diff options
author | Brecht Van Lommel <brecht@blender.org> | 2020-06-01 01:11:17 +0300 |
---|---|---|
committer | Brecht Van Lommel <brecht@blender.org> | 2020-06-24 16:17:36 +0300 |
commit | 669befdfbe487f76c65f54e3da0013d140d56893 (patch) | |
tree | 9aa766cca084b9d65fbc694f5fa0a08cbcbcce70 /intern/cycles/device | |
parent | 0a3bde63006c66b8b8531ed5eccca9bdf5e5dc20 (diff) |
Cycles: add Intel OpenImageDenoise support for viewport denoising
Compared to Optix denoise, this is usually slower since there is no GPU
acceleration. Some optimizations may still be possible, in avoid copies
to the GPU and/or denoising less often.
The main thing is that this adds viewport denoising support for computers
without an NVIDIA GPU (as long as the CPU supports SSE 4.1, which is nearly
all of them).
Ref T76259
Diffstat (limited to 'intern/cycles/device')
-rw-r--r-- | intern/cycles/device/CMakeLists.txt | 12 | ||||
-rw-r--r-- | intern/cycles/device/device.cpp | 12 | ||||
-rw-r--r-- | intern/cycles/device/device_cpu.cpp | 108 | ||||
-rw-r--r-- | intern/cycles/device/device_task.h | 1 |
4 files changed, 123 insertions, 10 deletions
diff --git a/intern/cycles/device/CMakeLists.txt b/intern/cycles/device/CMakeLists.txt index aa5b65a2b73..ca366722eb7 100644 --- a/intern/cycles/device/CMakeLists.txt +++ b/intern/cycles/device/CMakeLists.txt @@ -99,6 +99,18 @@ if(WITH_CYCLES_DEVICE_MULTI) add_definitions(-DWITH_MULTI) endif() +if(WITH_OPENIMAGEDENOISE) + add_definitions(-DWITH_OPENIMAGEDENOISE) + add_definitions(-DOIDN_STATIC_LIB) + list(APPEND INC_SYS + ${OPENIMAGEDENOISE_INCLUDE_DIRS} + ) + list(APPEND LIB + ${OPENIMAGEDENOISE_LIBRARIES} + ${TBB_LIBRARIES} + ) +endif() + include_directories(${INC}) include_directories(SYSTEM ${INC_SYS}) diff --git a/intern/cycles/device/device.cpp b/intern/cycles/device/device.cpp index 73415d5f9c6..9dbb33980b4 100644 --- a/intern/cycles/device/device.cpp +++ b/intern/cycles/device/device.cpp @@ -706,6 +706,18 @@ void DeviceInfo::add_denoising_devices(DenoiserType denoiser_type) denoisers = denoiser_type; } } + else if (denoiser_type == DENOISER_OPENIMAGEDENOISE && type != DEVICE_CPU) { + /* Convert to a special multi device with separate denoising devices. */ + if (multi_devices.empty()) { + multi_devices.push_back(*this); + } + + /* Add CPU denoising devices. */ + DeviceInfo cpu_device = Device::available_devices(DEVICE_MASK_CPU).front(); + denoising_devices.push_back(cpu_device); + + denoisers = denoiser_type; + } } CCL_NAMESPACE_END diff --git a/intern/cycles/device/device_cpu.cpp b/intern/cycles/device/device_cpu.cpp index 1f760a15530..8f68e66a1b4 100644 --- a/intern/cycles/device/device_cpu.cpp +++ b/intern/cycles/device/device_cpu.cpp @@ -51,6 +51,7 @@ #include "util/util_function.h" #include "util/util_logging.h" #include "util/util_map.h" +#include "util/util_openimagedenoise.h" #include "util/util_opengl.h" #include "util/util_optimization.h" #include "util/util_progress.h" @@ -177,6 +178,10 @@ class CPUDevice : public Device { #ifdef WITH_OSL OSLGlobals osl_globals; #endif +#ifdef WITH_OPENIMAGEDENOISE + oidn::DeviceRef oidn_device; + oidn::FilterRef oidn_filter; +#endif bool use_split_kernel; @@ -943,6 +948,70 @@ class CPUDevice : public Device { } } + void denoise_openimagedenoise(DeviceTask &task, RenderTile &rtile) + { +#ifdef WITH_OPENIMAGEDENOISE + assert(openimagedenoise_supported()); + + /* Only one at a time, since OpenImageDenoise itself is multithreaded. */ + static thread_mutex mutex; + thread_scoped_lock lock(mutex); + + /* Create device and filter, cached for reuse. */ + if (!oidn_device) { + oidn_device = oidn::newDevice(); + oidn_device.commit(); + } + if (!oidn_filter) { + oidn_filter = oidn_device.newFilter("RT"); + } + + /* Copy pixels from compute device to CPU (no-op for CPU device). */ + rtile.buffers->buffer.copy_from_device(); + + /* Set images with appropriate stride for our interleaved