Add Mali™-G720 and Mali™-G620 as GpuTargets

This patch adds adds the latest Gpus as Gpu Target and sets up kernel selection heuristics for MatMul to address some nightly issues.

Resolves: COMPMID-6766

Change-Id: I29dbb08c5ecfb3fcd63230b0b1675ab557074aca
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10902
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 24 insertions, 65 deletions

diff --git a/src/gpu/cl/kernels/gemm/native/ClGemmNativeKernelConfig.h b/src/gpu/cl/kernels/gemm/native/ClGemmNativeKernelConfig.h
index 955bb3c01a..22aa1e2034 100644
--- a/src/gpu/cl/kernels/gemm/native/ClGemmNativeKernelConfig.h
+++ b/src/gpu/cl/kernels/gemm/native/ClGemmNativeKernelConfig.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CL_GEMM_NATIVE_KERNEL_CONFIGURATION_H
-#define ARM_COMPUTE_CL_GEMM_NATIVE_KERNEL_CONFIGURATION_H
+#ifndef ACL_SRC_GPU_CL_KERNELS_GEMM_NATIVE_CLGEMMNATIVEKERNELCONFIG_H
+#define ACL_SRC_GPU_CL_KERNELS_GEMM_NATIVE_CLGEMMNATIVEKERNELCONFIG_H
 
 #include "src/gpu/cl/kernels/gemm/IClGemmKernelConfig.h"
 #include "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.h"
@@ -58,6 +58,7 @@ public:
             case GPUTarget::BIFROST:
                 return std::make_unique<ClGemmDefaultConfigNativeBifrost>(gpu);
             case GPUTarget::VALHALL:
+            case GPUTarget::FIFTHGEN:
                 return std::make_unique<ClGemmDefaultConfigNativeValhall>(gpu);
             default:
                 ARM_COMPUTE_ERROR("Not supported GPU target");
@@ -68,4 +69,4 @@ public:
 } // namespace kernels
 } // namespace opencl
 } // namespace arm_compute
-#endif /*ARM_COMPUTE_CL_GEMM_NATIVE_KERNEL_CONFIGURATION_H */
+#endif // ACL_SRC_GPU_CL_KERNELS_GEMM_NATIVE_CLGEMMNATIVEKERNELCONFIG_H
diff --git a/src/gpu/cl/kernels/gemm/reshaped/ClGemmReshapedKernelConfig.h b/src/gpu/cl/kernels/gemm/reshaped/ClGemmReshapedKernelConfig.h
index 83928b3f4f..6327ee3027 100644
--- a/src/gpu/cl/kernels/gemm/reshaped/ClGemmReshapedKernelConfig.h
+++ b/src/gpu/cl/kernels/gemm/reshaped/ClGemmReshapedKernelConfig.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CL_GEMM_RESHAPED_KERNEL_CONFIGURATION_H
-#define ARM_COMPUTE_CL_GEMM_RESHAPED_KERNEL_CONFIGURATION_H
+#ifndef ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_CLGEMMRESHAPEDKERNELCONFIG_H
+#define ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_CLGEMMRESHAPEDKERNELCONFIG_H
 
 #include "src/gpu/cl/kernels/gemm/IClGemmKernelConfig.h"
 #include "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedBifrost.h"
@@ -56,6 +56,7 @@ public:
             case GPUTarget::BIFROST:
                 return std::make_unique<ClGemmDefaultConfigReshapedBifrost>(gpu);
             case GPUTarget::VALHALL:
+            case GPUTarget::FIFTHGEN:
                 return std::make_unique<ClGemmDefaultConfigReshapedValhall>(gpu);
             default:
                 ARM_COMPUTE_ERROR("Not supported GPU target");
@@ -66,4 +67,4 @@ public:
 } // namespace kernels
 } // namespace opencl
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_GEMM_RESHAPED_KERNEL_CONFIGURATION_H */
+#endif // ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_CLGEMMRESHAPEDKERNELCONFIG_H
diff --git a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmReshapedOnlyRhsKernelConfig.h b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmReshapedOnlyRhsKernelConfig.h
index e07ad993ed..1f0c5c2d87 100644
--- a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmReshapedOnlyRhsKernelConfig.h
+++ b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmReshapedOnlyRhsKernelConfig.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CL_GEMM_RESHAPED_ONLY_RHS_KERNEL_CONFIGURATION_H
-#define ARM_COMPUTE_CL_GEMM_RESHAPED_ONLY_RHS_KERNEL_CONFIGURATION_H
+#ifndef ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_ONLY_RHS_CLGEMMRESHAPEDONLYRHSKERNELCONFIG_H
+#define ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_ONLY_RHS_CLGEMMRESHAPEDONLYRHSKERNELCONFIG_H
 
