error*
platform__get_devices(clobj_t _plat, clobj_t **_devices,
uint32_t *num_devices, cl_device_type devtype)
{
auto plat = static_cast<platform*>(_plat);
return c_handle_error([&] {
*num_devices = 0;
try {
pyopencl_call_guarded(clGetDeviceIDs, plat, devtype, 0, nullptr,
buf_arg(*num_devices));
} catch (const clerror &e) {
if (e.code() != CL_DEVICE_NOT_FOUND)
throw e;
*num_devices = 0;
}
if (*num_devices == 0) {
*_devices = nullptr;
return;
}
pyopencl_buf<cl_device_id> devices(*num_devices);
pyopencl_call_guarded(clGetDeviceIDs, plat, devtype, devices,
buf_arg(*num_devices));
*_devices = buf_to_base<device>(devices).release();
});
}
error*
enqueue_copy_buffer(clobj_t *evt, clobj_t _queue, clobj_t _src, clobj_t _dst,
ptrdiff_t byte_count, size_t src_offset, size_t dst_offset,
const clobj_t *_wait_for, uint32_t num_wait_for)
{
auto queue = static_cast<command_queue*>(_queue);
auto src = static_cast<memory_object*>(_src);
auto dst = static_cast<memory_object*>(_dst);
return c_handle_error([&] {
if (byte_count < 0) {
size_t byte_count_src = 0;
size_t byte_count_dst = 0;
pyopencl_call_guarded(
clGetMemObjectInfo, src, CL_MEM_SIZE,
sizeof(byte_count), &byte_count_src, nullptr);
pyopencl_call_guarded(
clGetMemObjectInfo, src, CL_MEM_SIZE,
sizeof(byte_count), &byte_count_dst, nullptr);
byte_count = std::min(byte_count_src, byte_count_dst);
}
const auto wait_for = buf_from_class<event>(_wait_for,
num_wait_for);
retry_mem_error([&] {
pyopencl_call_guarded(
clEnqueueCopyBuffer, queue, src, dst, src_offset,
dst_offset, byte_count, wait_for, event_out(evt));
});
});
}
error*
enqueue_migrate_mem_object_ext(clobj_t *evt, clobj_t _queue,
const clobj_t *_mem_obj, uint32_t num_mem_obj,
cl_mem_migration_flags_ext flags,
const clobj_t *_wait_for, uint32_t num_wait_for)
{
const auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
const auto mem_obj = buf_from_class<memory_object>(_mem_obj, num_mem_obj);
auto queue = static_cast<command_queue*>(_queue);
return c_handle_error([&] {
#if PYOPENCL_CL_VERSION >= 0x1020
// {{{ get platform
cl_device_id dev;
pyopencl_call_guarded(clGetCommandQueueInfo, queue, CL_QUEUE_DEVICE,
size_arg(dev), nullptr);
cl_platform_id plat;
pyopencl_call_guarded(clGetDeviceInfo, dev, CL_DEVICE_PLATFORM,
size_arg(plat), nullptr);
// }}}
#endif
auto clEnqueueMigrateMemObjectEXT =
pyopencl_get_ext_fun(plat, clEnqueueMigrateMemObjectEXT);
retry_mem_error([&] {
pyopencl_call_guarded(clEnqueueMigrateMemObjectsEXT, queue,
mem_obj, flags, wait_for,
event_out(evt));
});
});
}
pyopencl_buf<clobj_t>
program::all_kernels()
{
cl_uint num_knls;
pyopencl_call_guarded(clCreateKernelsInProgram, this, 0, nullptr,
buf_arg(num_knls));
pyopencl_buf<cl_kernel> knls(num_knls);
pyopencl_call_guarded(clCreateKernelsInProgram, this, knls,
buf_arg(num_knls));
return buf_to_base<kernel>(knls, true);
}
error*
get_platforms(clobj_t **_platforms, uint32_t *num_platforms)
{
return c_handle_error([&] {
*num_platforms = 0;
pyopencl_call_guarded(clGetPlatformIDs, 0, nullptr,
