Use the stable CKW API in the GPU dynamic fusion backend
- Refactor all kernels to work with the CKW stable API
- Add support for sub-tile in the op_load/op_store CKW operator
- Fix mismatch in resize
- Add comments in all kernels written with CKW to help developers
understand the structure of the code
- Add texture image support in depthwise convolution written with CKW
- Add support for different block sizes in depthwise convolution
- Remove the use of the dynamic fusion helper functions.
- Add support for floor in the op_unary() of CKW
Resolves: COMPMID-6708, COMPMID-6743, COMPMID-6530
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Change-Id: I8104ce4d04a3138a1aeb0b84940e1f1c89e76069
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10914
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp
index a98ebed..7d16f35 100644
--- a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp
+++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp
@@ -34,15 +34,16 @@
#include "src/cl/CLTile.h"
#include "src/ITensor.h"
#include "src/Tensor3dMapper.h"
+#include "src/TileView.h"
namespace ckw
{
-bool CLMemoryOpBufferHelper::validate(const CLKernelWriter *writer,
- const ITensor *tensor,
- const TensorSampler *sampler,
- const Tensor3dMapper *mapper,
- MemoryOperation op,
- const CLTile *dst)
+bool CLMemoryOpBufferHelper::validate(const CLKernelWriter *writer,
+ const ITensor *tensor,
+ const TensorSampler *sampler,
+ const Tensor3dMapper *mapper,
+ MemoryOperation op,
+ const TileView<CLTile> &dst)
{
CKW_UNUSED(writer, tensor, mapper, op, dst);
@@ -100,17 +101,14 @@
* The outermost block is x, then z and then y. This is why, if/else's covering for y are initialized
* at each row write. In some addressing modes, such as None, no if/else conditions are written.
*/
-void CLMemoryOpBufferHelper::initialize(const CLTile *dst, const CLTile *x, const CLTile *z, const CLTile *b)
+void CLMemoryOpBufferHelper::initialize(const CLTile *x, const CLTile *z, const CLTile *b)
{
- _dst = dst;
-
CKW_ASSERT(validate(_writer, _tensor, _sampler, _mapper.get(), _op, _dst));
- _ls_width_full = dst->info().width();
- _coord_x = x->scalar(0, 0).str;
- _coord_z = z->scalar(0, 0).str;
- _coord_b = b->scalar(0, 0).str;
- _coord_orig_z = _coord_z;
+ _coord_x = x->scalar(0, 0).str;
+ _coord_z = z->scalar(0, 0).str;
+ _coord_b = b->scalar(0, 0).str;
+ _coord_orig_z = _coord_z;
out_of_bound_initialize_x(_coord_x);
out_of_bound_initialize_z(_coord_z);
@@ -121,7 +119,7 @@
// The only check required is on Y.
out_of_bound_initialize_y(coord_y);
- const std::string dst = _dst->vector(row_id).str;
+ const std::string dst = _dst.vector(row_id).str;
const std::string address = to_buffer_address(_coord_x, coord_y, _coord_z, _coord_b);
const std::string ls_buf = to_statement(_op, _ls_width_full, dst, address);
@@ -133,10 +131,17 @@
// The left over load/store will be written in the finalize stage
if (_ls_width_part.size() != 0)
{
- int32_t col_start = 0;
+ int32_t col_start = 0;
+ const TileArea original_area = _dst.area();
+
for (int32_t partial_width : _ls_width_part)
{
- const std::string dst = _dst->vector(row_id, col_start, partial_width).str;
+ // Set the active area
+ const TileArea area(original_area.row_start(), original_area.row_end(), col_start,
+ col_start + partial_width);
+ _dst.area(area);
+
+ const std::string dst = _dst.vector(row_id).str;
const std::string coord_x = _coord_x + " + " + std::to_string(col_start);
const std::string address = to_buffer_address(coord_x, coord_y, _coord_z, _coord_b);
const std::string statement = to_statement(_op, partial_width, dst, address);
@@ -144,6 +149,8 @@
col_start += partial_width;
}
+ // Restore the original area
+ _dst.area(original_area);
}
}
@@ -304,7 +311,7 @@
CKW_ASSERT(tensor_storage == TensorStorageType::BufferUint8Ptr);
const std::string ptr_buf = _tensor->storage(tensor_storage).val;
- const std::string dst_type = cl_data_type_rounded_up_to_valid_vector_width(_dst->info().data_type(), 1);
+ const std::string dst_type = cl_data_type_rounded_up_to_valid_vector_width(_dst.data_type(), 1);
std::string address;
address += "(__global ";
diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h
index 4e1a842..a6b3272 100644
--- a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h
+++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h
@@ -22,8 +22,8 @@
* SOFTWARE.
