浏览代码

Fix OpenCL bandwidth estimation.

Force synchronization with clFinish because clEnqueueWriteBuffer in blocking mode does not ensure that the transfer has been performed.
Sylvain Henry 12 年之前
父节点
当前提交
59d5f426c5
共有 1 个文件被更改,包括 5 次插入0 次删除
  1. 5 0
      src/core/perfmodel/perfmodel_bus.c

+ 5 - 0
src/core/perfmodel/perfmodel_bus.c

@@ -361,6 +361,7 @@ static void measure_bandwidth_between_host_and_dev_on_cpu_with_opencl(int dev, i
 	memset(h_buffer, 0, size);
 	memset(h_buffer, 0, size);
         err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
         err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
         if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
         if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+        clFinish(queue);
 	/* hack to avoid third party libs to rebind threads */
 	/* hack to avoid third party libs to rebind threads */
 	_starpu_bind_thread_on_cpu(config, cpu);
 	_starpu_bind_thread_on_cpu(config, cpu);
 
 
@@ -375,6 +376,7 @@ static void measure_bandwidth_between_host_and_dev_on_cpu_with_opencl(int dev, i
 	{
 	{
                 err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
                 err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	}
 	gettimeofday(&end, NULL);
 	gettimeofday(&end, NULL);
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
@@ -387,6 +389,7 @@ static void measure_bandwidth_between_host_and_dev_on_cpu_with_opencl(int dev, i
 	{
 	{
                 err = clEnqueueReadBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
                 err = clEnqueueReadBuffer(queue, d_buffer, CL_TRUE, 0, size, h_buffer, 0, NULL, NULL);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	}
 	gettimeofday(&end, NULL);
 	gettimeofday(&end, NULL);
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
@@ -399,6 +402,7 @@ static void measure_bandwidth_between_host_and_dev_on_cpu_with_opencl(int dev, i
 	{
 	{
 		err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, 1, h_buffer, 0, NULL, NULL);
 		err = clEnqueueWriteBuffer(queue, d_buffer, CL_TRUE, 0, 1, h_buffer, 0, NULL, NULL);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	}
 	gettimeofday(&end, NULL);
 	gettimeofday(&end, NULL);
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
@@ -411,6 +415,7 @@ static void measure_bandwidth_between_host_and_dev_on_cpu_with_opencl(int dev, i
 	{
 	{
 		err = clEnqueueReadBuffer(queue, d_buffer, CL_TRUE, 0, 1, h_buffer, 0, NULL, NULL);
 		err = clEnqueueReadBuffer(queue, d_buffer, CL_TRUE, 0, 1, h_buffer, 0, NULL, NULL);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	}
 	gettimeofday(&end, NULL);
 	gettimeofday(&end, NULL);
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));