void
pocl_basic_run
(void *data,
_cl_command_node* cmd)
{
struct data *d;
struct pocl_argument *al;
size_t x, y, z;
unsigned i;
cl_kernel kernel = cmd->command.run.kernel;
struct pocl_context *pc = &cmd->command.run.pc;
assert (data != NULL);
d = (struct data *) data;
d->current_kernel = kernel;
void **arguments = (void**)malloc(
sizeof(void*) * (kernel->num_args + kernel->num_locals)
);
/* Process the kernel arguments. Convert the opaque buffer
pointers to real device pointers, allocate dynamic local
memory buffers, etc. */
for (i = 0; i < kernel->num_args; ++i)
{
al = &(cmd->command.run.arguments[i]);
if (kernel->arg_info[i].is_local)
{
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = pocl_memalign_alloc(MAX_EXTENDED_ALIGNMENT, al->size);
}
else if (kernel->arg_info[i].type == POCL_ARG_TYPE_POINTER)
{
/* It's legal to pass a NULL pointer to clSetKernelArguments. In
that case we must pass the same NULL forward to the kernel.
Otherwise, the user must have created a buffer with per device
pointers stored in the cl_mem. */
if (al->value == NULL)
{
arguments[i] = malloc (sizeof (void *));
*(void **)arguments[i] = NULL;
}
else
arguments[i] = &((*(cl_mem *) (al->value))->device_ptrs[cmd->device->dev_id].mem_ptr);
}
else if (kernel->arg_info[i].type == POCL_ARG_TYPE_IMAGE)
{
dev_image_t di;
fill_dev_image_t (&di, al, cmd->device);
void* devptr = pocl_memalign_alloc(MAX_EXTENDED_ALIGNMENT, sizeof(dev_image_t));
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = devptr;
pocl_basic_write (data, &di, devptr, 0, sizeof(dev_image_t));
}
else if (kernel->arg_info[i].type == POCL_ARG_TYPE_SAMPLER)
{
dev_sampler_t ds;
fill_dev_sampler_t(&ds, al);
void* devptr = pocl_memalign_alloc(MAX_EXTENDED_ALIGNMENT, sizeof(dev_sampler_t));
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = devptr;
pocl_basic_write (data, &ds, devptr, 0, sizeof(dev_sampler_t));
}
else
{
arguments[i] = al->value;
}
}
for (i = kernel->num_args;
i < kernel->num_args + kernel->num_locals;
++i)
{
al = &(cmd->command.run.arguments[i]);
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = pocl_memalign_alloc(MAX_EXTENDED_ALIGNMENT, al->size);
}
for (z = 0; z < pc->num_groups[2]; ++z)
{
for (y = 0; y < pc->num_groups[1]; ++y)
{
for (x = 0; x < pc->num_groups[0]; ++x)
{
pc->group_id[0] = x;
pc->group_id[1] = y;
pc->group_id[2] = z;
cmd->command.run.wg (arguments, pc);
}
}
}
for (i = 0; i < kernel->num_args; ++i)
{
if (kernel->arg_info[i].is_local)
{
POCL_MEM_FREE(*(void **)(arguments[i]));
POCL_MEM_FREE(arguments[i]);
}
else if (kernel->arg_info[i].type == POCL_ARG_TYPE_IMAGE ||
kernel->arg_info[i].type == POCL_ARG_TYPE_SAMPLER)
{
POCL_MEM_FREE(*(void **)(arguments[i]));
POCL_MEM_FREE(arguments[i]);
}
else if (kernel->arg_info[i].type == POCL_ARG_TYPE_POINTER && *(void**)arguments[i] == NULL)
{
POCL_MEM_FREE(arguments[i]);
}
}
for (i = kernel->num_args;
i < kernel->num_args + kernel->num_locals;
++i)
{
POCL_MEM_FREE(*(void **)(arguments[i]));
POCL_MEM_FREE(arguments[i]);
}
free(arguments);
}
void
pocl_basic_run
(void *data,
_cl_command_node* cmd)
{
struct data *d;
int error;
const char *module_fn;
char command[COMMAND_LENGTH];
char workgroup_string[WORKGROUP_STRING_LENGTH];
unsigned device;
struct pocl_argument *al;
size_t x, y, z;
unsigned i;
pocl_workgroup w;
char* tmpdir = cmd->command.run.tmp_dir;
cl_kernel kernel = cmd->command.run.kernel;
