Add support for CLVK

This patch enables CLVK through the graph API and inside the
CLScheduler. By default the Native platform is selected.
Selecting CLVK can be done via --target=clvk.

Resolves COMPMID-4205 and COMPMID-4206

Change-Id: Ic60744980c6b8a60e776627ea677ed46be88f656
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5475
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

7 files changed, 83 insertions, 16 deletions

diff --git a/src/core/CL/OpenCL.cpp b/src/core/CL/OpenCL.cpp
index dd9960a7b3..a7be534397 100644
--- a/src/core/CL/OpenCL.cpp
+++ b/src/core/CL/OpenCL.cpp
@@ -122,6 +122,7 @@ bool CLSymbols::load(const std::string &library)
     LOAD_FUNCTION_PTR(clGetDeviceIDs, handle);
     LOAD_FUNCTION_PTR(clGetMemObjectInfo, handle);
     LOAD_FUNCTION_PTR(clRetainEvent, handle);
+    LOAD_FUNCTION_PTR(clGetPlatformInfo, handle);
     LOAD_FUNCTION_PTR(clGetPlatformIDs, handle);
     LOAD_FUNCTION_PTR(clGetKernelWorkGroupInfo, handle);
     LOAD_FUNCTION_PTR(clGetCommandQueueInfo, handle);
@@ -865,6 +866,24 @@ cl_int clRetainEvent(cl_event event)
     }
 }
 
+cl_int clGetPlatformInfo(cl_platform_id   platform,
+                         cl_platform_info param_name,
+                         size_t           param_value_size,
+                         void            *param_value,
+                         size_t          *param_value_size_ret)
+{
+    arm_compute::CLSymbols::get().load_default();
+    auto func = arm_compute::CLSymbols::get().clGetPlatformInfo_ptr;
+    if(func != nullptr)
+    {
+        return func(platform, param_name, param_value_size, param_value, param_value_size_ret);
+    }
+    else
+    {
+        return CL_OUT_OF_RESOURCES;
+    }
+}
+
 cl_int clGetPlatformIDs(cl_uint num_entries, cl_platform_id *platforms, cl_uint *num_platforms)
 {
     arm_compute::CLSymbols::get().load_default();
diff --git a/src/graph/GraphManager.cpp b/src/graph/GraphManager.cpp
index 9d53172dc8..ab7aac6230 100644
--- a/src/graph/GraphManager.cpp
+++ b/src/graph/GraphManager.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -57,6 +57,18 @@ void GraphManager::finalize_graph(Graph &graph, GraphContext &ctx, PassManager &
     // Force target to all graph construct
     // TODO (COMPMID-2014) : Support heterogeneous execution
     Target forced_target = target;
+
+    // In case CLVK is selected, use the CL backend and
+    // update config
+    if(target == Target::CLVK)
+    {
+        forced_target       = Target::CL;
+        GraphConfig config  = ctx.config();
+        config.backend_type = CLBackendType::Clvk;
+
+        ctx.set_config(config);
+    }
+
     if(!is_target_supported(target))
     {
         forced_target = get_default_target();
diff --git a/src/graph/TypeLoader.cpp b/src/graph/TypeLoader.cpp
index 8efd66f2c5..3c51289dba 100644
--- a/src/graph/TypeLoader.cpp
+++ b/src/graph/TypeLoader.cpp
@@ -59,6 +59,7 @@ Target target_from_name(const std::string &name)
     {
         { "neon", Target::NEON },
         { "cl", Target::CL },
+        { "clvk", Target::CLVK },
     };
 
 #ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED
diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp
index f8e22ca7a0..eafda98669 100644
--- a/src/graph/backends/CL/CLDeviceBackend.cpp
+++ b/src/graph/backends/CL/CLDeviceBackend.cpp
@@ -65,7 +65,7 @@ bool file_exists(const std::string &filename)
 static detail::BackendRegistrar<CLDeviceBackend> CLDeviceBackend_registrar(Target::CL);
 
