int call_pocl_build( cl_device_id device,
const char* source_file_name,
const char* binary_file_name,
const char* device_tmpdir,
const char* user_options )
{
int error;
const char *pocl_build_script;
char command[COMMAND_LENGTH];
if (getenv("POCL_BUILDING") != NULL)
pocl_build_script = BUILDDIR "/scripts/" POCL_BUILD;
else if (access(PKGDATADIR "/" POCL_BUILD, X_OK) == 0)
pocl_build_script = PKGDATADIR "/" POCL_BUILD;
else
pocl_build_script = POCL_BUILD;
if (device->llvm_target_triplet != NULL)
{
error = snprintf(command, COMMAND_LENGTH,
"USER_OPTIONS=\"%s\" %s -t %s -o %s %s",
user_options,
pocl_build_script,
device->llvm_target_triplet,
binary_file_name, source_file_name);
}
else
{
error = snprintf(command, COMMAND_LENGTH,
"USER_OPTIONS=\"%s\" %s -o %s %s",
user_options,
pocl_build_script,
binary_file_name, source_file_name);
}
if (error < 0)
return CL_OUT_OF_HOST_MEMORY;
/* call the customized build command, if needed for the
device driver */
if (device->build_program != NULL)
{
error = device->build_program
(device->data, source_file_name, binary_file_name,
command, user_options, device_tmpdir);
}
else
{
error = system(command);
}
return error;
}
cl_int
clGetDeviceInfo(cl_device_id device,
cl_device_info param_name,
size_t param_value_size,
void * param_value,
size_t * param_value_size_ret)
{
#ifdef DBG_API
std::cerr << "clGetDeviceInfo\n";
#endif
if (!device->isA(Coal::Object::T_Device))
return CL_INVALID_DEVICE;
Coal::DeviceInterface *iface = (Coal::DeviceInterface *)device;
return iface->info(param_name, param_value_size, param_value,
param_value_size_ret);
}
// Command Queue APIs
cl_command_queue
clCreateCommandQueue(cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int * errcode_ret)
{
cl_int default_errcode_ret;
// No errcode_ret ?
if (!errcode_ret)
errcode_ret = &default_errcode_ret;
if (!device->isA(Coal::Object::T_Device))
{
*errcode_ret = CL_INVALID_DEVICE;
return 0;
}
if (!context->isA(Coal::Object::T_Context))
{
*errcode_ret = CL_INVALID_CONTEXT;
return 0;
}
*errcode_ret = CL_SUCCESS;
Coal::CommandQueue *queue = new Coal::CommandQueue(
(Coal::Context *)context,
(Coal::DeviceInterface *)device,
properties,
errcode_ret);
if (*errcode_ret != CL_SUCCESS)
{
// Initialization failed, destroy context
delete queue;
return 0;
}
return (_cl_command_queue *)queue;
}
// Command Queue APIs
cl_command_queue
clCreateCommandQueue(cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int * errcode_ret)
{
#ifdef DBG_API
std::cerr << "Entering clCreateCommandQueue\n";
#endif
cl_int default_errcode_ret;
// No errcode_ret ?
if (!errcode_ret)
errcode_ret = &default_errcode_ret;
#ifdef DBG_API
std::cerr << "Check if the device is an object\n";
#endif
if (!device->isA(Coal::Object::T_Device))
{
std::cerr << "INVALID_DEVICE\n";
*errcode_ret = CL_INVALID_DEVICE;
return 0;
}
if (!context->isA(Coal::Object::T_Context))
{
#ifdef DBG_OUTPUT
std::cout << "!!! ERROR: INVALID_CONTEXT" << std::endl;
#endif
*errcode_ret = CL_INVALID_CONTEXT;
return 0;
}
#ifdef DBG_API
std::cerr << "Attempt to initialise device\n";
#endif
if (!device->init()) {
#ifdef DBG_OUTPUT
std::cout << "!!!ERROR: Device initialisation failed!\n";
#endif
*errcode_ret = CL_DEVICE_NOT_AVAILABLE;
return 0;
}
*errcode_ret = CL_SUCCESS;
Coal::CommandQueue *queue = new Coal::CommandQueue(
(Coal::Context *)context,
(Coal::DeviceInterface *)device,
properties,
errcode_ret);
if (*errcode_ret != CL_SUCCESS)
{
#ifdef DBG_OUTPUT
std::cout << "!!! ERROR: CommandQueue create failed" << std::endl;
#endif
// Initialization failed, destroy context
delete queue;
return 0;
}
return (_cl_command_queue *)queue;
}
PUBLIC cl_int
clGetDeviceInfo(cl_device_id dev, cl_device_info param,
size_t size, void *buf, size_t *size_ret) {
if (!dev)
return CL_INVALID_DEVICE;
switch (param) {
case CL_DEVICE_TYPE:
