COMPMID-2097: Implement a heuristic to dispatch CLGEMMReshapedOnlyRHS kernel from CLGEMM
Change-Id: I4170a80647b02501aa669e2c0347ddc39888ee76
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/928
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/core/CL/ICLGEMMKernelConfiguration.h b/arm_compute/core/CL/ICLGEMMKernelConfiguration.h
new file mode 100644
index 0000000..2e6d495
--- /dev/null
+++ b/arm_compute/core/CL/ICLGEMMKernelConfiguration.h
@@ -0,0 +1,68 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_ICLGEMMKERNELCONFIGURATION_H__
+#define __ARM_COMPUTE_ICLGEMMKERNELCONFIGURATION_H__
+
+#include "arm_compute/core/GPUTarget.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+/** Basic interface for the GEMM kernel configuration */
+class ICLGEMMKernelConfiguration
+{
+public:
+ /** Constructor
+ *
+ * @param[in] arch GPU target
+ */
+ ICLGEMMKernelConfiguration(GPUTarget arch)
+ : _target(arch)
+ {
+ }
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ ICLGEMMKernelConfiguration(const ICLGEMMKernelConfiguration &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ ICLGEMMKernelConfiguration &operator=(const ICLGEMMKernelConfiguration &) = delete;
+ /** Default Move Constructor. */
+ ICLGEMMKernelConfiguration(ICLGEMMKernelConfiguration &&) = default;
+ /** Default move assignment operator */
+ ICLGEMMKernelConfiguration &operator=(ICLGEMMKernelConfiguration &&) = default;
+ /** Virtual destructor */
+ virtual ~ICLGEMMKernelConfiguration() = default;
+ /** Given M, N, K and B, this method returns the @ref GEMMLHSMatrixInfo and @ref GEMMRHSMatrixInfo to be used
+ *
+ * @param[in] m Number of rows LHS matrix
+ * @param[in] n Number of columns RHS matrix
+ * @param[in] k Number of columns LHS matrix or number of rows RHS matrix
+ * @param[in] b Batch size
+ * @param[in] data_type Data type
+ */
+ virtual std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type) = 0;
+
+protected:
+ GPUTarget _target;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_ICLGEMMKERNELCONFIGURATION_H__ */
diff --git a/arm_compute/core/CL/gemm/CLGEMMHelpers.h b/arm_compute/core/CL/gemm/CLGEMMHelpers.h
new file mode 100644
index 0000000..d263712
--- /dev/null
+++ b/arm_compute/core/CL/gemm/CLGEMMHelpers.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMHELPERS_H__
+#define __ARM_COMPUTE_CLGEMMHELPERS_H__
+
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+namespace cl_gemm
+{
+/** Configure @ref GEMMLHSMatrixInfo and @ref GEMMRHSMatrixInfo
+ *
+ * @param[in] m Number of rows (M) in the LHS matrix not reshaped
+ * @param[in] n Number of columns (N) in the RHS matrix not reshaped
+ * @param[in] m0 Number of rows processed by each thread/work-item
+ * @param[in] n0 Number of columns processed by each thread/work-item
+ * @param[in] k0 Number of inner accumulation performed by each thread/work-item
+ * @param[in] v0 Number of vertical blocks of size (m0xk0) stored on the same output row
+ * @param[in] h0 Number of horizontal blocks of size (k0xn0) stored on the same output row
+ * @param[in] lhs_interleave True if the v0 (m0xk0) blocks have to be interleaved in the output row
+ * @param[in] rhs_interleave True if the h0 (k0xn0) blocks have to be interleaved in the output row
+ * @param[in] lhs_transpose True if the (m0xk0) block has to be transposed before been stored
+ * @param[in] rhs_transpose True if the (k0xn0) block has to be transposed before been stored
+ *
+ * @return @ref GEMMLHSMatrixInfo and @ref GEMMRHSMatrixInfo
+ */
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
+ bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose);
+} // namespace cl_gemm
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMHELPERS_H__ */
diff --git a/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h b/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h
new file mode 100644
index 0000000..105a58a
--- /dev/null
+++ b/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATION_H__
+#define __ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATION_H__
+
+#include "arm_compute/core/CL/ICLGEMMKernelConfiguration.h"
+#include "arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+namespace cl_gemm
+{
+/** CLGEMMReshaped factory class */
+class CLGEMMReshapedKernelConfigurationFactory final
+{
+public:
+ /** Static method to call the CLGEMMReshaped kernel configuration class accordingly with the GPU architecture
+ *
+ * @param[in] arch GPU target
+ *
+ * @return CLGEMMReshaped kernel configuration class
+ */
+ static std::unique_ptr<ICLGEMMKernelConfiguration> create(GPUTarget arch)
+ {
+ switch(get_arch_from_target(arch))
+ {