 #include "src/gpu/cl/kernels/gemm/IClGemmKernelConfig.h"
 #include "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyBifrost.h"
@@ -56,6 +56,7 @@ public:
             case GPUTarget::BIFROST:
                 return std::make_unique<ClGemmDefaultConfigReshapedRhsOnlyBifrost>(gpu);
             case GPUTarget::VALHALL:
+            case GPUTarget::FIFTHGEN:
                 return std::make_unique<ClGemmDefaultConfigReshapedRhsOnlyValhall>(gpu);
             default:
                 ARM_COMPUTE_ERROR("Not supported GPU target");
@@ -66,4 +67,4 @@ public:
 } // namespace kernels
 } // namespace opencl
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_GEMM_RESHAPED_ONLY_RHS_KERNEL_CONFIGURATION_H */
+#endif // ACL_SRC_GPU_CL_KERNELS_GEMM_RESHAPED_ONLY_RHS_CLGEMMRESHAPEDONLYRHSKERNELCONFIG_H
diff --git a/src/gpu/cl/operators/ClMatMul.cpp b/src/gpu/cl/operators/ClMatMul.cpp
index 43303001d0..28a2aa2540 100644
--- a/src/gpu/cl/operators/ClMatMul.cpp
+++ b/src/gpu/cl/operators/ClMatMul.cpp
@@ -34,6 +34,7 @@
 #include "src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h"
 #include "src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.h"
 #include "src/runtime/heuristics/matmul_native/ClMatMulNativeKernelConfig.h"
+#include "src/runtime/heuristics/matmul_native/ClMatMulNativeKernelVariant.h"
 #include "src/runtime/heuristics/matmul_native/IClMatMulNativeKernelConfig.h"
 
 using namespace arm_compute::cl_matmul;
@@ -42,57 +43,6 @@ namespace arm_compute
 {
 namespace opencl
 {
-namespace
-{
-enum class MatMulKernelType
-{
-    /** Native matrix multiplication for FP types */
-    NATIVE_FP,
-
-    /** Native matrix multiplication for quantized types */
-    NATIVE_QUANTIZED,
-
-    /** Native matrix multiplication using MMUL extension for FP types */
-    NATIVE_MMUL_FP,
-
-    /** Native matrix multiplication using MMUL extension for Quantized types */
-    NATIVE_MMUL_QUANTIZED
-};
-
-MatMulKernelType get_matmul_kernel(const ITensorInfo         *lhs,
-                                   const ITensorInfo         *rhs,
-                                   const MatMulInfo          &matmul_info,
-                                   const ActivationLayerInfo &act_info)
-{
-    ARM_COMPUTE_UNUSED(lhs, rhs, matmul_info, act_info);
-
-    const bool is_quantized      = is_data_type_quantized_asymmetric(lhs->data_type());
-    const bool is_mmul_supported = arm_matrix_multiply_supported(CLKernelLibrary::get().get_device());
-
-    const int k = matmul_info.adj_lhs() ? lhs->tensor_shape().y() : lhs->tensor_shape().x();
-
-    if (is_quantized)
-    {
-        // MMUL kernel works only when K is a multiple of 16
-        if (is_mmul_supported && !act_info.enabled() && k % 16 == 0)
-        {
-            return MatMulKernelType::NATIVE_MMUL_QUANTIZED;
-        }
-
-        return MatMulKernelType::NATIVE_QUANTIZED;
-    }
-    else
-    {
-        // MMUL kernel works only when K is a multiple of 4
-        if (is_mmul_supported && !act_info.enabled() && k % 4 == 0)
-        {
-            return MatMulKernelType::NATIVE_MMUL_FP;
-        }
-
-        return MatMulKernelType::NATIVE_FP;
-    }
-}
-} // namespace
 using namespace arm_compute::opencl::kernels;
 
 ClMatMul::ClMatMul()
@@ -117,7 +67,10 @@ Status ClMatMul::validate(const ITensorInfo         *lhs,
 
     const MatMulKernelInfo kernel_info = t->configure(lhs, rhs, matmul_info);
 
-    switch (get_matmul_kernel(lhs, rhs, matmul_info, act_info))
+    const auto             kernel_selector = ClMatMulNativeKernelVariantFactory::create(gpu_target);
+    const MatMulKernelType kernel_type     = kernel_selector->select_kernel(lhs, rhs, matmul_info, act_info);
+
+    switch (kernel_type)
     {
         case MatMulKernelType::NATIVE_FP:
             return ClMatMulNativeKernel::validate(lhs, rhs, nullptr /* bias */, dst, kernel_info, act_info);
@@ -149,7 +102,10 @@ void ClMatMul::configure(const CLCompileContext    &compile_context,
     const auto             kernel_config = ClMatMulNativeKernelConfigurationFactory::create(gpu_target);
     const MatMulKernelInfo kernel_info   = kernel_config->configure(lhs, rhs, matmul_info);
 
-    switch (get_matmul_kernel(lhs, rhs, matmul_info, act_info))
+    const auto             kernel_selector = ClMatMulNativeKernelVariantFactory::create(gpu_target);
+    const MatMulKernelType kernel_type     = kernel_selector->select_kernel(lhs, rhs, matmul_info, act_info);
+
+    switch (kernel_type)
     {
         case MatMulKernelType::NATIVE_FP:
         {