*/
-#ifndef CKW_SRC_CL_CLMEMORYOPBUFFERHELPER_H
-#define CKW_SRC_CL_CLMEMORYOPBUFFERHELPER_H
+#ifndef CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H
+#define CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H
#include "src/cl/helpers/ICLMemoryOpHelper.h"
@@ -37,6 +37,8 @@
// Forward Declarations
class CLKernelWriter;
class CLTile;
+template <class CLTile>
+class TileView;
enum class MemoryOperation;
/** Helper class to write memory operations (like load/store) in OpenCL
@@ -45,19 +47,23 @@
{
public:
/** Constructor similar to @ref ICLMemoryOpHelper() */
- CLMemoryOpBufferHelper(CLKernelWriter *writer, ITensor *tensor, TensorSampler *sampler, MemoryOperation op)
- : ICLMemoryOpHelper(writer, tensor, sampler, op)
+ CLMemoryOpBufferHelper(CLKernelWriter *writer,
+ ITensor *tensor,
+ TensorSampler *sampler,
+ MemoryOperation op,
+ const TileView<CLTile> &dst)
+ : ICLMemoryOpHelper(writer, tensor, sampler, op, dst)
{
}
/** Copy constructor */
- CLMemoryOpBufferHelper(const CLMemoryOpBufferHelper &) = default;
+ CLMemoryOpBufferHelper(const CLMemoryOpBufferHelper &) = delete;
/** Assignment operator overload */
- CLMemoryOpBufferHelper &operator=(const CLMemoryOpBufferHelper &) = default;
+ CLMemoryOpBufferHelper &operator=(const CLMemoryOpBufferHelper &) = delete;
// Methods overridden
- void initialize(const CLTile *dst, const CLTile *x, const CLTile *z, const CLTile *b) override;
+ void initialize(const CLTile *x, const CLTile *z, const CLTile *b) override;
void write_row(int32_t row_id, const std::string &coord_y) override;
void finalize() override;
@@ -78,12 +84,12 @@
std::vector<LeftoverDescriptor> _leftovers_x{};
std::string _coord_orig_z{};
- static bool validate(const CLKernelWriter *writer,
- const ITensor *tensor,
- const TensorSampler *sampler,
- const Tensor3dMapper *mapper,
- MemoryOperation op,
- const CLTile *dst);
+ static bool validate(const CLKernelWriter *writer,
+ const ITensor *tensor,
+ const TensorSampler *sampler,
+ const Tensor3dMapper *mapper,
+ MemoryOperation op,
+ const TileView<CLTile> &dst);
void out_of_bound_initialize_x(const std::string &coord);
void out_of_bound_finalize_x();
@@ -99,4 +105,4 @@
};
} // namespace ckw
-#endif /* CKW_SRC_CL_CLMEMORYOPBUFFERHELPER_H */
+#endif // CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H
diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp
index b7d146b..f392cd8 100644
--- a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp
+++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp
@@ -33,18 +33,15 @@
#include "src/cl/CLTile.h"
#include "src/ITensor.h"
#include "src/Tensor3dMapper.h"
+#include "src/TileView.h"
namespace ckw
{
-void CLMemoryOpImage2dHelper::initialize(const CLTile *dst, const CLTile *x, const CLTile *z, const CLTile *b)
+void CLMemoryOpImage2dHelper::initialize(const CLTile *x, const CLTile *z, const CLTile *b)
{
- CKW_ASSERT(validate(_writer, _tensor, _sampler, _mapper.get(), _op, dst));
-
- _dst = dst;
- _ls_width_full = dst->info().width();
- _coord_x = x->scalar(0, 0).str;
- _coord_z = z->scalar(0, 0).str;
- _coord_b = b->scalar(0, 0).str;
+ _coord_x = x->scalar(0, 0).str;
+ _coord_z = z->scalar(0, 0).str;
+ _coord_b = b->scalar(0, 0).str;
}
void CLMemoryOpImage2dHelper::write_row(int32_t row_id, const std::string &coord_y)
@@ -52,7 +49,7 @@
// The only check required is on Y.