+ case GPUTarget::BIFROST:
+ return support::cpp14::make_unique<CLGEMMReshapedKernelConfigurationBifrost>(arch);
+ default:
+ return nullptr;
+ }
+ }
+};
+} // namespace cl_gemm
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATION_H__ */
diff --git a/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.h b/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.h
new file mode 100644
index 0000000..a0aae19
--- /dev/null
+++ b/arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATIONBIFROST_H__
+#define __ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATIONBIFROST_H__
+
+#include "arm_compute/core/CL/ICLGEMMKernelConfiguration.h"
+
+namespace arm_compute
+{
+namespace cl_gemm
+{
+/** Bifrost based OpenCL GEMMReshaped configuration */
+class CLGEMMReshapedKernelConfigurationBifrost final : public ICLGEMMKernelConfiguration
+{
+public:
+ /** Constructor
+ *
+ * @param[in] arch GPU target
+ */
+ CLGEMMReshapedKernelConfigurationBifrost(GPUTarget arch);
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLGEMMReshapedKernelConfigurationBifrost(const CLGEMMReshapedKernelConfigurationBifrost &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLGEMMReshapedKernelConfigurationBifrost &operator=(const CLGEMMReshapedKernelConfigurationBifrost &) = delete;
+ /** Default Move Constructor. */
+ CLGEMMReshapedKernelConfigurationBifrost(CLGEMMReshapedKernelConfigurationBifrost &&) = default;
+ /** Default move assignment operator */
+ CLGEMMReshapedKernelConfigurationBifrost &operator=(CLGEMMReshapedKernelConfigurationBifrost &&) = default;
+
+ // Inherited overridden method
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type) override;
+
+private:
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+};
+} // namespace cl_gemm
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMRESHAPEDKERNELCONFIGURATIONBIFROST_H__ */
diff --git a/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h b/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h
new file mode 100644
index 0000000..b9bf150
--- /dev/null
+++ b/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATION_H__
+#define __ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATION_H__
+
+#include "arm_compute/core/CL/ICLGEMMKernelConfiguration.h"
+#include "arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+namespace cl_gemm
+{
+/** CLGEMMReshapedOnlyRHS factory class */
+class CLGEMMReshapedOnlyRHSKernelConfigurationFactory final
+{
+public:
+ /** Static method to call the CLGEMMReshapedOnlyRHS kernel configuration class accordingly with the GPU architecture
+ *
+ * @param[in] arch GPU target
+ *
+ * @return CLGEMMReshapedOnlyRHS kernel configuration class
+ */
+ static std::unique_ptr<ICLGEMMKernelConfiguration> create(GPUTarget arch)
+ {
+ switch(get_arch_from_target(arch))
+ {
+ case GPUTarget::BIFROST:
+ return support::cpp14::make_unique<CLGEMMReshapedOnlyRHSKernelConfigurationBifrost>(arch);
+ default:
+ return nullptr;
+ }
+ }
+};
+} // namespace cl_gemm
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATION_H__ */
diff --git a/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.h b/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.h
new file mode 100644
index 0000000..3bed118
--- /dev/null
+++ b/arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATIONBIFROST_H__
+#define __ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATIONBIFROST_H__
+
+#include "arm_compute/core/CL/ICLGEMMKernelConfiguration.h"
+
+namespace arm_compute
+{
+namespace cl_gemm
+{
+/** Bifrost based OpenCL GEMMReshapedOnlyRHS configuration */
+class CLGEMMReshapedOnlyRHSKernelConfigurationBifrost final : public ICLGEMMKernelConfiguration
+{
+public:
+ /** Constructor
+ *
+ * @param[in] arch GPU target
+ */
+ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost(GPUTarget arch);
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost(const CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &operator=(const CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &) = delete;
+ /** Default Move Constructor. */
+ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost(CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &&) = default;
+ /** Default move assignment operator */
+ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &operator=(CLGEMMReshapedOnlyRHSKernelConfigurationBifrost &&) = default;
+
+ // Inherited overridden method
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type) override;
+
+private:
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+};
+} // namespace cl_gemm
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMRESHAPEDONLYRHSKERNELCONFIGURATIONBIFROST_H__ */