Port DirectConv2d to CKW backend
Ports the direct convolution 2D kernel from the experimental Dynamic Fusion interface to use the new Compute Kernel Writer backend for OpenCL code generation.
Support is for FP16/FP32 only.
Resolves: COMPMID-6259
Change-Id: Ia8d7b9cb789737b22b1d877cd798a73eda0ce4ab
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10059
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/compute_kernel_writer/prototype/include/ckw/KernelWriter.h b/compute_kernel_writer/prototype/include/ckw/KernelWriter.h
index c116e62..72f85c7 100644
--- a/compute_kernel_writer/prototype/include/ckw/KernelWriter.h
+++ b/compute_kernel_writer/prototype/include/ckw/KernelWriter.h
@@ -129,11 +129,38 @@
/** Load the data from the tensor memory to the tile using the sampling information.
*
+ * @param[out] tile The tile to be loaded.
+ * @param[in] tensor The tensor to be read.
+ * @param[in] sampler The tensor sampling information.
+ * @param[in] dilation_y Dilation in the Y dimension.
+ */
+ void op_load(TileOperand &tile, const TensorOperand &tensor, const TensorTileSampler &sampler, const TileOperand &dilation_y = TileOperand("dil_y", 1));
+
+ /** Load the data from the tensor memory to the tile using the indirect buffer approach and respective of the sampling information.
+ *
* @param[out] tile The tile to be loaded.
* @param[in] tensor The tensor to be read.
* @param[in] sampler The tensor sampling information.
*/
- void op_load(TileOperand &tile, TensorOperand &tensor, const TensorTileSampler &sampler);
+ void op_load_indirect(TileOperand &tile, const TensorOperand &tensor, const TensorTileSampler &sampler);
+
+ /** Construct an indirection buffer in @p tile containing the precalculated addresses of elements in the source tensor.
+ *
+ * @param[out] tile The tile to be loaded.
+ * @param[in] tensor The tensor the be read.
+ * @param[in] sampler The tensor sampling information.
+ * @param[in] x The X coordinate.
+ * @param[in] y The Y coordinate.
+ * @param[in] x_off Offset in the X dimension.
+ * @param[in] y_off Offset in the Y dimension.
+ */
+ void util_get_indirect_buffer(TileOperand &tile,
+ const TensorOperand &tensor,
+ const TensorTileSampler &sampler,
+ const TileOperand &x,
+ const TileOperand &y,
+ const TileOperand &x_off,
+ const TileOperand &y_off);
/** Store the tile to the tensor using the specified sampling information.
*
diff --git a/compute_kernel_writer/prototype/include/ckw/TileOperand.h b/compute_kernel_writer/prototype/include/ckw/TileOperand.h
index c071707..24ee373 100644
--- a/compute_kernel_writer/prototype/include/ckw/TileOperand.h
+++ b/compute_kernel_writer/prototype/include/ckw/TileOperand.h
@@ -37,6 +37,8 @@
class Kernel;
+using TileContainer = std::vector<std::vector<std::string>>;
+
/** Tile operand which can be either scalar, vector or 2D tile. */
class TileOperand : public OperandBase
{
@@ -69,6 +71,13 @@
*/
TileOperand(const ::std::string &name, float value);
+ /** Initialize a new instance of @ref TileOperand for compile-time constant variable.
+ *
+ * @param[in] name The name of the tile.
+ * @param[in] value The value of the tile.
+ */
+ TileOperand(const ::std::string &name, const ::std::vector<std::vector<std::string>> &value, DataType dt);
+
/** Prohibit copy of tile operand. */
TileOperand(const TileOperand &) = delete;
@@ -96,13 +105,21 @@
/** Get the scalar value of the tile.
*
* The tile must have the shape of 1, 1 (i.e. scalar).
+ *
+ * @return Scalar value as a string.
*/
- ScalarValue scalar_value() const;
+ std::string scalar_value() const;
+
+ /** Get the values of the tile.
+ *
+ * @return 2D container of values.
+ */
+ const TileContainer &value() const;
private:
- TileInfo _info;
- ScalarValue _value{};
- bool _constant;
+ TileInfo _info;
+ TileContainer _value{};
+ bool _constant;
};
} // namespace ckw