Use cl_mem for OpenCL buffer pointers, as that's what they actually are, and to avoid confusing beginners with CPU buffers.

doc/starpu.texi

@@ -732,7 +732,7 @@ void scal_cpu_func(void *buffers[], void *cl_arg)
 
     /* length of the vector */
     unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
-    /* local copy of the vector pointer */
+    /* CPU copy of the vector pointer */
     float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
 
     for (i = 0; i < n; i++)
@@ -811,7 +811,7 @@ extern "C" void scal_cuda_func(void *buffers[], void *_args)
 
     /* length of the vector */
     unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
-    /* local copy of the vector pointer */
+    /* CUDA copy of the vector pointer */
     float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
     unsigned threads_per_block = 64;
     unsigned nblocks = (n + threads_per_block-1) / threads_per_block;
@@ -861,8 +861,8 @@ void scal_opencl_func(void *buffers[], void *_args)
 
     /* length of the vector */
     unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
-    /* local copy of the vector pointer */
-    float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
+    /* OpenCL copy of the vector pointer */
+    cl_mem val = (cl_mem) STARPU_VECTOR_GET_PTR(buffers[0]);
 
 @i{    id = starpu_worker_get_id();}
 @i{    devid = starpu_worker_get_devid(id);}
@@ -871,7 +871,7 @@ void scal_opencl_func(void *buffers[], void *_args)
 @i{                    "vector_mult_opencl", devid);   /* @b{Name of the codelet defined above} */}
 @i{    if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);}
 
-@i{    err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);}
+@i{    err = clSetKernelArg(kernel, 0, sizeof(val), &val);}
 @i{    err |= clSetKernelArg(kernel, 1, sizeof(n), &n);}
 @i{    err |= clSetKernelArg(kernel, 2, sizeof(*factor), factor);}
 @i{    if (err) STARPU_OPENCL_REPORT_ERROR(err);}

doc/vector_scal_opencl.texi

@@ -13,8 +13,8 @@ void scal_opencl_func(void *buffers[], void *_args)
 
     /* length of the vector */
     unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
-    /* local copy of the vector pointer */
-    float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
+    /* OpenCL copy of the vector pointer */
+    cl_mem val = (cl_mem)STARPU_VECTOR_GET_PTR(buffers[0]);
 
     id = starpu_worker_get_id();
     devid = starpu_worker_get_devid(id);
@@ -23,7 +23,7 @@ void scal_opencl_func(void *buffers[], void *_args)
                                     devid);
     if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
-    err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+    err = clSetKernelArg(kernel, 0, sizeof(val), &val);
     err |= clSetKernelArg(kernel, 1, sizeof(n), &n);
     err |= clSetKernelArg(kernel, 2, sizeof(*factor), factor);
     if (err) STARPU_OPENCL_REPORT_ERROR(err);

examples/basic_examples/block_opencl.c

@@ -26,7 +26,7 @@ void opencl_codelet(void *descr[], void *_args)
 	cl_command_queue queue;
 	cl_event event;
 	int id, devid, err;
-	float *block = (float *)STARPU_BLOCK_GET_PTR(descr[0]);
+	cl_mem block = (cl_mem)STARPU_BLOCK_GET_PTR(descr[0]);
 	int nx = (int)STARPU_BLOCK_GET_NX(descr[0]);
 	int ny = (int)STARPU_BLOCK_GET_NY(descr[0]);
 	int nz = (int)STARPU_BLOCK_GET_NZ(descr[0]);
@@ -41,13 +41,13 @@ void opencl_codelet(void *descr[], void *_args)
         if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &block);
-	err = clSetKernelArg(kernel, 1, sizeof(int), &nx);
-	err = clSetKernelArg(kernel, 2, sizeof(int), &ny);
-	err = clSetKernelArg(kernel, 3, sizeof(int), &nz);
+	err = clSetKernelArg(kernel, 0, sizeof(block), &block);
+	err = clSetKernelArg(kernel, 1, sizeof(nx), &nx);
+	err = clSetKernelArg(kernel, 2, sizeof(ny), &ny);
+	err = clSetKernelArg(kernel, 3, sizeof(nz), &nz);
 	err = clSetKernelArg(kernel, 4, sizeof(ldy), &ldy);
 	err = clSetKernelArg(kernel, 5, sizeof(ldz), &ldz);
-	err = clSetKernelArg(kernel, 6, sizeof(float), multiplier);
+	err = clSetKernelArg(kernel, 6, sizeof(*multiplier), multiplier);
         if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{