buf_arg(*num_platforms));
pyopencl_buf<cl_platform_id> platforms(*num_platforms);
pyopencl_call_guarded(clGetPlatformIDs, platforms,
buf_arg(*num_platforms));
*_platforms = buf_to_base<platform>(platforms).release();
});
}
error*
enqueue_write_buffer_rect(clobj_t *evt, clobj_t _queue, clobj_t _mem, void *buf,
const size_t *_buf_orig, size_t buf_orig_l,
const size_t *_host_orig, size_t host_orig_l,
const size_t *_reg, size_t reg_l,
const size_t *_buf_pitches, size_t buf_pitches_l,
const size_t *_host_pitches, size_t host_pitches_l,
const clobj_t *_wait_for, uint32_t num_wait_for,
int block, void *pyobj)
{
#if PYOPENCL_CL_VERSION >= 0x1010
const auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
auto queue = static_cast<command_queue*>(_queue);
auto mem = static_cast<memory_object*>(_mem);
ConstBuffer<size_t, 3> buf_orig(_buf_orig, buf_orig_l);
ConstBuffer<size_t, 3> host_orig(_host_orig, host_orig_l);
ConstBuffer<size_t, 3> reg(_reg, reg_l, 1);
ConstBuffer<size_t, 2> buf_pitches(_buf_pitches, buf_pitches_l);
ConstBuffer<size_t, 2> host_pitches(_host_pitches, host_pitches_l);
return c_handle_retry_mem_error([&] {
pyopencl_call_guarded(
clEnqueueWriteBufferRect, queue, mem, bool(block), buf_orig,
host_orig, reg, buf_pitches[0], buf_pitches[1], host_pitches[0],
host_pitches[1], buf, wait_for, nanny_event_out(evt, pyobj));
});
#else
PYOPENCL_UNSUPPORTED(clEnqueueWriteBufferRect, "CL 1.0")
#endif
}
error*
context__get_supported_image_formats(clobj_t _ctx, cl_mem_flags flags,
cl_mem_object_type image_type,
generic_info *out)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_error([&] {
cl_uint num;
pyopencl_call_guarded(clGetSupportedImageFormats, ctx, flags,
image_type, 0, nullptr, buf_arg(num));
pyopencl_buf<cl_image_format> formats(num);
pyopencl_call_guarded(clGetSupportedImageFormats, ctx, flags,
image_type, formats, buf_arg(num));
*out = pyopencl_convert_array_info(cl_image_format, formats);
});
}
error*
enqueue_copy_buffer_rect(clobj_t *evt, clobj_t _queue, clobj_t _src,
clobj_t _dst, const size_t *_src_orig,
size_t src_orig_l, const size_t *_dst_orig,
size_t dst_orig_l, const size_t *_reg, size_t reg_l,
const size_t *_src_pitches, size_t src_pitches_l,
const size_t *_dst_pitches, size_t dst_pitches_l,
const clobj_t *_wait_for, uint32_t num_wait_for)
{
#if PYOPENCL_CL_VERSION >= 0x1010
const auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
auto queue = static_cast<command_queue*>(_queue);
auto src = static_cast<memory_object*>(_src);
auto dst = static_cast<memory_object*>(_dst);
ConstBuffer<size_t, 3> src_orig(_src_orig, src_orig_l);
ConstBuffer<size_t, 3> dst_orig(_dst_orig, dst_orig_l);
ConstBuffer<size_t, 3> reg(_reg, reg_l, 1);
ConstBuffer<size_t, 2> src_pitches(_src_pitches, src_pitches_l);
ConstBuffer<size_t, 2> dst_pitches(_dst_pitches, dst_pitches_l);
return c_handle_retry_mem_error([&] {
pyopencl_call_guarded(
clEnqueueCopyBufferRect, queue, src, dst, src_orig, dst_orig,
reg, src_pitches[0], src_pitches[1], dst_pitches[0],
