ker

4 files changed, 60 insertions, 1 deletions

diff --git a/backend/src/driver/cl_gen_driver.c b/backend/src/driver/cl_gen_driver.c
index 51da3228..1f44d03a 100644
--- a/backend/src/driver/cl_gen_driver.c
+++ b/backend/src/driver/cl_gen_driver.c
@@ -28,4 +28,6 @@ _cl_driver clgenDriver = {
   .build_program = GenBuildProgram,
   .get_program_kernel_names = GenGetProgramKernelNames,
   .release_program = GenReleaseProgram,
+  .create_kernel = GenCreateKernel,
+  .release_kernel = GenReleaseKernel,
 };
diff --git a/backend/src/driver/cl_gen_driver.h b/backend/src/driver/cl_gen_driver.h
index a5dca8ab..4dfe3786 100644
--- a/backend/src/driver/cl_gen_driver.h
+++ b/backend/src/driver/cl_gen_driver.h
@@ -39,6 +39,9 @@ cl_int GenBuildProgram(cl_program program, const cl_device_id device, const char
 cl_int GenGetProgramKernelNames(cl_program program, const cl_device_id device, char *names,
                                 cl_uint name_sz, cl_uint* ret_sz, cl_uint* ker_num);
 cl_int GenReleaseProgram(cl_program program, const cl_device_id device);
+cl_int GenCreateKernel(cl_kernel kernel, const cl_device_id device);
+cl_int GenReleaseKernel(cl_kernel kernel, const cl_device_id device);
+
 
 /*****************************************************************************/
 int dri2OpenX11(void);
diff --git a/backend/src/driver/cl_gen_kernel.cpp b/backend/src/driver/cl_gen_kernel.cpp
index 86e34c6b..5b0f57a7 100644
--- a/backend/src/driver/cl_gen_kernel.cpp
+++ b/backend/src/driver/cl_gen_kernel.cpp
@@ -23,6 +23,7 @@ extern "C" {  // for the C header files
 #include "cl_program.h"
 #include "cl_kernel.h"
 #include "cl_gen_driver.h"
+#include "cl_gen_devices.h"
 #ifdef __cplusplus
 }
 #endif /* __cplusplus */
@@ -33,6 +34,32 @@ extern "C" {  // for the C header files
 
 using namespace gbe;
 
+static size_t genGetKernelMaxWorkGroupSize(cl_kernel kernel, Kernel* ker, const cl_device_id device)
+{
+  size_t work_group_size, thread_cnt;
+  int simd_width = ker->getSIMDWidth();
+  GenGPUDevice* gpuDev = reinterpret_cast<GenGPUDevice*>(getGenDevicePrivate(device));
+  GBE_ASSERT(gpuDev);
+
+  if (!ker->getUseSLM()) {
+    if (!IS_BAYTRAIL_T(gpuDev->device_id) || simd_width == 16)
+      work_group_size = simd_width * 64;
+    else
+      work_group_size = device->max_compute_unit * gpuDev->max_thread_per_unit * simd_width;
+
+  } else {
+    thread_cnt = device->max_compute_unit *
+                 gpuDev->max_thread_per_unit / gpuDev->sub_slice_count;
+    if(thread_cnt > 64)
+      thread_cnt = 64;
+    work_group_size = thread_cnt * simd_width;
+  }
+
+  if(work_group_size > device->max_work_group_size)
+    work_group_size = device->max_work_group_size;
+  return work_group_size;
+}
+
 extern "C"
 cl_int GenCreateKernel(cl_kernel kernel, const cl_device_id device)
 {
@@ -46,11 +73,28 @@ cl_int GenCreateKernel(cl_kernel kernel, const cl_device_id device)
   if (ker == NULL)
     return CL_INVALID_KERNEL_NAME;
 
-  /* Set kernel parameters. If the parameters have alread been set, check and return
+  /* Set kernel args. If the arg have alread been set, check and return
      CL_INVALID_KERNEL_DEFINITION if failed. */
 
 
+  cl_kernel_workgroup_info wgInfo = &kernel->wg_info[findIndexByDevice(kernel->program->ctx, device)];
+  ker->getCompileWorkGroupSize(wgInfo->compile_wg_sz);
+  wgInfo->local_mem_sz = ker->getSLMSize();
+  wgInfo->private_mem_sz = ker->getStackSize();
+  wgInfo->work_group_sz = genGetKernelMaxWorkGroupSize(kernel, ker, device);
+
   setGenKernelPrivate(kernel, device, ker);
   return CL_SUCCESS;
 }
 
+extern "C"
+cl_int GenReleaseKernel(cl_kernel kernel, const cl_device_id device)
+{
+  Kernel* ker = reinterpret_cast<Kernel*>(getGenKernelPrivate(kernel, device));
+  if (ker == NULL)
+    return CL_INVALID_VALUE;
+
+  setGenKernelPrivate(kernel, device, NULL);
+  return CL_SUCCESS;
+}
+
diff --git a/libclapi/cl_kernel.c b/libclapi/cl_kernel.c
index 50875ab1..cea15ec2 100644
--- a/libclapi/cl_kernel.c
+++ b/libclapi/cl_kernel.c
@@ -251,3 +251,13 @@ error:
   return kernel;
 }
 
+cl_int
+clReleaseKernel(cl_kernel kernel)
+{
+  cl_int err = CL_SUCCESS;
+  CHECK_KERNEL(kernel);
+  cl_release_kernel(kernel);
+error:
+  return err;
+}
+