examples/basic_examples/variable_kernels_opencl.c

@@ -21,7 +21,7 @@
 extern struct starpu_opencl_program opencl_program;
 void opencl_codelet(void *descr[], void *_args)
 {
-	float *val = (float *)STARPU_VARIABLE_GET_PTR(descr[0]);
+	cl_mem val = (cl_mem)STARPU_VARIABLE_GET_PTR(descr[0]);
 	cl_kernel kernel;
 	cl_command_queue queue;
 	cl_event event;
@@ -34,7 +34,7 @@ void opencl_codelet(void *descr[], void *_args)
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+	err = clSetKernelArg(kernel, 0, sizeof(val), &val);
 	if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{

examples/basic_examples/vector_scal_opencl.c

@@ -36,8 +36,8 @@ void scal_opencl_func(void *buffers[], void *_args)
 
 	/* length of the vector */
 	unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
-	/* local copy of the vector pointer */
-	float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
+	/* OpenCL copy of the vector pointer */
+	cl_mem val = (cl_mem)STARPU_VECTOR_GET_PTR(buffers[0]);
 
 	id = starpu_worker_get_id();
 	devid = starpu_worker_get_devid(id);
@@ -45,7 +45,7 @@ void scal_opencl_func(void *buffers[], void *_args)
 	err = starpu_opencl_load_kernel(&kernel, &queue, &opencl_program, "vector_mult_opencl", devid);
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+	err = clSetKernelArg(kernel, 0, sizeof(val), &val);
 	err |= clSetKernelArg(kernel, 1, sizeof(n), &n);
 	err |= clSetKernelArg(kernel, 2, sizeof(*factor), factor);
 	if (err) STARPU_OPENCL_REPORT_ERROR(err);

examples/filters/fblock_opencl.c

@@ -28,7 +28,7 @@ void opencl_func(void *buffers[], void *cl_arg)
 	cl_event event;
 
         int *factor = cl_arg;
-	int *block = (int *)STARPU_BLOCK_GET_PTR(buffers[0]);
+	cl_mem block = (cl_mem)STARPU_BLOCK_GET_PTR(buffers[0]);
 	int nx = (int)STARPU_BLOCK_GET_NX(buffers[0]);
 	int ny = (int)STARPU_BLOCK_GET_NY(buffers[0]);
 	int nz = (int)STARPU_BLOCK_GET_NZ(buffers[0]);
@@ -42,7 +42,7 @@ void opencl_func(void *buffers[], void *cl_arg)
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &block);
+	err = clSetKernelArg(kernel, 0, sizeof(block), &block);
 	err = clSetKernelArg(kernel, 1, sizeof(nx), &nx);
 	err = clSetKernelArg(kernel, 2, sizeof(ny), &ny);
 	err = clSetKernelArg(kernel, 3, sizeof(nz), &nz);

examples/incrementer/incrementer_kernels_opencl.c

@@ -21,7 +21,7 @@
 extern struct starpu_opencl_program opencl_program;
 void opencl_codelet(void *descr[], void *_args)
 {
-	float *val = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
+	cl_mem val = (cl_mem)STARPU_VECTOR_GET_PTR(descr[0]);
 	cl_kernel kernel;
 	cl_command_queue queue;
 	cl_event event;
@@ -34,7 +34,7 @@ void opencl_codelet(void *descr[], void *_args)
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+	err = clSetKernelArg(kernel, 0, sizeof(val), &val);
 	if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{

examples/mandelbrot/mandelbrot.c

@@ -247,15 +247,15 @@ static void compute_block_opencl(void *descr[], void *cl_arg)
 
 	starpu_opencl_load_kernel(&kernel, &queue, &opencl_programs, "mandelbrot_kernel", devid);
 