pass storage. */ + const struct { + const char *name; + int offset; + } passes[] = {{"color", task.pass_denoising_data + DENOISING_PASS_COLOR}, + {"normal", task.pass_denoising_data + DENOISING_PASS_NORMAL}, + {"albedo", task.pass_denoising_data + DENOISING_PASS_ALBEDO}, + {"output", 0}, + { NULL, + 0 }}; + + for (int i = 0; passes[i].name; i++) { + const int64_t offset = rtile.offset + rtile.x + rtile.y * rtile.stride; + const int64_t buffer_offset = (offset * task.pass_stride + passes[i].offset) * sizeof(float); + const int64_t pixel_stride = task.pass_stride * sizeof(float); + const int64_t row_stride = rtile.stride * pixel_stride; + + oidn_filter.setImage(passes[i].name, + (char *)rtile.buffer + buffer_offset, + oidn::Format::Float3, + rtile.w, + rtile.h, + 0, + pixel_stride, + row_stride); + } + + /* Execute filter. */ + oidn_filter.set("hdr", true); + oidn_filter.set("srgb", false); + oidn_filter.commit(); + oidn_filter.execute(); + + /* todo: it may be possible to avoid this copy, but we have to ensure that + * when other code copies data from the device it doesn't overwrite the + * denoiser buffers. */ + rtile.buffers->buffer.copy_to_device(); +#else + (void)task; + (void)rtile; +#endif + } + void denoise_nlm(DenoisingTask &denoising, RenderTile &tile) { ProfilingHelper profiling(denoising.profiler, PROFILING_DENOISING); @@ -1018,7 +1087,10 @@ class CPUDevice : public Device { render(task, tile, kg); } else if (tile.task == RenderTile::DENOISE) { - if (task.denoising.type == DENOISER_NLM) { + if (task.denoising.type == DENOISER_OPENIMAGEDENOISE) { + denoise_openimagedenoise(task, tile); + } + else if (task.denoising.type == DENOISER_NLM) { if (denoising == NULL) { denoising = new DenoisingTask(this, task); denoising->profiler = &kg->profiler; @@ -1060,16 +1132,22 @@ class CPUDevice : public Device { tile.stride = task.stride; tile.buffers = task.buffers; - DenoisingTask denoising(this, task); + if (task.denoising.type == DENOISER_OPENIMAGEDENOISE) { + denoise_openimagedenoise(task, tile); + } + else { + DenoisingTask denoising(this, task); - ProfilingState denoising_profiler_state; - profiler.add_state(&denoising_profiler_state); - denoising.profiler = &denoising_profiler_state; + ProfilingState denoising_profiler_state; + profiler.add_state(&denoising_profiler_state); + denoising.profiler = &denoising_profiler_state; - denoise_nlm(denoising, tile); - task.update_progress(&tile, tile.w * tile.h); + denoise_nlm(denoising, tile); - profiler.remove_state(&denoising_profiler_state); + profiler.remove_state(&denoising_profiler_state); + } + + task.update_progress(&tile, tile.w * tile.h); } void thread_film_convert(DeviceTask &task) @@ -1143,10 +1221,17 @@ class CPUDevice : public Device { /* split task into smaller ones */ list<DeviceTask> tasks; - if (task.type == DeviceTask::SHADER) + if (task.type == DeviceTask::DENOISE_BUFFER && + task.denoising.type == DENOISER_OPENIMAGEDENOISE) { + /* Denoise entire buffer at once with OIDN, it has own threading. */ + tasks.push_back(task); + } + else if (task.type == DeviceTask::SHADER) { task.split(tasks, info.cpu_threads, 256); - else + } + else { task.split(tasks, info.cpu_threads); + } foreach (DeviceTask &task, tasks) { task_pool.push([=] { @@ -1351,6 +1436,9 @@ void device_cpu_info(vector<DeviceInfo> &devices) info.has_half_images = true; info.has_profiling = true; info.denoisers = DENOISER_NLM; + if (openimagedenoise_supported()) { + info.denoisers |= DENOISER_OPENIMAGEDENOISE; + } devices.insert(devices.begin(), info); } diff --git a/intern/cycles/device/device_task.h b/intern/cycles/device/device_task.h index 1ad8e0d9485..aeb5ff72e89 100644 --- a/intern/cycles/device/device_task.h +++ b/intern/cycles/device/device_task.h @@ -34,6 +34,7 @@ class Tile; enum DenoiserType { DENOISER_NLM = 1, DENOISER_OPTIX = 2, + DENOISER_OPENIMAGEDENOISE = 4, DENOISER_NUM, DENOISER_NONE = 0, |