dst_pitches[1], wait_for, event_out(evt));
});
#else
PYOPENCL_UNSUPPORTED(clEnqueueCopyBufferRect, "CL 1.0")
#endif
}
error*
platform__unload_compiler(clobj_t plat)
{
return c_handle_error([&] {
pyopencl_call_guarded(clUnloadPlatformCompiler,
static_cast<platform*>(plat));
});
}
error*
get_gl_object_info(clobj_t mem, cl_gl_object_type *otype, GLuint *gl_name)
{
auto globj = static_cast<memory_object*>(mem);
return c_handle_error([&] {
pyopencl_call_guarded(clGetGLObjectInfo, globj, buf_arg(*otype),
buf_arg(*gl_name));
});
}
error*
kernel__set_arg_null(clobj_t _knl, cl_uint arg_index)
{
auto knl = static_cast<kernel*>(_knl);
return c_handle_error([&] {
const cl_mem m = 0;
pyopencl_call_guarded(clSetKernelArg, knl, arg_index, size_arg(m));
});
}
// Kernel
error*
create_kernel(clobj_t *knl, clobj_t _prog, const char *name)
{
auto prog = static_cast<const program*>(_prog);
return c_handle_error([&] {
*knl = new kernel(pyopencl_call_guarded(clCreateKernel, prog,
name), false);
});
}
error*
memory_object__get_host_array(clobj_t _obj, void **hostptr, size_t *size)
{
auto obj = static_cast<memory_object*>(_obj);
return c_handle_error([&] {
cl_mem_flags flags;
pyopencl_call_guarded(clGetMemObjectInfo, obj, CL_MEM_FLAGS,
size_arg(flags), nullptr);
if (!(flags & CL_MEM_USE_HOST_PTR))
throw clerror("MemoryObject.get_host_array", CL_INVALID_VALUE,
"Only MemoryObject with USE_HOST_PTR "
"is supported.");
pyopencl_call_guarded(clGetMemObjectInfo, obj, CL_MEM_HOST_PTR,
size_arg(*hostptr), nullptr);
pyopencl_call_guarded(clGetMemObjectInfo, obj, CL_MEM_SIZE,
size_arg(*size), nullptr);
});
}
error*
kernel__set_arg_sampler(clobj_t _knl, cl_uint arg_index, clobj_t _samp)
{
auto knl = static_cast<kernel*>(_knl);
auto samp = static_cast<sampler*>(_samp);
return c_handle_error([&] {
pyopencl_call_guarded(clSetKernelArg, knl, arg_index,
size_arg(samp->data()));
});
}
error*
kernel__set_arg_buf(clobj_t _knl, cl_uint arg_index,
const void *buffer, size_t size)
{
auto knl = static_cast<kernel*>(_knl);
return c_handle_error([&] {
pyopencl_call_guarded(clSetKernelArg, knl, arg_index,
size_arg(buffer, size));
});
}
PYOPENCL_USE_RESULT static gl_texture*
create_from_gl_texture(const context *ctx, cl_mem_flags flags,
GLenum texture_target, GLint miplevel,
GLuint texture, unsigned dims)
{
if (dims == 2) {
cl_mem mem = pyopencl_call_guarded(clCreateFromGLTexture2D,
ctx, flags, texture_target,
miplevel, texture);
return pyopencl_convert_obj(gl_texture, clReleaseMemObject, mem);
} else if (dims == 3) {
cl_mem mem = pyopencl_call_guarded(clCreateFromGLTexture3D,
ctx, flags, texture_target,
miplevel, texture);
return pyopencl_convert_obj(gl_texture, clReleaseMemObject, mem);
} else {
throw clerror("Image", CL_INVALID_VALUE, "invalid dimension");
}
}
void
program::compile(const char *opts, const clobj_t *_devs, size_t num_devs,
const clobj_t *_prgs, const char *const *names,
size_t num_hdrs)
{
const auto devs = buf_from_class<device>(_devs, num_devs);
const auto prgs = buf_from_class<program>(_prgs, num_hdrs);