-	clSetKernelArg(kernel, 0, sizeof(cl_mem), &data);
-	clSetKernelArg(kernel, 1, sizeof(double), &leftX);
-	clSetKernelArg(kernel, 2, sizeof(double), &topY);
-	clSetKernelArg(kernel, 3, sizeof(double), &stepX);
-	clSetKernelArg(kernel, 4, sizeof(double), &stepY);
-	clSetKernelArg(kernel, 5, sizeof(int), &maxIt);
-	clSetKernelArg(kernel, 6, sizeof(int), &iby);
-	clSetKernelArg(kernel, 7, sizeof(int), &block_size);
-	clSetKernelArg(kernel, 8, sizeof(int), &width);
+	clSetKernelArg(kernel, 0, sizeof(data), &data);
+	clSetKernelArg(kernel, 1, sizeof(leftX), &leftX);
+	clSetKernelArg(kernel, 2, sizeof(topY), &topY);
+	clSetKernelArg(kernel, 3, sizeof(stepX), &stepX);
+	clSetKernelArg(kernel, 4, sizeof(stepY), &stepY);
+	clSetKernelArg(kernel, 5, sizeof(maxIt), &maxIt);
+	clSetKernelArg(kernel, 6, sizeof(iby), &iby);
+	clSetKernelArg(kernel, 7, sizeof(block_size), &block_size);
+	clSetKernelArg(kernel, 8, sizeof(width), &width);
 
 	unsigned dim = 16;
 	size_t local[2] = {dim, 1};

examples/matvecmult/matvecmult.c

@@ -29,9 +29,9 @@ void opencl_codelet(void *descr[], __attribute__ ((unused)) void *_args)
 	cl_kernel kernel;
 	cl_command_queue queue;
 	int id, devid, err, n;
-	float *matrix = (float *)STARPU_MATRIX_GET_PTR(descr[0]);
-	float *vector = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
-	float *mult = (float *)STARPU_VECTOR_GET_PTR(descr[2]);
+	cl_mem matrix = (cl_mem)STARPU_MATRIX_GET_PTR(descr[0]);
+	cl_mem vector = (cl_mem)STARPU_VECTOR_GET_PTR(descr[1]);
+	cl_mem mult = (cl_mem)STARPU_VECTOR_GET_PTR(descr[2]);
 	int nx = STARPU_MATRIX_GET_NX(descr[0]);
 	int ny = STARPU_MATRIX_GET_NY(descr[0]);
 	cl_event event;
@@ -43,11 +43,11 @@ void opencl_codelet(void *descr[], __attribute__ ((unused)) void *_args)
         if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
         n=0;
-        err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &matrix);
-        err |= clSetKernelArg(kernel, n++, sizeof(cl_mem), &vector);
-        err |= clSetKernelArg(kernel, n++, sizeof(int), (void*)&nx);
-        err |= clSetKernelArg(kernel, n++, sizeof(int), (void*)&ny);
-        err |= clSetKernelArg(kernel, n++, sizeof(cl_mem), &mult);
+        err = clSetKernelArg(kernel, n++, sizeof(matrix), &matrix);
+        err |= clSetKernelArg(kernel, n++, sizeof(vector), &vector);
+        err |= clSetKernelArg(kernel, n++, sizeof(nx), (void*)&nx);
+        err |= clSetKernelArg(kernel, n++, sizeof(ny), (void*)&ny);
+	err |= clSetKernelArg(kernel, n++, sizeof(mult), &mult);
         if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{

examples/spmv/spmv_kernels.c

@@ -30,15 +30,15 @@ void spmv_kernel_opencl(void *descr[], void *args)
 
 	uint32_t nnz = STARPU_CSR_GET_NNZ(descr[0]);
 	uint32_t nrow = STARPU_CSR_GET_NROW(descr[0]);
-	float *nzval = (float *)STARPU_CSR_GET_NZVAL(descr[0]);
+	cl_mem nzval = (cl_mem)STARPU_CSR_GET_NZVAL(descr[0]);
 	uint32_t *colind = STARPU_CSR_GET_COLIND(descr[0]);
 	uint32_t *rowptr = STARPU_CSR_GET_ROWPTR(descr[0]);
 	uint32_t firstentry = STARPU_CSR_GET_FIRSTENTRY(descr[0]);
 