return scalar_property<cl_device_type>(buf, size, size_ret, dev->type());
case CL_DEVICE_VENDOR_ID:
return scalar_property<cl_uint>(buf, size, size_ret, dev->vendor_id());
case CL_DEVICE_MAX_COMPUTE_UNITS:
return scalar_property<cl_uint>(buf, size, size_ret, 1);
case CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS:
return scalar_property<cl_uint>(buf, size, size_ret,
dev->max_block_size().size());
case CL_DEVICE_MAX_WORK_ITEM_SIZES:
return vector_property<size_t>(buf, size, size_ret,
dev->max_block_size());
case CL_DEVICE_MAX_WORK_GROUP_SIZE:
return scalar_property<size_t>(buf, size, size_ret,
dev->max_threads_per_block());
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR:
return scalar_property<cl_uint>(buf, size, size_ret, 16);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT:
return scalar_property<cl_uint>(buf, size, size_ret, 8);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT:
return scalar_property<cl_uint>(buf, size, size_ret, 4);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG:
return scalar_property<cl_uint>(buf, size, size_ret, 2);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT:
return scalar_property<cl_uint>(buf, size, size_ret, 4);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE:
return scalar_property<cl_uint>(buf, size, size_ret, 2);
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF:
return scalar_property<cl_uint>(buf, size, size_ret, 0);
case CL_DEVICE_MAX_CLOCK_FREQUENCY:
return scalar_property<cl_uint>(buf, size, size_ret, 0);
case CL_DEVICE_ADDRESS_BITS:
return scalar_property<cl_uint>(buf, size, size_ret, 32);
case CL_DEVICE_MAX_READ_IMAGE_ARGS:
return scalar_property<cl_uint>(buf, size, size_ret,
dev->max_images_read());
case CL_DEVICE_MAX_WRITE_IMAGE_ARGS:
return scalar_property<cl_uint>(buf, size, size_ret,
dev->max_images_write());
case CL_DEVICE_MAX_MEM_ALLOC_SIZE:
return scalar_property<cl_ulong>(buf, size, size_ret,
dev->max_mem_alloc_size());
case CL_DEVICE_IMAGE2D_MAX_WIDTH:
case CL_DEVICE_IMAGE2D_MAX_HEIGHT:
return scalar_property<size_t>(buf, size, size_ret,
1 << dev->max_image_levels_2d());
case CL_DEVICE_IMAGE3D_MAX_WIDTH:
case CL_DEVICE_IMAGE3D_MAX_HEIGHT:
case CL_DEVICE_IMAGE3D_MAX_DEPTH:
return scalar_property<size_t>(buf, size, size_ret,
1 << dev->max_image_levels_3d());
case CL_DEVICE_IMAGE_SUPPORT:
return scalar_property<cl_bool>(buf, size, size_ret, CL_TRUE);
case CL_DEVICE_MAX_PARAMETER_SIZE:
return scalar_property<size_t>(buf, size, size_ret,
dev->max_mem_input());
case CL_DEVICE_MAX_SAMPLERS:
return scalar_property<cl_uint>(buf, size, size_ret,
dev->max_samplers());
case CL_DEVICE_MEM_BASE_ADDR_ALIGN:
case CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE:
return scalar_property<cl_uint>(buf, size, size_ret, 128);
case CL_DEVICE_SINGLE_FP_CONFIG:
return scalar_property<cl_device_fp_config>(buf, size, size_ret,
CL_FP_DENORM | CL_FP_INF_NAN | CL_FP_ROUND_TO_NEAREST);
case CL_DEVICE_GLOBAL_MEM_CACHE_TYPE:
return scalar_property<cl_device_mem_cache_type>(buf, size, size_ret,
CL_NONE);
case CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE:
return scalar_property<cl_uint>(buf, size, size_ret, 0);
case CL_DEVICE_GLOBAL_MEM_CACHE_SIZE:
return scalar_property<cl_ulong>(buf, size, size_ret, 0);
case CL_DEVICE_GLOBAL_MEM_SIZE:
return scalar_property<cl_ulong>(buf, size, size_ret,
dev->max_mem_global());
case CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE:
return scalar_property<cl_ulong>(buf, size, size_ret,
dev->max_const_buffer_size());
case CL_DEVICE_MAX_CONSTANT_ARGS:
return scalar_property<cl_uint>(buf, size, size_ret,
dev->max_const_buffers());
case CL_DEVICE_LOCAL_MEM_TYPE:
return scalar_property<cl_device_local_mem_type>(buf, size, size_ret,
CL_LOCAL);
case CL_DEVICE_LOCAL_MEM_SIZE:
return scalar_property<cl_ulong>(buf, size, size_ret,