out_of_bound_initialize_y(coord_y);
- const std::string dst = _dst->vector(row_id).str;
+ const std::string dst = _dst.vector(row_id).str;
const std::string sampler = to_ls_image2d_sampler();
const std::string coord = to_ls_image2d_address(_coord_x, coord_y, _coord_z, _coord_b);
const std::string ls_buf = to_ls_image2d(_op, _ls_width_full, dst, sampler, coord);
@@ -66,16 +63,16 @@
{
}
-bool CLMemoryOpImage2dHelper::validate(const CLKernelWriter *writer,
- const ITensor *tensor,
- const TensorSampler *sampler,
- const Tensor3dMapper *mapper,
- MemoryOperation op,
- const CLTile *dst)
+bool CLMemoryOpImage2dHelper::validate(const CLKernelWriter *writer,
+ const ITensor *tensor,
+ const TensorSampler *sampler,
+ const Tensor3dMapper *mapper,
+ MemoryOperation op,
+ const TileView<CLTile> &dst)
{
CKW_UNUSED(writer, tensor, mapper);
- if (dst->info().width() != 4)
+ if (dst.width() != 4)
{
return false;
}
@@ -95,7 +92,7 @@
{
return false;
}
- if ((dst->info().data_type() != DataType::Fp32) && (dst->info().data_type() != DataType::Fp16))
+ if ((dst.data_type() != DataType::Fp32) && (dst.data_type() != DataType::Fp16))
{
return false;
}
@@ -143,10 +140,12 @@
const std::string &address) const
{
CKW_UNUSED(vector_width);
+ CKW_ASSERT_MSG(_dst.data_type() == DataType::Fp32 || _dst.data_type() == DataType::Fp16,
+ "Image2d only supports floating-point data type");
const TensorStorageType tensor_storage = _sampler->storage();
const std::string image2d_obj = _tensor->storage(tensor_storage).val;
- const std::string post_fix = _dst->info().data_type() == DataType::Fp32 ? "f" : "h";
+ const std::string post_fix = _dst.data_type() == DataType::Fp32 ? "f" : "h";
switch (op)
{
diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h
index fd9b097..6c42c13 100644
--- a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h
+++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h
@@ -35,6 +35,8 @@
// Forward Declarations
class CLKernelWriter;
class CLTile;
+template <class CLTile>
+class TileView;
enum class MemoryOperation;
/** Helper class to write memory operations (like load/store) in OpenCL for Image2d type */
@@ -42,29 +44,33 @@
{
public:
/** Constructor similar to @ref ICLMemoryOpHelper() */
- CLMemoryOpImage2dHelper(CLKernelWriter *writer, ITensor *tensor, TensorSampler *sampler, MemoryOperation op)
- : ICLMemoryOpHelper(writer, tensor, sampler, op)
+ CLMemoryOpImage2dHelper(CLKernelWriter *writer,
+ ITensor *tensor,
+ TensorSampler *sampler,
+ MemoryOperation op,
+ const TileView<CLTile> &dst)
+ : ICLMemoryOpHelper(writer, tensor, sampler, op, dst)
{
}
/** Copy constructor */
- CLMemoryOpImage2dHelper(const CLMemoryOpImage2dHelper &) = default;
+ CLMemoryOpImage2dHelper(const CLMemoryOpImage2dHelper &) = delete;
/** Assignment operator overload */
- CLMemoryOpImage2dHelper &operator=(const CLMemoryOpImage2dHelper &) = default;
+ CLMemoryOpImage2dHelper &operator=(const CLMemoryOpImage2dHelper &) = delete;
// Methods overridden
- void initialize(const CLTile *dst, const CLTile *x, const CLTile *z, const CLTile *b) override;
+ void initialize(const CLTile *x, const CLTile *z, const CLTile *b) override;
void write_row(int32_t row_id, const std::string &coord_y) override;
void finalize() override;
private:
- static bool validate(const CLKernelWriter *writer,