-	float *vecin = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
+	cl_mem vecin = (cl_mem)STARPU_VECTOR_GET_PTR(descr[1]);
 	uint32_t nx_in = STARPU_VECTOR_GET_NX(descr[1]);
 
-	float *vecout = (float *)STARPU_VECTOR_GET_PTR(descr[2]);
+	cl_mem vecout = (cl_mem)STARPU_VECTOR_GET_PTR(descr[2]);
 	uint32_t nx_out = STARPU_VECTOR_GET_NX(descr[2]);
 
         id = starpu_worker_get_id();
@@ -49,16 +49,16 @@ void spmv_kernel_opencl(void *descr[], void *args)
 
 	err = 0;
         n=0;
-	err = clSetKernelArg(kernel, n++, sizeof(uint32_t), &nnz);
-	err = clSetKernelArg(kernel, n++, sizeof(uint32_t), &nrow);
-	err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &nzval);
-	err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &colind);
-	err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &rowptr);
-	err = clSetKernelArg(kernel, n++, sizeof(uint32_t), &firstentry);
-	err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &vecin);
-	err = clSetKernelArg(kernel, n++, sizeof(uint32_t), &nx_in);
-	err = clSetKernelArg(kernel, n++, sizeof(cl_mem), &vecout);
-	err = clSetKernelArg(kernel, n++, sizeof(uint32_t), &nx_out);
+	err = clSetKernelArg(kernel, n++, sizeof(nnz), &nnz);
+	err = clSetKernelArg(kernel, n++, sizeof(nrow), &nrow);
+	err = clSetKernelArg(kernel, n++, sizeof(nzval), &nzval);
+	err = clSetKernelArg(kernel, n++, sizeof(colind), &colind);
+	err = clSetKernelArg(kernel, n++, sizeof(rowptr), &rowptr);
+	err = clSetKernelArg(kernel, n++, sizeof(firstentry), &firstentry);
+	err = clSetKernelArg(kernel, n++, sizeof(vecin), &vecin);
+	err = clSetKernelArg(kernel, n++, sizeof(nx_in), &nx_in);
+	err = clSetKernelArg(kernel, n++, sizeof(vecout), &vecout);
+	err = clSetKernelArg(kernel, n++, sizeof(nx_out), &nx_out);
         if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{

tests/datawizard/sync_and_notify_data_opencl.c

@@ -22,7 +22,7 @@ extern struct starpu_opencl_program opencl_code;
 
 void opencl_codelet_incA(void *descr[], __attribute__ ((unused)) void *_args)
 {
-        unsigned *val = (unsigned *)STARPU_VECTOR_GET_PTR(descr[0]);
+	cl_mem val = (cl_mem)STARPU_VECTOR_GET_PTR(descr[0]);
 	cl_kernel kernel;
 	cl_command_queue queue;
 	cl_event event;
@@ -35,7 +35,7 @@ void opencl_codelet_incA(void *descr[], __attribute__ ((unused)) void *_args)
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+	err = clSetKernelArg(kernel, 0, sizeof(val), &val);
 	if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{
@@ -53,7 +53,7 @@ void opencl_codelet_incA(void *descr[], __attribute__ ((unused)) void *_args)
 
 void opencl_codelet_incC(void *descr[], __attribute__ ((unused)) void *_args)
 {
-	unsigned *val = (unsigned *)STARPU_VECTOR_GET_PTR(descr[0]);
+	cl_mem val = (cl_mem)STARPU_VECTOR_GET_PTR(descr[0]);
 	cl_kernel kernel;
 	cl_command_queue queue;
 	cl_event event;
@@ -66,7 +66,7 @@ void opencl_codelet_incC(void *descr[], __attribute__ ((unused)) void *_args)
 	if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
 
 	err = 0;
-	err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &val);
+	err = clSetKernelArg(kernel, 0, sizeof(val), &val);
 	if (err) STARPU_OPENCL_REPORT_ERROR(err);
 
 	{