/* This function is called for each user statement in the AST,
* i.e., for each kernel body statement, copy statement or sync statement.
*/
static __isl_give isl_printer *opencl_print_kernel_stmt(
__isl_take isl_printer *p,
__isl_take isl_ast_print_options *print_options,
__isl_keep isl_ast_node *node, void *user)
{
isl_id *id;
struct ppcg_kernel_stmt *stmt;
id = isl_ast_node_get_annotation(node);
stmt = isl_id_get_user(id);
isl_id_free(id);
isl_ast_print_options_free(print_options);
switch (stmt->type) {
case ppcg_kernel_copy:
return ppcg_kernel_print_copy(p, stmt);
case ppcg_kernel_sync:
return opencl_print_sync(p, stmt);
case ppcg_kernel_domain:
return print_opencl_kernel_domain(p, stmt);
}
return p;
}
/* Print the user statement of the host code to "p".
*
* The host code may contain original user statements, kernel launches,
* statements that copy data to/from the device and statements
* the initialize or clear the device.
* The original user statements and the kernel launches have
* an associated annotation, while the other statements do not.
* The latter are handled by print_device_node.
* The annotation on the user statements is called "user".
*
* In case of a kernel launch, print a block of statements that
* defines the grid and the block and then launches the kernel.
*/
static __isl_give isl_printer *print_host_user(__isl_take isl_printer *p,
__isl_take isl_ast_print_options *print_options,
__isl_keep isl_ast_node *node, void *user)
{
isl_id *id;
int is_user;
struct ppcg_kernel *kernel;
struct ppcg_kernel_stmt *stmt;
struct print_host_user_data *data;
isl_ast_print_options_free(print_options);
data = (struct print_host_user_data *) user;
id = isl_ast_node_get_annotation(node);
if (!id)
{
//p = isl_printer_print_str(p,"marker_NO_ID_CASE");
return print_device_node(p, node, data->prog);
}
is_user = !strcmp(isl_id_get_name(id), "user");
kernel = is_user ? NULL : isl_id_get_user(id);
stmt = is_user ? isl_id_get_user(id) : NULL;
isl_id_free(id);
if (is_user)
return ppcg_kernel_print_domain(p, stmt);
p = ppcg_start_block(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "dim3 k");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, "_dimBlock");
print_reverse_list(isl_printer_get_file(p),
kernel->n_block, kernel->block_dim);
p = isl_printer_print_str(p, ";");
p = isl_printer_end_line(p);
p = print_grid(p, kernel);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "kernel");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, " <<<k");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, "_dimGrid, k");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, "_dimBlock>>> (");
p = print_kernel_arguments(p, data->prog, kernel, 0);
p = isl_printer_print_str(p, ");");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "cudaCheckKernel();");
p = isl_printer_end_line(p);
p = ppcg_end_block(p);
p = isl_printer_start_line(p);
p = isl_printer_end_line(p);
p = copy_data_from_device_to_device(p,kernel);
printf("printing kernel");
print_kernel(data->prog, kernel, data->cuda);
printf("printing kernel done");
return p;
}
/* Print the user statement of the host code to "p".
*
* The host code may contain original user statements, kernel launches and
* statements that copy data to/from the device.
* The original user statements and the kernel launches have
* an associated annotation, while the data copy statements do not.
* The latter are handled by print_to_from_device.
* The annotation on the user statements is called "user".
*
* In case of a kernel launch, print a block of statements that
* defines the grid and the work group and then launches the kernel.
*
* A grid is composed of many work groups (blocks), each work group holds
* many work-items (threads).
*
* global_work_size[kernel->n_block] represents the total number of work
* items. It points to an array of kernel->n_block unsigned
* values that describe the total number of work-items that will execute
* the kernel. The total number of work-items is computed as:
* global_work_size[0] *...* global_work_size[kernel->n_block - 1].
*
* The size of each work group (i.e. the number of work-items in each work
* group) is described using block_size[kernel->n_block]. The total
* number of work-items in a block (work-group) is computed as:
* block_size[0] *... * block_size[kernel->n_block - 1].
