Kaynağa Gözat

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 yıl önce
ebeveyn
işleme
59d5f426c5
1 değiştirilmiş dosya ile 5 ekleme ve 0 silme
  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);
         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);
+        clFinish(queue);
 	/* hack to avoid third party libs to rebind threads */
 	_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);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	gettimeofday(&end, NULL);
 	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);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	gettimeofday(&end, NULL);
 	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);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	gettimeofday(&end, NULL);
 	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);
                 if (STARPU_UNLIKELY(err != CL_SUCCESS)) STARPU_OPENCL_REPORT_ERROR(err);
+                clFinish(queue);
 	}
 	gettimeofday(&end, NULL);
 	timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));