[Bf-blender-cvs] [817873cc83] master: Cycles: CUDA implementation of split kernel
Mai Lavelle
noreply at git.blender.org
Wed Mar 8 07:53:23 CET 2017
Commit: 817873cc83034c460f1be6bf410c95ff009f3ae2
Author: Mai Lavelle
Date: Tue Feb 14 05:50:29 2017 -0500
Branches: master
https://developer.blender.org/rB817873cc83034c460f1be6bf410c95ff009f3ae2
Cycles: CUDA implementation of split kernel
===================================================================
M intern/cycles/blender/addon/properties.py
M intern/cycles/blender/addon/ui.py
M intern/cycles/blender/blender_python.cpp
M intern/cycles/device/device_cuda.cpp
M intern/cycles/kernel/CMakeLists.txt
M intern/cycles/kernel/kernel_compat_cuda.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/kernels/cuda/kernel.cu
A intern/cycles/kernel/kernels/cuda/kernel_config.h
A intern/cycles/kernel/kernels/cuda/kernel_split.cu
M intern/cycles/kernel/split/kernel_split_data.h
M intern/cycles/util/util_debug.cpp
M intern/cycles/util/util_debug.h
===================================================================
diff --git a/intern/cycles/blender/addon/properties.py b/intern/cycles/blender/addon/properties.py
index 1f0b712c93..ca10973431 100644
--- a/intern/cycles/blender/addon/properties.py
+++ b/intern/cycles/blender/addon/properties.py
@@ -668,6 +668,7 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
cls.debug_use_cpu_split_kernel = BoolProperty(name="Split Kernel", default=False)
cls.debug_use_cuda_adaptive_compile = BoolProperty(name="Adaptive Compile", default=False)
+ cls.debug_use_cuda_split_kernel = BoolProperty(name="Split Kernel", default=False)
cls.debug_opencl_kernel_type = EnumProperty(
name="OpenCL Kernel Type",
diff --git a/intern/cycles/blender/addon/ui.py b/intern/cycles/blender/addon/ui.py
index 8d3fe87759..7c1e3e270f 100644
--- a/intern/cycles/blender/addon/ui.py
+++ b/intern/cycles/blender/addon/ui.py
@@ -1523,6 +1523,7 @@ class CyclesRender_PT_debug(CyclesButtonsPanel, Panel):
col = layout.column()
col.label('CUDA Flags:')
col.prop(cscene, "debug_use_cuda_adaptive_compile")
+ col.prop(cscene, "debug_use_cuda_split_kernel")
col = layout.column()
col.label('OpenCL Flags:')
diff --git a/intern/cycles/blender/blender_python.cpp b/intern/cycles/blender/blender_python.cpp
index ed410e15e7..75118c4374 100644
--- a/intern/cycles/blender/blender_python.cpp
+++ b/intern/cycles/blender/blender_python.cpp
@@ -70,6 +70,7 @@ bool debug_flags_sync_from_scene(BL::Scene b_scene)
flags.cpu.split_kernel = get_boolean(cscene, "debug_use_cpu_split_kernel");
/* Synchronize CUDA flags. */
flags.cuda.adaptive_compile = get_boolean(cscene, "debug_use_cuda_adaptive_compile");
+ flags.cuda.split_kernel = get_boolean(cscene, "debug_use_cuda_split_kernel");
/* Synchronize OpenCL kernel type. */
switch(get_enum(cscene, "debug_opencl_kernel_type")) {
case 0:
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index 1e5ce7875b..74f36022b3 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -22,6 +22,7 @@
#include "device.h"
#include "device_intern.h"
+#include "device_split_kernel.h"
#include "buffers.h"
@@ -43,6 +44,8 @@
#include "util_types.h"
#include "util_time.h"
+#include "split/kernel_split_data.h"
+
CCL_NAMESPACE_BEGIN
#ifndef WITH_CUDA_DYNLOAD
@@ -79,6 +82,29 @@ int cuewCompilerVersion(void)
} /* namespace */