struct pocl_context *pc = &cmd->command.run.pc;
assert (data != NULL);
d = (struct data *) data;
module_fn = llvm_codegen (tmpdir);
d->current_dlhandle = lt_dlopen (module_fn);
if (d->current_dlhandle == NULL)
{
printf ("pocl error: lt_dlopen(\"%s\") failed with '%s'.\n", module_fn, lt_dlerror());
printf ("note: missing symbols in the kernel binary might be reported as 'file not found' errors.\n");
abort();
}
d->current_kernel = kernel;
/* Find which device number within the context correspond
to current device. */
for (i = 0; i < kernel->context->num_devices; ++i)
{
if (kernel->context->devices[i]->data == data)
{
device = i;
break;
}
}
snprintf (workgroup_string, WORKGROUP_STRING_LENGTH,
"_%s_workgroup", kernel->function_name);
w = (pocl_workgroup) lt_dlsym (d->current_dlhandle, workgroup_string);
if (w == NULL)
{
printf("pocl error: could not load the work-group function '%s' in module '%s'.\n",
workgroup_string, module_fn);
abort();
}
free ((void*) module_fn);
void *arguments[kernel->num_args + kernel->num_locals];
/* Process the kernel arguments. Convert the opaque buffer
pointers to real device pointers, allocate dynamic local
memory buffers, etc. */
for (i = 0; i < kernel->num_args; ++i)
{
al = &(cmd->command.run.arguments[i]);
if (kernel->arg_is_local[i])
{
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = pocl_basic_malloc(data, 0, al->size, NULL);
}
else if (kernel->arg_is_pointer[i])
{
/* It's legal to pass a NULL pointer to clSetKernelArguments. In
that case we must pass the same NULL forward to the kernel.
Otherwise, the user must have created a buffer with per device
pointers stored in the cl_mem. */
if (al->value == NULL)
{
arguments[i] = malloc (sizeof (void *));
*(void **)arguments[i] = NULL;
}
else
arguments[i] = &((*(cl_mem *) (al->value))->device_ptrs[device]);
}
else if (kernel->arg_is_image[i])
{
dev_image2d_t di;
cl_mem mem = *(cl_mem*)al->value;
di.data = &((*(cl_mem *) (al->value))->device_ptrs[device]);
di.data = ((*(cl_mem *) (al->value))->device_ptrs[device]);
di.width = mem->image_width;
di.height = mem->image_height;
di.rowpitch = mem->image_row_pitch;
di.order = mem->image_channel_order;
di.data_type = mem->image_channel_data_type;
void* devptr = pocl_basic_malloc(data, 0, sizeof(dev_image2d_t), NULL);
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = devptr;
pocl_basic_write (data, &di, devptr, sizeof(dev_image2d_t));
}
else if (kernel->arg_is_sampler[i])
{
dev_sampler_t ds;
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = pocl_basic_malloc(data, 0, sizeof(dev_sampler_t), NULL);
pocl_basic_write (data, &ds, *(void**)arguments[i], sizeof(dev_sampler_t));
}
else
{
arguments[i] = al->value;
}
}
for (i = kernel->num_args;
i < kernel->num_args + kernel->num_locals;
++i)
{
al = &(cmd->command.run.arguments[i]);
arguments[i] = malloc (sizeof (void *));
*(void **)(arguments[i]) = pocl_basic_malloc(data, 0, al->size, NULL);
}
for (z = 0; z < pc->num_groups[2]; ++z)
{
for (y = 0; y < pc->num_groups[1]; ++y)
{
for (x = 0; x < pc->num_groups[0]; ++x)
{
pc->group_id[0] = x;
pc->group_id[1] = y;
pc->group_id[2] = z;
w (arguments, pc);
}
}
}
for (i = 0; i < kernel->num_args; ++i)
{
if (kernel->arg_is_local[i])
pocl_basic_free(data, 0, *(void **)(arguments[i]));
}
for (i = kernel->num_args;
i < kernel->num_args + kernel->num_locals;
++i)
pocl_basic_free(data, 0, *(void **)(arguments[i]));
}