dev->max_mem_local());
case CL_DEVICE_ERROR_CORRECTION_SUPPORT:
return scalar_property<cl_bool>(buf, size, size_ret, CL_FALSE);
case CL_DEVICE_PROFILING_TIMER_RESOLUTION:
return scalar_property<size_t>(buf, size, size_ret, 0);
case CL_DEVICE_ENDIAN_LITTLE:
return scalar_property<cl_bool>(buf, size, size_ret,
dev->endianness() == PIPE_ENDIAN_LITTLE);
case CL_DEVICE_AVAILABLE:
case CL_DEVICE_COMPILER_AVAILABLE:
return scalar_property<cl_bool>(buf, size, size_ret, CL_TRUE);
case CL_DEVICE_EXECUTION_CAPABILITIES:
return scalar_property<cl_device_exec_capabilities>(buf, size, size_ret,
CL_EXEC_KERNEL);
case CL_DEVICE_QUEUE_PROPERTIES:
return scalar_property<cl_command_queue_properties>(buf, size, size_ret,
CL_QUEUE_PROFILING_ENABLE);
case CL_DEVICE_NAME:
return string_property(buf, size, size_ret, dev->device_name());
case CL_DEVICE_VENDOR:
return string_property(buf, size, size_ret, dev->vendor_name());
case CL_DRIVER_VERSION:
return string_property(buf, size, size_ret, PACKAGE_VERSION);
case CL_DEVICE_PROFILE:
return string_property(buf, size, size_ret, "FULL_PROFILE");
case CL_DEVICE_VERSION:
return string_property(buf, size, size_ret,
"OpenCL 1.1 MESA " PACKAGE_VERSION);
case CL_DEVICE_EXTENSIONS:
return string_property(buf, size, size_ret, "");
case CL_DEVICE_PLATFORM:
return scalar_property<cl_platform_id>(buf, size, size_ret,
&dev->platform);
case CL_DEVICE_HOST_UNIFIED_MEMORY:
return scalar_property<cl_bool>(buf, size, size_ret, CL_TRUE);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR:
return scalar_property<cl_uint>(buf, size, size_ret, 16);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT:
return scalar_property<cl_uint>(buf, size, size_ret, 8);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_INT:
return scalar_property<cl_uint>(buf, size, size_ret, 4);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG:
return scalar_property<cl_uint>(buf, size, size_ret, 2);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT:
return scalar_property<cl_uint>(buf, size, size_ret, 4);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE:
return scalar_property<cl_uint>(buf, size, size_ret, 2);
case CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF:
return scalar_property<cl_uint>(buf, size, size_ret, 0);
case CL_DEVICE_OPENCL_C_VERSION:
return string_property(buf, size, size_ret, "OpenCL C 1.1");
default:
return CL_INVALID_VALUE;
}
}
extern cl_int
pocl_write_image (cl_mem image,
cl_device_id device_id,
const size_t * origin_, /*[3]*/
const size_t * region_, /*[3]*/
size_t host_row_pitch,
size_t host_slice_pitch,
const void * ptr)
{
if (image == NULL)
return CL_INVALID_MEM_OBJECT;
if ((ptr == NULL) ||
(region_ == NULL))
return CL_INVALID_VALUE;
int width = image->image_width;
int height = image->image_height;
cl_channel_order order = image->image_channel_order;
cl_channel_type type = image->image_channel_data_type;
size_t dev_elem_size = sizeof(cl_float);
int dev_channels = 4;
int host_elem_size;
int host_channels;
pocl_get_image_information (image, &host_channels, &host_elem_size);
size_t origin[3] = { origin_[0]*dev_elem_size*dev_channels, origin_[1], origin_[2] };
size_t region[3] = { region_[0]*dev_elem_size*dev_channels, region_[1], region_[2] };
size_t image_row_pitch = width*dev_elem_size*dev_channels;
size_t image_slice_pitch = 0;
if ((region[0]*region[1]*region[2] > 0) &&
(region[0]-1 +
image_row_pitch * (region[1]-1) +
image_slice_pitch * (region[2]-1) >= image->size))
return CL_INVALID_VALUE;
cl_float* temp = malloc( width*height*dev_channels*dev_elem_size );
if (temp == NULL)
return CL_OUT_OF_HOST_MEMORY;
int x, y, k;
for (y=0; y<height; y++)
for (x=0; x<width*dev_channels; x++)
temp[x+y*width*dev_channels] = 0.f;
for (y=0; y<height; y++)
{
for (x=0; x<width; x++)
{
cl_float elem[4]; //TODO 0,0,0,0 for some modes?