- const ITensor *tensor,
- const TensorSampler *sampler,
- const Tensor3dMapper *mapper,
- MemoryOperation op,
- const CLTile *dst);
+ static bool validate(const CLKernelWriter *writer,
+ const ITensor *tensor,
+ const TensorSampler *sampler,
+ const Tensor3dMapper *mapper,
+ MemoryOperation op,
+ const TileView<CLTile> &dst);
void out_of_bound_initialize_y(const std::string &coord);
void out_of_bound_finalize_y();
diff --git a/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h b/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h
index f46fee9..a5b679a 100644
--- a/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h
+++ b/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h
@@ -28,6 +28,7 @@
#include "ckw/TensorSampler.h"
#include "src/Tensor3dMapper.h"
+#include "src/TileView.h"
#include <cstdint>
#include <memory>
@@ -55,18 +56,24 @@
* @param[in] tensor @ref ckw::ITensor object to perform the memory operation on
* @param[in] sampler @ref ckw::TensorSampler object that tells how to sample a tensor
* @param[in] op The memory operation to be done (e.g. Load/Store)
+ * @param[in] dst The tile to perform the memory operation on
*/
- ICLMemoryOpHelper(CLKernelWriter *writer, ITensor *tensor, TensorSampler *sampler, MemoryOperation op)
- : _writer(writer), _tensor(tensor), _sampler(sampler), _op(op)
+ ICLMemoryOpHelper(CLKernelWriter *writer,
+ ITensor *tensor,
+ TensorSampler *sampler,
+ MemoryOperation op,
+ const TileView<CLTile> &dst)
+ : _writer(writer), _tensor(tensor), _sampler(sampler), _op(op), _dst(dst)
{
- _mapper = std::make_unique<Tensor3dMapper>(tensor, sampler->format());
+ _mapper = std::make_unique<Tensor3dMapper>(tensor, sampler->format());
+ _ls_width_full = _dst.width();
}
/** Copy constructor */
- ICLMemoryOpHelper(const ICLMemoryOpHelper &) = default;
+ ICLMemoryOpHelper(const ICLMemoryOpHelper &) = delete;
/** Assignment operator overload */
- ICLMemoryOpHelper &operator=(const ICLMemoryOpHelper &) = default;
+ ICLMemoryOpHelper &operator=(const ICLMemoryOpHelper &) = delete;
/** Destructor */
virtual ~ICLMemoryOpHelper() = default;
@@ -75,12 +82,11 @@
* the batch offset as a tile object, and initializes the code inside
* the writer object.
*
- * @param[in] dst tile object to perform the memory operation on
* @param[in] x tile object that describes the x-coordinate of the tensor involved
* @param[in] z tile object that describes the z-coordinate of the tensor involved
* @param[in] b tile object that describes the batch offset of the tensor involved
*/
- virtual void initialize(const CLTile *dst, const CLTile *x, const CLTile *z, const CLTile *b) = 0;
+ virtual void initialize(const CLTile *x, const CLTile *z, const CLTile *b) = 0;
/** Method that writes the actual code to the writer that performs the mentioned memory
* operation on the tile initialized. It writes the code for a specific row given in the
@@ -104,7 +110,7 @@
TensorSampler *_sampler{nullptr};
MemoryOperation _op;
std::unique_ptr<Tensor3dMapper> _mapper{nullptr};
- const CLTile *_dst{nullptr};
+ TileView<CLTile> _dst{};
int32_t _ls_width_full{0};
std::string _coord_x{};
std::string _coord_z{};
@@ -112,4 +118,4 @@
};
} // namespace ckw
-#endif /* CKW_SRC_CL_HELPERS_ICLMEMORYOPHELPER_H */
+#endif // CKW_SRC_CL_HELPERS_ICLMEMORYOPHELPER_H