COMPMID-2043: Add support for "dummy threads" in CLGEMMReshaped
Change-Id: I89403b97503fbb99f6a32f5d62b8c535ab26a7be
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/877
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/core/CL/CLHelpers.h b/arm_compute/core/CL/CLHelpers.h
index 78427c3..16fe09f 100644
--- a/arm_compute/core/CL/CLHelpers.h
+++ b/arm_compute/core/CL/CLHelpers.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -144,5 +144,14 @@
* @return preferred vector width
*/
size_t preferred_vector_width(const cl::Device &device, DataType dt);
+
+/** Helper function to check if "dummy work-items" are preferred to have a power of two NDRange
+ * In case dummy work-items is enabled, it is OpenCL kernel responsibility to check if the work-item is out-of range or not
+ *
+ * @param[in] device A CL device
+ *
+ * @return True if dummy work-items should be preferred to dispatch the NDRange
+ */
+bool preferred_dummy_work_items_support(const cl::Device &device);
}
#endif /* __ARM_COMPUTE_CLHELPERS_H__ */
diff --git a/arm_compute/core/CL/ICLKernel.h b/arm_compute/core/CL/ICLKernel.h
index f542358..590f892 100644
--- a/arm_compute/core/CL/ICLKernel.h
+++ b/arm_compute/core/CL/ICLKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -308,14 +308,16 @@
*
* @note If kernel->kernel() is empty then the function will return without adding anything to the queue.
*
- * @param[in,out] queue OpenCL command queue.
- * @param[in] kernel Kernel to enqueue
- * @param[in] window Window the kernel has to process.
- * @param[in] lws_hint Local workgroup size requested. Default is based on the device target.
+ * @param[in,out] queue OpenCL command queue.
+ * @param[in] kernel Kernel to enqueue
+ * @param[in] window Window the kernel has to process.
+ * @param[in] lws_hint (Optional) Local workgroup size requested. Default is based on the device target.
+ * @param[in] use_dummy_work_items (Optional) Use dummy work items in order to have two dimensional power of two NDRange. Default is false
+ * Note: it is kernel responsibility to check if the work-item is out-of-range
*
* @note If any dimension of the lws is greater than the global workgroup size then no lws will be passed.
*/
-void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint = CLKernelLibrary::get().default_ndrange());
+void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint = CLKernelLibrary::get().default_ndrange(), bool use_dummy_work_items = false);
/** Add the passed array's parameters to the object's kernel's arguments starting from the index idx.
*
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h
index 1cf7236..d361236 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h
@@ -81,6 +81,7 @@
bool _slide_matrix_b;
bool _reinterpret_output_as_3d;
unsigned int _k;
+ bool _use_dummy_work_items;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
index cb23b96..b0d245f 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
@@ -84,6 +84,7 @@
bool _slide_matrix_b;
bool _reinterpret_output_as_3d;
unsigned int _k;
+ bool _use_dummy_work_items;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h
index 7471594..b3ee435 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h
@@ -84,6 +84,7 @@
bool _slide_matrix_b;
bool _reinterpret_input_as_3d;
bool _reinterpret_output_as_3d;
+ bool _use_dummy_work_items;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h
index 91d85be..c7c7110 100644
--- a/arm_compute/core/Helpers.h
+++ b/arm_compute/core/Helpers.h
@@ -756,6 +756,34 @@
{
return x >= 0 ? x % m : (x % m + m) % m;
}
+
+/** Given an integer value, this function returns the next power of two
+ *
+ * @param[in] x Input value
+ *
+ * @return the next power of two
+ */
+inline unsigned int get_next_power_two(unsigned int x)
+{
+ // Decrement by 1
+ x--;
+
+ // Shift right by 1
+ x |= x >> 1u;
+ // Shift right by 2
+ x |= x >> 2u;
+ // Shift right by 4
+ x |= x >> 4u;
+ // Shift right by 8
+ x |= x >> 8u;
+ // Shift right by 16
+ x |= x >> 16u;
+
+ // Increment by 1
+ x++;
+
+ return x;
+}
} // namespace arm_compute
#include "arm_compute/core/Helpers.inl"