pyopencl_call_guarded(clCompileProgram, this, devs, opts, prgs,
buf_arg(names, num_hdrs), nullptr, nullptr);
}
error*
kernel__set_arg_mem(clobj_t _knl, cl_uint arg_index, clobj_t _mem)
{
auto knl = static_cast<kernel*>(_knl);
auto mem = static_cast<memory_object*>(_mem);
return c_handle_error([&] {
pyopencl_call_guarded(clSetKernelArg, knl, arg_index,
size_arg(mem->data()));
});
}
// Buffer
error*
create_buffer(clobj_t *buffer, clobj_t _ctx, cl_mem_flags flags,
size_t size, void *hostbuf)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_retry_mem_error([&] {
auto mem = pyopencl_call_guarded(clCreateBuffer, ctx,
flags, size, hostbuf);
*buffer = new_buffer(mem);
});
}
error*
program__build(clobj_t _prog, const char *options,
cl_uint num_devices, const clobj_t *_devices)
{
auto prog = static_cast<const program*>(_prog);
const auto devices = buf_from_class<device>(_devices, num_devices);
return c_handle_error([&] {
pyopencl_call_guarded(clBuildProgram, prog, devices, options,
nullptr, nullptr);
});
}
error*
create_from_gl_buffer(clobj_t *ptr, clobj_t _ctx,
cl_mem_flags flags, GLuint bufobj)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_error([&] {
cl_mem mem = pyopencl_call_guarded(clCreateFromGLBuffer,
ctx, flags, bufobj);
*ptr = pyopencl_convert_obj(gl_buffer, clReleaseMemObject, mem);
});
}
error*
create_image_from_desc(clobj_t *img, clobj_t _ctx, cl_mem_flags flags,
cl_image_format *fmt, cl_image_desc *desc, void *buf)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_error([&] {
auto mem = pyopencl_call_guarded(clCreateImage, ctx, flags, fmt,
desc, buf);
*img = new_image(mem, fmt);
});
}
PYOPENCL_USE_RESULT buffer*
buffer::get_sub_region(size_t orig, size_t size, cl_mem_flags flags) const
{
cl_buffer_region reg = {orig, size};
auto mem = retry_mem_error([&] {
return pyopencl_call_guarded(clCreateSubBuffer, PYOPENCL_CL_CASTABLE_THIS, flags,
CL_BUFFER_CREATE_TYPE_REGION, ®);
});
return new_buffer(mem);
}
// Image
error*
create_image_2d(clobj_t *img, clobj_t _ctx, cl_mem_flags flags,
cl_image_format *fmt, size_t width, size_t height,
size_t pitch, void *buf)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_retry_mem_error([&] {
auto mem = pyopencl_call_guarded(clCreateImage2D, ctx, flags, fmt,
width, height, pitch, buf);
*img = new_image(mem, fmt);
});
}
// Program
error*
create_program_with_source(clobj_t *prog, clobj_t _ctx, const char *_src)
{
auto ctx = static_cast<context*>(_ctx);
return c_handle_error([&] {
const auto &src = _src;
const size_t length = strlen(src);
cl_program result = pyopencl_call_guarded(
clCreateProgramWithSource, ctx, len_arg(src), buf_arg(length));
*prog = new_program(result, KND_SOURCE);
});
}
error*
enqueue_task(clobj_t *evt, clobj_t _queue, clobj_t _knl,
const clobj_t *_wait_for, uint32_t num_wait_for)
{
auto queue = static_cast<command_queue*>(_queue);
auto knl = static_cast<kernel*>(_knl);
const auto wait_for = buf_from_class<event>(_wait_for, num_wait_for);
return c_handle_retry_mem_error([&] {