 CLDeviceBackend::CLDeviceBackend()
-    : _context_count(0), _tuner(), _gemm_heuristics(), _allocator(nullptr), _tuner_file()
+    : _context_count(0), _tuner(), _gemm_heuristics(), _allocator(nullptr), _tuner_file(), _backend_type(CLBackendType::Native)
 {
 }
 
@@ -87,7 +87,7 @@ void CLDeviceBackend::set_kernel_tuning_mode(CLTunerMode tuning_mode)
 void CLDeviceBackend::initialize_backend()
 {
     // Setup Scheduler
-    CLScheduler::get().default_init(&_tuner, &_gemm_heuristics);
+    CLScheduler::get().default_init(&_tuner, &_gemm_heuristics, _backend_type);
     // Create allocator with new context
     _allocator = std::make_unique<CLBufferAllocator>(nullptr /* legacy path for CLCoreRuntimeContext */);
 }
@@ -108,6 +108,7 @@ void CLDeviceBackend::setup_backend_context(GraphContext &ctx)
     _context_count++;
     if(_context_count == 1)
     {
+        _backend_type = ctx.config().backend_type;
         initialize_backend();
     }
 
diff --git a/src/runtime/CL/CLHelpers.cpp b/src/runtime/CL/CLHelpers.cpp
index 5f1842f76d..5b4bbbcde0 100644
--- a/src/runtime/CL/CLHelpers.cpp
+++ b/src/runtime/CL/CLHelpers.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -85,14 +85,48 @@ void initialise_context_properties(const cl::Platform &platform, const cl::Devic
 
 namespace arm_compute
 {
-std::tuple<cl::Context, cl::Device, cl_int>
-create_opencl_context_and_device()
+cl::Platform select_preferable_platform(CLBackendType cl_backend_type)
 {
-    ARM_COMPUTE_ERROR_ON(!opencl_is_available());
     std::vector<cl::Platform> platforms;
     cl::Platform::get(&platforms);
     ARM_COMPUTE_ERROR_ON_MSG(platforms.size() == 0, "Couldn't find any OpenCL platform");
-    cl::Platform            p = platforms[0];
+
+    cl::Platform selected_platform{ nullptr };
+
+    // If the user has selected the Native platform, return the first available.
+    switch(cl_backend_type)
+    {
+        case CLBackendType::Native:
+            selected_platform = platforms[0];
+            break;
+        case CLBackendType::Clvk:
+            for(auto p : platforms)
+            {
+                std::string res = p.getInfo<CL_PLATFORM_NAME>();
+                if(res.find("clvk") != std::string::npos)
+                {
+                    selected_platform = p;
+                    break;
+                }
+            }
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported backend type");
+    }
+
+    if(!selected_platform())
+    {
+        ARM_COMPUTE_ERROR("No valid platform found");
+    }
+
+    return selected_platform;
+}
+
+std::tuple<cl::Context, cl::Device, cl_int>
+create_opencl_context_and_device(CLBackendType cl_backend_type)
+{
+    ARM_COMPUTE_ERROR_ON(!opencl_is_available());
+    cl::Platform            p = select_preferable_platform(cl_backend_type);
     cl::Device              device;
     std::vector<cl::Device> platform_devices;
     p.getDevices(CL_DEVICE_TYPE_DEFAULT, &platform_devices);
diff --git a/src/runtime/CL/CLRuntimeContext.cpp b/src/runtime/CL/CLRuntimeContext.cpp
index 9d46126ee4..0c1d011f9a 100644
--- a/src/runtime/CL/CLRuntimeContext.cpp
+++ b/src/runtime/CL/CLRuntimeContext.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,10 +29,10 @@
 namespace arm_compute
 {
 CLRuntimeContext::CLRuntimeContext()
-    : _gpu_owned_scheduler(std::make_unique<CLScheduler>()), _gpu_scheduler(_gpu_owned_scheduler.get()), _symbols(), _core_context()
+    : _gpu_owned_scheduler(std::make_unique<CLScheduler>()), _gpu_scheduler(_gpu_owned_scheduler.get()), _symbols(), _core_context(), _backend_type()
 {
     _symbols.load_default();
-    auto ctx_dev_err = create_opencl_context_and_device();
+    auto ctx_dev_err = create_opencl_context_and_device(_backend_type);
     ARM_COMPUTE_ERROR_ON_MSG(std::get<2>(ctx_dev_err) != CL_SUCCESS, "Failed to create OpenCL context");
     auto             ctx   = std::get<0>(ctx_dev_err);
     auto             dev   = std::get<1>(ctx_dev_err);
diff --git a/src/runtime/CL/CLScheduler.cpp b/src/runtime/CL/CLScheduler.cpp
index f228cf6513..cb5f04ce8b 100644
--- a/src/runtime/CL/CLScheduler.cpp
+++ b/src/runtime/CL/CLScheduler.cpp
@@ -24,7 +24,6 @@
 #include "arm_compute/runtime/CL/CLScheduler.h"
 