#endif /* WITH_CUDA_DYNLOAD */
+class CUDADevice;
+
+class CUDASplitKernel : public DeviceSplitKernel {
+ CUDADevice *device;
+public:
+ explicit CUDASplitKernel(CUDADevice *device);
+
+ virtual bool enqueue_split_kernel_data_init(const KernelDimensions& dim,
+ RenderTile& rtile,
+ int num_global_elements,
+ device_memory& kernel_globals,
+ device_memory& kernel_data_,
+ device_memory& split_data,
+ device_memory& ray_state,
+ device_memory& queue_index,
+ device_memory& use_queues_flag,
+ device_memory& work_pool_wgs);
+
+ virtual SplitKernelFunction* get_split_kernel_function(string kernel_name, const DeviceRequestedFeatures&);
+ virtual int2 split_kernel_local_size();
+ virtual int2 split_kernel_global_size(DeviceTask *task);
+};
+
class CUDADevice : public Device
{
public:
@@ -259,11 +285,16 @@ public:
return DebugFlags().cuda.adaptive_compile;
}
+ bool use_split_kernel()
+ {
+ return DebugFlags().cuda.split_kernel;
+ }
+
/* Common NVCC flags which stays the same regardless of shading model,
* kernel sources md5 and only depends on compiler or compilation settings.
*/
string compile_kernel_get_common_cflags(
- const DeviceRequestedFeatures& requested_features)
+ const DeviceRequestedFeatures& requested_features, bool split=false)
{
const int cuda_version = cuewCompilerVersion();
const int machine = system_cpu_bits();
@@ -288,6 +319,11 @@ public:
#ifdef WITH_CYCLES_DEBUG
cflags += " -D__KERNEL_DEBUG__";
#endif
+
+ if(split) {
+ cflags += " -D__SPLIT__";
+ }
+
return cflags;
}
@@ -321,7 +357,7 @@ public:
return true;
}
- string compile_kernel(const DeviceRequestedFeatures& requested_features)
+ string compile_kernel(const DeviceRequestedFeatures& requested_features, bool split=false)
{
/* Compute cubin name. */
int major, minor;
@@ -330,7 +366,8 @@ public:
/* Attempt to use kernel provided with Blender. */
if(!use_adaptive_compilation()) {
- const string cubin = path_get(string_printf("lib/kernel_sm_%d%d.cubin",
+ const string cubin = path_get(string_printf(split ? "lib/kernel_split_sm_%d%d.cubin"
+ : "lib/kernel_sm_%d%d.cubin",
major, minor));
VLOG(1) << "Testing for pre-compiled kernel " << cubin << ".";
if(path_exists(cubin)) {
@@ -340,7 +377,7 @@ public:
}
const string common_cflags =
- compile_kernel_get_common_cflags(requested_features);
+ compile_kernel_get_common_cflags(requested_features, split);
/* Try to use locally compiled kernel. */
const string kernel_path = path_get("kernel");
@@ -351,7 +388,8 @@ public:
*/
const string cubin_md5 = util_md5_string(kernel_md5 + common_cflags);
- const string cubin_file = string_printf("cycles_kernel_sm%d%d_%s.cubin",
+ const string cubin_file = string_printf(split ? "cycles_kernel_split_sm%d%d_%s.cubin"
+ : "cycles_kernel_sm%d%d_%s.cubin",
major, minor,
cubin_md5.c_str());
const string cubin = path_cache_get(path_join("kernels", cubin_file));
@@ -386,7 +424,7 @@ public:
const char *nvcc = cuewCompilerPath();
const string kernel = path_join(kernel_path,
path_join("kernels",
- path_join("cuda", "kernel.cu")));
+ path_join("cuda", split ? "kernel_split.cu" : "kernel.cu")));
double starttime = time_dt();
printf("Compiling CUDA kernel ...\n");
@@ -434,7 +472,7 @@ public:
return false;
/* get kernel */
- string cubin = compile_kernel(requested_features);
+ string cubin = compile_kernel(requested_features, use_split_kernel());