pyopencl_call_guarded(clEnqueueTask, queue, knl, wait_for,
event_out(evt));
});
}
error*
kernel__set_arg_svm_pointer(clobj_t _knl, cl_uint arg_index, void *value)
{
#if PYOPENCL_CL_VERSION >= 0x2000
auto knl = static_cast<kernel*>(_knl);
return c_handle_error([&] {
pyopencl_call_guarded(clSetKernelArgSVMPointer, knl, arg_index, value);
});
#else
PYOPENCL_UNSUPPORTED_BEFORE(clSetKernelArgSVMPointer, "CL 2.0")
#endif
}
void
context::get_version(cl_context ctx, int *major, int *minor)
{
cl_device_id s_buff[16];
size_t size;
pyopencl_buf<cl_device_id> d_buff(0);
cl_device_id *devs = s_buff;
pyopencl_call_guarded(clGetContextInfo, ctx, CL_CONTEXT_DEVICES,
0, nullptr, buf_arg(size));
if (PYOPENCL_UNLIKELY(!size)) {
throw clerror("Context.get_version", CL_INVALID_VALUE,
"Cannot get devices from context.");
}
if (PYOPENCL_UNLIKELY(size > sizeof(s_buff))) {
d_buff.resize(size / sizeof(cl_device_id));
devs = d_buff.get();
}
pyopencl_call_guarded(clGetContextInfo, ctx, CL_CONTEXT_DEVICES,
size_arg(devs, size), buf_arg(size));
device::get_version(devs[0], major, minor);
}
generic_info
program::get_info(cl_uint param) const
{
switch ((cl_program_info)param) {
case CL_PROGRAM_CONTEXT:
return pyopencl_get_opaque_info(context, Program, this, param);
case CL_PROGRAM_REFERENCE_COUNT:
case CL_PROGRAM_NUM_DEVICES:
return pyopencl_get_int_info(cl_uint, Program, this, param);
case CL_PROGRAM_DEVICES:
return pyopencl_get_opaque_array_info(device, Program, this, param);
case CL_PROGRAM_SOURCE:
return pyopencl_get_str_info(Program, this, param);
case CL_PROGRAM_BINARY_SIZES:
return pyopencl_get_array_info(size_t, Program, this, param);
case CL_PROGRAM_BINARIES: {
auto sizes = pyopencl_get_vec_info(size_t, Program, this,
CL_PROGRAM_BINARY_SIZES);
pyopencl_buf<char*> result_ptrs(sizes.len());
for (size_t i = 0;i < sizes.len();i++) {
result_ptrs[i] = (char*)malloc(sizes[i]);
}
try {
pyopencl_call_guarded(clGetProgramInfo, this, CL_PROGRAM_BINARIES,
sizes.len() * sizeof(char*),
result_ptrs.get(), nullptr);
} catch (...) {
for (size_t i = 0;i < sizes.len();i++) {
free(result_ptrs[i]);
}
}
pyopencl_buf<generic_info> gis(sizes.len());
for (size_t i = 0;i < sizes.len();i++) {
gis[i].value = result_ptrs[i];
gis[i].dontfree = 0;
gis[i].opaque_class = CLASS_NONE;
gis[i].type = _copy_str(std::string("char[") +
tostring(sizes[i]) + "]");
}
return pyopencl_convert_array_info(generic_info, gis);
}
#if PYOPENCL_CL_VERSION >= 0x1020
case CL_PROGRAM_NUM_KERNELS:
return pyopencl_get_int_info(size_t, Program, this, param);
case CL_PROGRAM_KERNEL_NAMES:
return pyopencl_get_str_info(Program, this, param);
#endif
default:
throw clerror("Program.get_info", CL_INVALID_VALUE);
}
}
// Context
error*
create_context_from_type(clobj_t *_ctx, const cl_context_properties *props,
cl_device_type dev_type)
{
// TODO debug print properties
return c_handle_error([&] {
*_ctx = new context(
pyopencl_call_guarded(
clCreateContextFromType,
const_cast<cl_context_properties*>(props),
dev_type, nullptr, nullptr), false);
});
}