 #include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/runtime/CL/CLHelpers.h"
 #include "arm_compute/runtime/CL/CLTuner.h"
 #include "src/core/CL/ICLKernel.h"
 
@@ -96,7 +95,7 @@ bool CLScheduler::is_initialised() const
 std::once_flag CLScheduler::_initialize_symbols;
 
 CLScheduler::CLScheduler()
-    : _context(), _queue(), _target(GPUTarget::MIDGARD), _is_initialised(false), _cl_tuner(nullptr), _gemm_heuristics(nullptr)
+    : _context(), _queue(), _target(GPUTarget::MIDGARD), _is_initialised(false), _cl_tuner(nullptr), _gemm_heuristics(nullptr), _backend_type(CLBackendType::Native)
 {
 }
 
@@ -119,14 +118,14 @@ void CLScheduler::default_init_with_context(cl::Device &device, cl::Context &ctx
     }
 }
 
-void CLScheduler::default_init(ICLTuner *cl_tuner, CLGEMMHeuristicsHandle *gemm_h)
+void CLScheduler::default_init(ICLTuner *cl_tuner, CLGEMMHeuristicsHandle *gemm_h, CLBackendType cl_backend_type)
 {
     if(!_is_initialised)
     {
         cl::Context ctx;
         cl::Device  dev;
         cl_int      err;
-        std::tie(ctx, dev, err) = create_opencl_context_and_device();
+        std::tie(ctx, dev, err) = create_opencl_context_and_device(cl_backend_type);
         ARM_COMPUTE_ERROR_ON_MSG(err != CL_SUCCESS, "Failed to create OpenCL context");
         cl::CommandQueue queue = cl::CommandQueue(ctx, dev);
         CLKernelLibrary::get().init("./cl_kernels/", ctx, dev);
@@ -143,7 +142,7 @@ void CLScheduler::set_context(cl::Context context)
     CLKernelLibrary::get().set_context(_context);
 }
 
-void CLScheduler::init(cl::Context context, cl::CommandQueue queue, const cl::Device &device, ICLTuner *cl_tuner, CLGEMMHeuristicsHandle *gemm_h)
+void CLScheduler::init(cl::Context context, cl::CommandQueue queue, const cl::Device &device, ICLTuner *cl_tuner, CLGEMMHeuristicsHandle *gemm_h, CLBackendType cl_backend_type)
 {
     set_context(std::move(context));
     _queue           = std::move(queue);
@@ -151,6 +150,7 @@ void CLScheduler::init(cl::Context context, cl::CommandQueue queue, const cl::De
     _is_initialised  = true;
     _cl_tuner        = cl_tuner;
     _gemm_heuristics = gemm_h;
+    _backend_type    = cl_backend_type;
 }
 
 void CLScheduler::enqueue_common(ICLKernel &kernel, ITensorPack &tensors, bool flush)