Add example floating point calculation

.gitignore

@@ -1,3 +1,4 @@
 *.o
 build
+*.user.*
 *.user

CMakeLists.txt

@@ -12,5 +12,5 @@ message(STATUS "OpenCL libraries: ${OPENCL_LIBRARIES}")
 add_executable(${PROJECT_NAME} "main.c")
 configure_file("saxpy.cl" "cl_kernels/saxpy.cl" COPYONLY)
 
-target_link_libraries(${PROJECT_NAME} ${OPENCL_LIBRARIES})
+target_link_libraries(${PROJECT_NAME} m ${OPENCL_LIBRARIES})
 include_directories(${PROJECT_NAME} ${OPENCL_INCLUDE_DIR})

main.c

@@ -3,7 +3,7 @@
 #include <stdlib.h>
 #include <stdint.h>
 #include <string.h>
-
+#include <math.h>
 
 void check_error(cl_int error)
 {
@@ -42,23 +42,43 @@ cl_program create_program(char *src, cl_context context)
 	return program;
 }
 
-float test_a[] = {21.321, 213.213, 3213.23};
+float *test_a, *test_b;
-float test_b[] = {21.21, 13.213, 17.2553};
 
+static void prepare_data()
+{
+	int i;
+	test_a = malloc(sizeof(float)*1000000000);
+	test_b = malloc(sizeof(float)*1000000000);
+
+	for (i = 0; i < 1000000000; i++) {
+		test_a[i] = (float)i;
+		test_b[i] = 0.0f;
+	}
+
+}
+
+float mult = 1.0f;
 int main(void)
 {
-
+	int i;
 	cl_int error;
+	cl_int ret;
 	cl_platform_id *platform_ids = NULL;
 	cl_device_id *device_ids = NULL;
 	cl_uint platform_id_count = 0;
 	cl_uint device_id_count = 0;
-	cl_context_properties context_properties[4];
 	cl_context context;
 	cl_program program;
 	cl_kernel kernel;
+	cl_command_queue queue;
 	cl_mem a_buff, b_buff;
 	char *source_code = NULL;
+	char temp_buff[100];
+	size_t data_size = 1000000000;
+	size_t work_size = 64;
+
+
+	prepare_data();
 
 	clGetPlatformIDs(0, NULL, &platform_id_count);
 	if (platform_id_count == 0)
@@ -73,6 +93,11 @@ int main(void)
 
 	printf("Platforms available: %u\n", (unsigned int)platform_id_count);
 
+	for (i = 0; i < platform_id_count; i++) {
+		temp_buff[0] = '\0';
+		clGetPlatformInfo(platform_ids[i], CL_PLATFORM_NAME, sizeof(temp_buff), temp_buff, NULL);
+		printf("Platform ID %d: %s\n", i, temp_buff);
+	}
 
 	clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_ALL, 0, NULL,
 			&device_id_count);
@@ -88,16 +113,32 @@ int main(void)
 	clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_ALL, device_id_count,
 			device_ids, NULL);
 
-	context_properties[0] = CL_CONTEXT_PLATFORM;
+	for (i = 0; i < device_id_count; i++) {
-	context_properties[1] = (cl_context_properties)platform_ids[0];
+		temp_buff[0] = '\0';
-	context_properties[2] = context_properties[3] = 0;
+		clGetDeviceInfo(device_ids[i], CL_DEVICE_NAME, sizeof(temp_buff), temp_buff, NULL);
+		printf("Device ID %d: %s\n", i, temp_buff);
+	}
 
 
-	context = clCreateContext(context_properties, device_id_count,
+	context = clCreateContext(NULL, 1, &device_ids[0], NULL, NULL, &error);
-				   device_ids, NULL,
-				   NULL, &error);
 	check_error(error);
 
+	queue = clCreateCommandQueue(context, device_ids[0], 0, &error);
+
+
+	/* Create memory buffers for data */
+	a_buff = clCreateBuffer(context, CL_MEM_READ_ONLY |
+				CL_MEM_COPY_HOST_PTR,
+				sizeof(test_a[0])*data_size, test_a, &error);
+	check_error(error);
+
+	b_buff = clCreateBuffer(context, CL_MEM_READ_WRITE |
+				CL_MEM_COPY_HOST_PTR,
+				sizeof(test_b[0])*data_size, test_b, &error);
+	check_error(error);
+
+
+
 	load_kernel_from_file("cl_kernels/saxpy.cl", &source_code);
 	program = create_program(source_code, context);
 
@@ -108,17 +149,28 @@ int main(void)
 
 	kernel = clCreateKernel(program, "SAXPY", &error);
 
+	/* Setting the arguments */
+	clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&a_buff);
+	clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&b_buff);
+	clSetKernelArg(kernel, 2, sizeof(float), (void *)&mult);
 
-	/* Create memory buffers for data */
+	ret = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &data_size, &work_size, 0, NULL, NULL);
-	a_buff = clCreateBuffer(context, CL_MEM_READ_ONLY |
+	check_error(ret);
-				CL_MEM_COPY_HOST_PTR,
-				sizeof(test_a), test_a, &error);
-	check_error(error);
 
-	b_buff = clCreateBuffer(context, CL_MEM_READ_WRITE |
+	clEnqueueReadBuffer(queue, b_buff, CL_TRUE, 0, 1000000000*sizeof(float), test_b, 0, NULL, NULL);
-				CL_MEM_COPY_HOST_PTR,
+
-				sizeof(test_b), test_b, &error);
+	ret = clFlush(queue);
-	check_error(error);
+	ret = clFinish(queue);
+	ret = clReleaseKernel(kernel);
+	ret = clReleaseProgram(program);
+	ret = clReleaseMemObject(a_buff);
+	ret = clReleaseMemObject(b_buff);
+	ret = clReleaseCommandQueue(queue);
+	ret = clReleaseContext(context);
+	free(test_a);
+	free(test_b);
+	free(platform_ids);
+	free(device_ids);
 
 	return 0;
 }