if(cubin == "")
return false;
@@ -1261,25 +1299,48 @@ public:
/* Upload Bindless Mapping */
load_bindless_mapping();
- /* keep rendering tiles until done */
- while(task->acquire_tile(this, tile)) {
- int start_sample = tile.start_sample;
- int end_sample = tile.start_sample + tile.num_samples;
+ if(!use_split_kernel()) {
+ /* keep rendering tiles until done */
+ while(task->acquire_tile(this, tile)) {
+ int start_sample = tile.start_sample;
+ int end_sample = tile.start_sample + tile.num_samples;
- for(int sample = start_sample; sample < end_sample; sample++) {
- if(task->get_cancel()) {
- if(task->need_finish_queue == false)
- break;
- }
+ for(int sample = start_sample; sample < end_sample; sample++) {
+ if(task->get_cancel()) {
+ if(task->need_finish_queue == false)
+ break;
+ }
- path_trace(tile, sample, branched);
+ path_trace(tile, sample, branched);
- tile.sample = sample + 1;
+ tile.sample = sample + 1;
- task->update_progress(&tile, tile.w*tile.h);
+ task->update_progress(&tile, tile.w*tile.h);
+ }
+
+ task->release_tile(tile);
}
+ }
+ else {
+ DeviceRequestedFeatures requested_features;
+ if(!use_adaptive_compilation()) {
+ requested_features.max_closure = 64;
+ }
+
+ CUDASplitKernel split_kernel(this);
+ split_kernel.load_kernels(requested_features);
+
+ while(task->acquire_tile(this, tile)) {
+ device_memory void_buffer;
+ split_kernel.path_trace(task, tile, void_buffer, void_buffer);
+
+ task->release_tile(tile);
- task->release_tile(tile);
+ if(task->get_cancel()) {
+ if(task->need_finish_queue == false)
+ break;
+ }
+ }
}
}
else if(task->type == DeviceTask::SHADER) {
@@ -1332,8 +1393,186 @@ public:
{
task_pool.cancel();
}
+
+ friend class CUDASplitKernelFunction;
+ friend class CUDASplitKernel;
};
+/* redefine the cuda_assert macro so it can be used outside of the CUDADevice class
+ * now that the definition of that class is complete
+ */
+#undef cuda_assert
+#define cuda_assert(stmt) \
+ { \
+ CUresult result = stmt; \
+ \
+ if(result != CUDA_SUCCESS) { \
+ string message = string_printf("CUDA error: %s in %s", cuewErrorString(result), #stmt); \
+ if(device->error_msg == "") \
+ device->error_msg = message; \
+ fprintf(stderr, "%s\n", message.c_str()); \
+ /*cuda_abort();*/ \
+ device->cuda_error_documentation(); \
+ } \
+ } (void)0
+
+/* split kernel */
+
+class CUDASplitKernelFunction : public SplitKernelFunction{
+ CUDADevice* device;
+ CUfunction func;
+public:
+ CUDASplitKernelFunction(CUDADevice *device, CUfunction func) : device(device), func(func) {}
+
+ /* enqueue the kernel, returns false if there is an error */
+ bool enqueue(const KernelDimensions &dim, device_memory &/*kg*/, device_memory &/*data*/)
+ {
+ return enqueue(dim, NULL);
+ }
+
+ /* enqueue the kernel, returns false if there is an error */
+ bool enqueue(const KernelDimensions &dim, void *args[])
+ {
+ device->cuda_push_context();
+
+ if(device->have_error())
+ return false;
+
+ /* we ignore dim.local_size for now, as this is faster */
+ int threads_per_block;
+ cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, func));
+
+ int xthreads = (int)sqrt(threads_per_block);
+ int ythreads = (int)sqrt(threads_per_block);
+
+ int xblocks = (dim.global_size[0] + xthreads - 1)/xthreads;
+ int yblocks = (d
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list