for (k=0; k<host_channels; k++)
{
if (type == CL_FLOAT)
elem[k] = ((float*)ptr)[k+(x+y*width)*host_channels];
else if (type==CL_UNORM_INT8)
{
cl_uchar foo = ((cl_uchar*)ptr)[k+(x+y*width)*host_channels];
elem[k] = (float)(foo) * (1.f/255.f);
}
else
POCL_ABORT_UNIMPLEMENTED();
}
if (order == CL_RGBA)
for (k=0; k<4; k++)
temp[(x+y*width)*dev_channels+k] = elem[k];
else if (order == CL_R)
{
temp[(x+y*width)*dev_channels+0] = elem[0];
temp[(x+y*width)*dev_channels+1] = 0.f;
temp[(x+y*width)*dev_channels+2] = 0.f;
temp[(x+y*width)*dev_channels+3] = 1.f;
}
}
}
device_id->write_rect(device_id->data, temp,
image->device_ptrs[device_id->dev_id],
origin, origin, region,
image_row_pitch, image_slice_pitch,
image_row_pitch, image_slice_pitch);
free (temp);
return CL_SUCCESS;
}
extern cl_int
pocl_read_image (cl_mem image,
cl_device_id device_id,
const size_t * origin_, /*[3]*/
const size_t * region_, /*[3]*/
size_t host_row_pitch,
size_t host_slice_pitch,
void * ptr)
{
if (image == NULL)
return CL_INVALID_MEM_OBJECT;
if ((ptr == NULL) ||
(region_ == NULL))
return CL_INVALID_VALUE;
int width = image->image_width;
int height = image->image_height;
int dev_elem_size = sizeof(cl_float);
int dev_channels = 4;
size_t origin[3] = { origin_[0]*dev_elem_size*dev_channels, origin_[1], origin_[2] };
size_t region[3] = { region_[0]*dev_elem_size*dev_channels, region_[1], region_[2] };
size_t image_row_pitch = width*dev_elem_size*dev_channels;
size_t image_slice_pitch = 0;
if ((region[0]*region[1]*region[2] > 0) &&
(region[0]-1 +
image_row_pitch * (region[1]-1) +
image_slice_pitch * (region[2]-1) >= image->size))
return CL_INVALID_VALUE;
int i, j, k;
cl_channel_order order = image->image_channel_order;
cl_channel_type type = image->image_channel_data_type;
cl_float* temp = malloc( width*height*dev_channels*dev_elem_size );
if (temp == NULL)
return CL_OUT_OF_HOST_MEMORY;
int host_channels, host_elem_size;
pocl_get_image_information(image, &host_channels, &host_elem_size);
if (host_row_pitch == 0) {
host_row_pitch = width*host_channels;
}
size_t buffer_origin[3] = { 0, 0, 0 };
device_id->read_rect(device_id->data, temp,
image->device_ptrs[device_id->dev_id],
origin, origin, region,
image_row_pitch, image_slice_pitch,
image_row_pitch, image_slice_pitch);
for (j=0; j<height; j++) {
for (i=0; i<width; i++) {
cl_float elem[4];
for (k=0; k<4; k++)
elem[k]=0;
if (order == CL_RGBA) {
for (k=0; k<4; k++)
elem[k] = temp[i*dev_channels + j*width*dev_channels + k];
}
else if (order == CL_R) { // host_channels == 1
elem[0] = temp[i*dev_channels + j*width*dev_channels + 0];
}
if (type == CL_UNORM_INT8)
{ // host_channels == 4
for (k=0; k<host_channels; k++)
{
((cl_uchar*)ptr)[i*host_channels + j*host_row_pitch + k]
= (unsigned char)(255*elem[k]);
}
}
else if (type == CL_FLOAT)
{
for (k=0; k<host_channels; k++)
{
POCL_ABORT_UNIMPLEMENTED();
((cl_float*)ptr)[i*host_channels + j*host_row_pitch + k]
= elem[k];
}
}
else
POCL_ABORT_UNIMPLEMENTED();
}
}
free (temp);
return CL_SUCCESS;
}
YamiStatus
OclContext::createBufferFromFdIntel(const cl_import_buffer_info_intel* info, cl_mem* mem)
{
return m_device->createBufferFromFdIntel(m_context, info, mem);
}
bool OclContext::releaseKernel(OclKernelMap& kernelMap)
{
return m_device->releaseKernel(kernelMap);
}
bool OclContext::createKernel(const char* name, OclKernelMap& kernelMap)
{
return m_device->createKernel(m_context, name, kernelMap);
}