[Bf-blender-cvs] [a7cc6e015cf] master: Cycles: Additional Metal kernel specialisation exposed through UI
Michael Jones
noreply at git.blender.org
Thu Jan 5 00:36:54 CET 2023
Commit: a7cc6e015cf98feca22cfc08c3356807b989a9fe
Author: Michael Jones
Date: Wed Jan 4 16:01:24 2023 +0000
Branches: master
https://developer.blender.org/rBa7cc6e015cf98feca22cfc08c3356807b989a9fe
Cycles: Additional Metal kernel specialisation exposed through UI
This patch adds a new "Kernel Optimization Level" dropdown menu to control Metal kernel specialisation. Currently this defaults to "full" optimisation, on the assumption that the changes proposed in D16371 will address usability concerns around app responsiveness and shader cache housekeeping.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16514
===================================================================
M intern/cycles/blender/addon/properties.py
M intern/cycles/blender/device.cpp
M intern/cycles/blender/device.h
M intern/cycles/blender/python.cpp
M intern/cycles/blender/sync.cpp
M intern/cycles/device/device.h
M intern/cycles/device/metal/device_impl.mm
M intern/cycles/device/metal/kernel.h
M intern/cycles/device/metal/kernel.mm
M intern/cycles/device/metal/queue.mm
===================================================================
diff --git a/intern/cycles/blender/addon/properties.py b/intern/cycles/blender/addon/properties.py
index eff6384c85e..a27a75e48fa 100644
--- a/intern/cycles/blender/addon/properties.py
+++ b/intern/cycles/blender/addon/properties.py
@@ -1543,6 +1543,17 @@ class CyclesPreferences(bpy.types.AddonPreferences):
default=False,
)
+ kernel_optimization_level: EnumProperty(
+ name="Kernel Optimization",
+ description="Kernels can be optimized based on scene content. Optimized kernels are requested at the start of a render. If optimized kernels are not available, rendering will proceed using generic kernels until the optimized set is available in the cache. This can result in additional CPU usage for a brief time (tens of seconds).",
+ default='FULL',
+ items=(
+ ('OFF', "Off", "Disable kernel optimization. Slowest rendering, no extra background CPU usage"),
+ ('INTERSECT', "Intersection only", "Optimize only intersection kernels. Faster rendering, negligible extra background CPU usage"),
+ ('FULL', "Full", "Optimize all kernels. Fastest rendering, may result in extra background CPU usage"),
+ ),
+ )
+
def find_existing_device_entry(self, device):
for device_entry in self.devices:
if device_entry.id == device[2] and device_entry.type == device[1]:
@@ -1711,10 +1722,12 @@ class CyclesPreferences(bpy.types.AddonPreferences):
if compute_device_type == 'METAL':
import platform
# MetalRT only works on Apple Silicon at present, pending argument encoding fixes on AMD
+ # Kernel specialization is only viable on Apple Silicon at present due to relative compilation speed
if platform.machine() == 'arm64':
- row = layout.row()
- row.use_property_split = True
- row.prop(self, "use_metalrt")
+ col = layout.column()
+ col.use_property_split = True
+ col.prop(self, "kernel_optimization_level")
+ col.prop(self, "use_metalrt")
def draw(self, context):
self.draw_impl(self.layout, context)
diff --git a/intern/cycles/blender/device.cpp b/intern/cycles/blender/device.cpp
index 22beca898f1..96e7bdd03aa 100644
--- a/intern/cycles/blender/device.cpp
+++ b/intern/cycles/blender/device.cpp
@@ -30,7 +30,7 @@ int blender_device_threads(BL::Scene &b_scene)
return 0;
}
-DeviceInfo blender_device_info(BL::Preferences &b_preferences, BL::Scene &b_scene, bool background)
+DeviceInfo blender_device_info(BL::Preferences &b_preferences, BL::Scene &b_scene, bool background, bool preview)
{
PointerRNA cscene = RNA_pointer_get(&b_scene.ptr, "cycles");
@@ -113,6 +113,18 @@ DeviceInfo blender_device_info(BL::Preferences &b_preferences, BL::Scene &b_scen
device.use_metalrt = true;
}
+ if (preview) {
+ /* Disable specialization for preview renders. */
+ device.kernel_optimization_level = KERNEL_OPTIMIZATION_LEVEL_OFF;
+ }
+ else {
+ device.kernel_optimization_level = (KernelOptimizationLevel)get_enum(
+ cpreferences,
+ "kernel_optimization_level",
+ KERNEL_OPTIMIZATION_NUM_LEVELS,
+ KERNEL_OPTIMIZATION_LEVEL_FULL);
+ }
+
return device;
}
diff --git a/intern/cycles/blender/device.h b/intern/cycles/blender/device.h
index 7a762261829..08655743eeb 100644
--- a/intern/cycles/blender/device.h
