,
+ GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Vector),
+ comp_group.is_intermediate_tensor(_bias),
+ "bias");
+ }
+ vtable.declare_variable(
+ _dst,
+ GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
+ comp_group.is_intermediate_tensor(_dst),
+ "dst");
+}
+
+TagLUT ClTemplateDirectConv2d::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
+{
+ TagLUT lut{};
+ // Arguments and global shared variables
+ lut["src"] = vtable.get_variable(_src);
+ lut["weight"] = vtable.get_variable(_weight);
+
+ if(_bias && _bias->has_valid_id()) // optional bias
+ {
+ lut["bias"] = vtable.get_variable(_bias);
+ lut["BIA_DATA_TYPE"] = get_cl_type_from_data_type(_bias->data_type());
+ }
+ lut["dst"] = vtable.get_variable(_dst);
+
+ const auto dst_argument = vtable.get_variable(comp_group.get_dst_tensors()[0]);
+ lut["arg_dst"] = dst_argument.uniq_name;
+
+ // Local build options
+ lut["meta_kernel_id"] = id();
+ lut["ACC_DATA_TYPE"] = _src->data_type();
+ lut["SRC_DATA_TYPE"] = _src->data_type();
+ lut["WEI_DATA_TYPE"] = _weight->data_type();
+
+ lut["SRC_TENSOR_TYPE"] = "BUFFER";
+ switch(vtable.get_variable(_weight).kernel_argument_info.type)
+ {
+ case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D:
+ case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D:
+ case GpuKernelArgumentInfo::Type::Tensor_4D_t_Image:
+ {
+ lut["WEI_TENSOR_TYPE"] = "IMAGE";
+ break;
+ }
+ default:
+ {
+ lut["WEI_TENSOR_TYPE"] = "BUFFER";
+ break;
+ }
+ }
+ const auto width_idx = 1;
+ const auto height_idx = 2;
+ lut["WEI_WIDTH"] = _weight->dimension(width_idx);
+ lut["WEI_HEIGHT"] = _weight->dimension(height_idx);
+
+ lut["STRIDE_X"] = _attributes.stride().x();
+ lut["STRIDE_Y"] = _attributes.stride().y();
+
+ lut["PAD_LEFT"] = _attributes.pad().left;
+ lut["PAD_TOP"] = _attributes.pad().top;
+
+ lut["ZERO_VALUE"] = 0;
+
+ return lut;
+}
+
+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, n0) 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 partial_store_n0 = _dst->dimension(0) % n0;
+
+ CLBuildOptions build_opts{};
+ if(_settings.fast_relaxed_math())
+ {
+ build_opts.add_option("-cl-fast-relaxed-math");
+ }
+ else
+ {
+ // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations
+ // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations
+ build_opts.add_option("-cl-unsafe-math-optimizations");
+ }
+ build_opts.add_option("-DIS_TILED");
+ build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
+ build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
+ build_opts.add_option("-DK0=" + support::cpp11::to_string(k0));
+ build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
+
+ return build_opts;
+}
+
+std::string ClTemplateDirectConv2d::get_config_id() const
+{
+ const DataType data_type = _src->data_type();
+ const DataLayout data_layout = _src->data_layout();
+
+ const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+ const unsigned int kernel_size = _weight->dimension(width_idx);
+
+ std::string config_id{};
+ config_id += lower_string(string_from_data_type(data_type));
+ config_id += "_";
+ config_id += support::cpp11::to_string(kernel_size);
+ config_id += "_";
+ config_id += support::cpp11::to_string(_attributes.stride().x());
+ config_id += "_";
+ config_id += support::cpp11::to_string(_attributes.stride().y());
+ config_id += "_";
+ config_id += support::cpp11::to_string(_dst->dimension(width_idx));
+ config_id += "_";
+ config_id += support::cpp11::to_string(_dst->dimension(height_idx));
+ config_id += "_";
+ config_id += lower_string(string_from_data_layout(data_layout));
+ return config_id;
+}
+
+std::set<std::string> ClTemplateDirectConv2d::get_headers_list() const
+{
+ return std::set<std::string>{ "helpers.h", "tile_helpers.h" };
+}
+
+Window ClTemplateDirectConv2d::get_window() const
+{
+ ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
+
+ const auto output_shape = _dst->tensor_shape();
+
+ 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;
+
+ // Create and configure kernel window
+ Window win = calculate_max_window(output_shape, Steps(vec_size, num_rows));
+
+ 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));
+ win.set(Window::DimZ, Window::Dimension(0, output_shape.total_size_upper(3), 1));
+
+ return win;
+}
+
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute