blender
/
blender
136
407
Fork 589
Code Issues 6.3k Pull Requests 779 Projects 19 Wiki Activity

Cycles: Use feature-selective base kernel compilation when using split kernel 

The idea is to make all kernels as small as possible to work around possible
issues with buggy drivers which might fail building feature-complete kernels.

It's indeed just a workaround to make at last simple test scenes to render
on OpenCL. Real fix should happen from the driver side.
This commit is contained in:
Sergey Sharybin 2015-07-18 15:09:25 +02:00
parent 5e4a8c6a87
commit 36a952e3e4
1 changed files with 99 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

116
intern/cycles/device/device_opencl.cpp
View File

@ -930,7 +930,7 @@ public:
return md5.get_hex();
}
bool load_kernels(const DeviceRequestedFeatures& /*requested_features*/)
bool load_kernels(const DeviceRequestedFeatures& requested_features)
{
/* Verify if device was initialized. */
if(!device_initialized) {
@ -940,20 +940,23 @@ public:
/* Try to use cached kernel. */
thread_scoped_lock cache_locker;
cpProgram = OpenCLCache::get_program(cpPlatform,
cdDevice,
OpenCLCache::OCL_DEV_BASE_PROGRAM,
cache_locker);
cpProgram = load_cached_kernel(requested_features,
OpenCLCache::OCL_DEV_BASE_PROGRAM,
cache_locker);
if(!cpProgram) {
VLOG(2) << "No cached OpenCL kernel.";
/* Verify we have right opencl version. */
if(!opencl_version_check())
return false;
string build_flags = build_options_for_base_program(requested_features);
/* Calculate md5 hashes to detect changes. */
string kernel_path = path_get("kernel");
string kernel_md5 = path_files_md5_hash(kernel_path);
string device_md5 = device_md5_hash();
string device_md5 = device_md5_hash(build_flags);
/* Path to cached binary.
*
@ -978,27 +981,41 @@ public:
}
/* If binary kernel exists already, try use it. */
if(path_exists(clbin) && load_binary(kernel_path, clbin, "", &cpProgram)) {
if(path_exists(clbin) && load_binary(kernel_path,
clbin,
build_flags,
&cpProgram)) {
/* Kernel loaded from binary, nothing to do. */
VLOG(2) << "Loaded kernel from " << clbin << ".";
}
else {
string init_kernel_source = "#include \"kernels/opencl/kernel.cl\" // " + kernel_md5 + "\n";
/* If does not exist or loading binary failed, compile kernel. */
if(!compile_kernel(kernel_path, init_kernel_source, "", &cpProgram, debug_src))
if(!compile_kernel(kernel_path,
init_kernel_source,
build_flags,
&cpProgram,
debug_src))
{
return false;
}
/* Save binary for reuse. */
if(!save_binary(&cpProgram, clbin))
if(!save_binary(&cpProgram, clbin)) {
return false;
}
}
/* Cache the program. */
OpenCLCache::store_program(cpPlatform,
cdDevice,
cpProgram,
OpenCLCache::OCL_DEV_BASE_PROGRAM,
cache_locker);
store_cached_kernel(cpPlatform,
cdDevice,
cpProgram,
OpenCLCache::OCL_DEV_BASE_PROGRAM,
cache_locker);
}
else {
VLOG(2) << "Found cached OpenCL kernel.";
}
/* Find kernels. */
@ -1032,9 +1049,9 @@ public:
}
if(ckFilmConvertByteKernel)
clReleaseKernel(ckFilmConvertByteKernel);
clReleaseKernel(ckFilmConvertByteKernel);
if(ckFilmConvertHalfFloatKernel)
clReleaseKernel(ckFilmConvertHalfFloatKernel);
clReleaseKernel(ckFilmConvertHalfFloatKernel);
if(ckShaderKernel)
clReleaseKernel(ckShaderKernel);
if(ckBakeKernel)
@ -1354,7 +1371,6 @@ public:
virtual void thread_run(DeviceTask * /*task*/) = 0;
protected:
string kernel_build_options(const string *debug_src = NULL)
{
string build_options = " -cl-fast-relaxed-math ";
@ -1546,6 +1562,44 @@ protected:
}
return build_options;
}
/* ** Those guys are for workign around some compiler-specific bugs ** */
virtual cl_program load_cached_kernel(
const DeviceRequestedFeatures& /*requested_features*/,
OpenCLCache::ProgramName program_name,
thread_scoped_lock& cache_locker)
{
return OpenCLCache::get_program(cpPlatform,
cdDevice,
program_name,
cache_locker);
}
virtual void store_cached_kernel(cl_platform_id platform,
cl_device_id device,
cl_program program,
OpenCLCache::ProgramName program_name,
thread_scoped_lock& cache_locker)
{
OpenCLCache::store_program(platform,
device,
program,
program_name,
cache_locker);
}
virtual string build_options_for_base_program(
const DeviceRequestedFeatures& /*requested_features*/)
{
/* TODO(sergey): By default we compile all features, meaning
* mega kernel is not getting feature-based optimizations.
*
* Ideally we need always compile kernel with as less features
* enabed as possible to keep performance at it's max.
*/
return "";
}
};
class OpenCLDeviceMegaKernel : public OpenCLDeviceBase
@ -3490,6 +3544,34 @@ protected:
opencl_assert_err(ciErr, "clCreateBuffer");
return ptr;
}
/* ** Those guys are for workign around some compiler-specific bugs ** */
cl_program load_cached_kernel(
const DeviceRequestedFeatures& /*requested_features*/,
OpenCLCache::ProgramName /*program_name*/,
thread_scoped_lock /*cache_locker*/)
{
VLOG(2) << "Skip loading kernel from cache, "
<< "not supported by split kernel.";
return NULL;
}
void store_cached_kernel(cl_platform_id /*platform*/,
cl_device_id /*device*/,
cl_program /*program*/,
OpenCLCache::ProgramName /*program_name*/,
thread_scoped_lock& /*slot_locker*/)
{
VLOG(2) << "Skip storing kernel in cache, "
<< "not supported by split kernel.";
}
string build_options_for_base_program(
const DeviceRequestedFeatures& requested_features)
{
return build_options_from_requested_features(requested_features);
}
};
Device *device_opencl_create(DeviceInfo& info, Stats &stats, bool background)