*
* For more information check:
* http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
*/
static __isl_give isl_printer *opencl_print_host_user(
__isl_take isl_printer *p,
__isl_take isl_ast_print_options *print_options,
__isl_keep isl_ast_node *node, void *user)
{
isl_id *id;
int is_user;
struct ppcg_kernel *kernel;
struct ppcg_kernel_stmt *stmt;
struct print_host_user_data_opencl *data;
isl_ast_print_options_free(print_options);
data = (struct print_host_user_data_opencl *) user;
id = isl_ast_node_get_annotation(node);
if (!id)
return print_to_from_device(p, node, data->prog);
is_user = !strcmp(isl_id_get_name(id), "user");
kernel = is_user ? NULL : isl_id_get_user(id);
stmt = is_user ? isl_id_get_user(id) : NULL;
isl_id_free(id);
if (is_user)
return ppcg_kernel_print_domain(p, stmt);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "{");
p = isl_printer_end_line(p);
p = isl_printer_indent(p, 2);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "size_t global_work_size[");
if (kernel->n_block > 0)
p = isl_printer_print_int(p, kernel->n_block);
else
p = isl_printer_print_int(p, 1);
p = isl_printer_print_str(p, "] = {");
p = opencl_print_total_number_of_work_items_as_list(p, kernel);
p = isl_printer_print_str(p, "};");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "size_t block_size[");
if (kernel->n_block > 0)
p = isl_printer_print_int(p, kernel->n_block);
else
p = isl_printer_print_int(p, 1);
p = isl_printer_print_str(p, "] = {");
p = opencl_print_block_sizes(p, kernel);
p = isl_printer_print_str(p, "};");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "cl_kernel kernel");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, " = clCreateKernel(program, \"kernel");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, "\", &err);");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "openclCheckReturn(err);");
p = isl_printer_end_line(p);
opencl_set_kernel_arguments(p, data->prog, kernel);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "openclCheckReturn(clEnqueueNDRangeKernel"
"(queue, kernel");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, ", ");
if (kernel->n_block > 0)
p = isl_printer_print_int(p, kernel->n_block);
else
p = isl_printer_print_int(p, 1);
p = isl_printer_print_str(p, ", NULL, global_work_size, "
"block_size, "
"0, NULL, NULL));");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "openclCheckReturn("
"clReleaseKernel(kernel");
p = isl_printer_print_int(p, kernel->id);
p = isl_printer_print_str(p, "));");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "clFinish(queue);");
p = isl_printer_end_line(p);
p = isl_printer_indent(p, -2);
p = isl_printer_start_line(p);
p = isl_printer_print_str(p, "}");
p = isl_printer_end_line(p);
p = isl_printer_start_line(p);
p = isl_printer_end_line(p);
data->opencl->kprinter = opencl_print_kernel(data->prog, kernel,
data->opencl->kprinter);
return p;
}
isl_ast_node * ast_gen::after_for(isl_ast_node *node, isl_ast_build * builder)
{
if (verbose<ast_gen>::enabled())
cout << "-- After for" << endl;
bool is_deepest_loop = m_deepest_loop == m_current_loop;
bool is_requested_parallel = false;
{
auto iter_expr = isl_ast_node_for_get_iterator(node);
auto id = isl_ast_expr_get_id(iter_expr);
if (verbose<ast_gen>::enabled())
cout << " Loop iter: " << isl_id_get_name(id) << endl;
auto data = isl_id_get_user(id);
if (data == &m_parallel_loop_id)
{
if (verbose<ast_gen>::enabled())
cout << " Requested as parallel" << endl;
is_requested_parallel = true;
}
id = isl_id_free(id);
isl_ast_expr_free(iter_expr);
}
auto id = isl_ast_node_get_annotation(node);
auto info = ast_node_info::get_from_id(id);
// Mark loop parallel if parallelizable and
// either requested by user or outermost parallizable.
if (!m_options.parallel)
{
if (verbose<ast_gen>::enabled())
cout << " Explicit parallelization not enabled." << endl;
}
else if (!info->is_parallelizable)
{
if (verbose<ast_gen>::enabled())
cout << " Not parallelizable." << endl;
}
else if (m_options.parallel_dim < 0)
{
if (m_num_parallelizable_loops != 1)
{
if (verbose<ast_gen>::enabled())
cout << " Not the outermost parallelizable loop." << endl;
}
else
{
info->is_parallel = true;
}
}
else
{
if (!is_requested_parallel)
{
if (verbose<ast_gen>::enabled())
cout << " Not the requested parallel loop." << endl;
}
else
{
if (verbose<ast_gen>::enabled())
cout << " Parallelized." << endl;
info->is_parallel = true;
}
}
// Mark loop vectorized if parallelizable and deepest.
if (m_options.vectorize && is_deepest_loop && info->is_parallelizable)
{
if (verbose<ast_gen>::enabled())
cout << "-- Loop vectorized." << endl;
info->is_vector = true;
}
if (info->is_parallel || info->is_vector)
{
store_parallel_accesses_for_current_dimension(builder);
}
if (info->is_parallelizable)
--m_num_parallelizable_loops;
--m_current_loop;
id = isl_id_free(id);
return node;
}
