Handle Intermediate tensors within the sketch
- Intermediate tensor info objects are not created by the user anymore. They're returned from create_op and reused. This will prevent allocation of the intermediate tensors in case of possible interface misuse.
- Sketch object handles intermediate tensor info pointers inside its implementation class via a unique pointer vector
- Conv2d operator is migrated into the new interface
Resolves: COMPMID-5776
Change-Id: I9422e3681eef4f2d2922f6d0a5d7786380837c6d
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8906
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@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/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h b/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
index afbe2b8..f19ad6d 100644
--- a/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
+++ b/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2022 Arm Limited.
+ * Copyright (c) 2022-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -83,7 +83,7 @@
*/
TensorInfo create_tensor_info(const ITensorInfo &tensor_info);
/** Create a default @ref TensorInfo associated with the workload sketch
- * It is usually used by a destination tensor whose @ref ITensorInfo is to be inferred automatically
+ * It is usually used by user input or output tensors
*
* @return TensorInfo Newly created tensor info
*/
diff --git a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h
index 76decfd..4ba237e 100644
--- a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h
+++ b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2022 Arm Limited.
+ * Copyright (c) 2022-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -60,32 +60,35 @@
* @param[in] src Source tensor
* @param[in] wei Weight tensor
* @param[in] bia (Optional) Bias tensor
- * @param[out] dst Destination tensor. If an uninitialized ITensorInfo is passed in, it will be auto-initialized
* @param[in] attributes Operator attributes
+ *
+ * @return pointer for the destination tensor
*/
- static void create_op(GpuWorkloadSketch &sketch,
- ITensorInfo *src,
- ITensorInfo *wei,
- ITensorInfo *bia,
- ITensorInfo *dst,
- const Attributes &attributes);
+ static ITensorInfo *create_op(GpuWorkloadSketch &sketch,
+ ITensorInfo *src,
+ ITensorInfo *wei,
+ ITensorInfo *bia,
+ const Attributes &attributes);
/** Check if the operator configuration is supported, irrespective of fusion
- * Similar to @ref GpuConv2d::create_op()
+ *
+ * @param[in] context Workload context within which the operator is running
+ * @param[in] src Source tensor
+ * @param[in] wei Weight tensor
+ * @param[in] bia (Optional) Bias tensor
+ * @param[in] attributes Operator attributes
*/
static Status is_supported_op(const GpuWorkloadContext &context,
const ITensorInfo *src,
const ITensorInfo *wei,
const ITensorInfo *bia,
- const ITensorInfo *dst,
const Attributes &attributes);
/** Check if the operator configuration is supported and if it can be fused into the workload sketch.
- * Similar to @ref GpuConv2d::create_op()
+ * Similar to @ref GpuConv2d::create_op()
*/
static Status validate_op(const GpuWorkloadSketch &sketch,
const ITensorInfo *src,
const ITensorInfo *wei,
const ITensorInfo *bia,
- const ITensorInfo *dst,
const Attributes &attributes);
};
} // namespace dynamic_fusion