+++ b/intern/cycles/blender/device.h
@@ -19,7 +19,8 @@ int blender_device_threads(BL::Scene &b_scene);
/* Convert Blender settings to device specification. */
DeviceInfo blender_device_info(BL::Preferences &b_preferences,
BL::Scene &b_scene,
- bool background);
+ bool background,
+ bool preview);
CCL_NAMESPACE_END
diff --git a/intern/cycles/blender/python.cpp b/intern/cycles/blender/python.cpp
index cfc7a78143c..96cb204be4b 100644
--- a/intern/cycles/blender/python.cpp
+++ b/intern/cycles/blender/python.cpp
@@ -754,7 +754,7 @@ static PyObject *denoise_func(PyObject * /*self*/, PyObject *args, PyObject *key
RNA_id_pointer_create((ID *)PyLong_AsVoidPtr(pyscene), &sceneptr);
BL::Scene b_scene(sceneptr);
- DeviceInfo device = blender_device_info(b_preferences, b_scene, true);
+ DeviceInfo device = blender_device_info(b_preferences, b_scene, true, true);
/* Get denoising parameters from view layer. */
PointerRNA viewlayerptr;
diff --git a/intern/cycles/blender/sync.cpp b/intern/cycles/blender/sync.cpp
index d87d094dc56..45fe4334f06 100644
--- a/intern/cycles/blender/sync.cpp
+++ b/intern/cycles/blender/sync.cpp
@@ -866,7 +866,7 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine &b_engine,
/* Device */
params.threads = blender_device_threads(b_scene);
- params.device = blender_device_info(b_preferences, b_scene, params.background);
+ params.device = blender_device_info(b_preferences, b_scene, params.background, b_engine.is_preview());
/* samples */
int samples = get_int(cscene, "samples");
diff --git a/intern/cycles/device/device.h b/intern/cycles/device/device.h
index ad625fc5a47..3923698b1cd 100644
--- a/intern/cycles/device/device.h
+++ b/intern/cycles/device/device.h
@@ -57,6 +57,14 @@ enum DeviceTypeMask {
#define DEVICE_MASK(type) (DeviceTypeMask)(1 << type)
+enum KernelOptimizationLevel {
+ KERNEL_OPTIMIZATION_LEVEL_OFF = 0,
+ KERNEL_OPTIMIZATION_LEVEL_INTERSECT = 1,
+ KERNEL_OPTIMIZATION_LEVEL_FULL = 2,
+
+ KERNEL_OPTIMIZATION_NUM_LEVELS
+};
+
class DeviceInfo {
public:
DeviceType type;
@@ -66,13 +74,15 @@ class DeviceInfo {
bool display_device; /* GPU is used as a display device. */
bool has_nanovdb; /* Support NanoVDB volumes. */
bool has_light_tree; /* Support light tree. */
- bool has_osl; /* Support Open Shading Language. */
- bool has_guiding; /* Support path guiding. */
- bool has_profiling; /* Supports runtime collection of profiling info. */
- bool has_peer_memory; /* GPU has P2P access to memory of another GPU. */
- bool has_gpu_queue; /* Device supports GPU queue. */
- bool use_metalrt; /* Use MetalRT to accelerate ray queries (Metal only). */
- DenoiserTypeMask denoisers; /* Supported denoiser types. */
+ bool has_osl; /* Support Open Shading Language. */
+ bool has_guiding; /* Support path guiding. */
+ bool has_profiling; /* Supports runtime collection of profiling info. */
+ bool has_peer_memory; /* GPU has P2P access to memory of another GPU. */
+ bool has_gpu_queue; /* Device supports GPU queue. */
+ bool use_metalrt; /* Use MetalRT to accelerate ray queries (Metal only). */
+ KernelOptimizationLevel kernel_optimization_level; /* Optimization level applied to path tracing
+ kernels (Metal only). */
+ DenoiserTypeMask denoisers; /* Supported denoiser types. */
int cpu_threads;
vector<DeviceInfo> multi_devices;
string error_msg;
diff --git a/intern/cycles/device/metal/device_impl.mm b/intern/cycles/device/metal/device_impl.mm
index a6966bf167d..01578155931 100644
--- a/intern/cycles/device/metal/device_impl.mm
+++ b/intern/cycles/device/metal/device_impl.mm
@@ -110,10 +110,6 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
case METAL_GPU_APPLE: {
max_threads_per_threadgroup = 512;
use_metalrt = info.use_metalrt;
-
- /* Specialize the intersection kernels on Apple GPUs by default as these can be built very
- * quickly. */
- kernel_specialization_level = PSO_SPECIALIZED_INTERSECT;
break;
}
}
@@ -126,6 +122,22 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
capture_enabled = true;
}
+ if (device_vendor == METAL_GPU_APPLE) {
+ /* Set kernel_specialization_level based on user prefs. */
+ switch (info.kernel_optimization_level) {
+ case KERNEL_OPTIMIZATION_LEVEL_OFF:
+ kernel_specialization_level = PSO_GENERIC;
+ break;
+ default:
+ case KERNEL_OPTIMIZATION_LEVEL_INTERSECT:
+ kernel_specialization_level = PSO_SPECIALIZED_INTERSECT;
+ break;
+ case KERNEL_OPTIMIZATION_LEVEL_FULL:
+ kernel_specialization_level = PSO_SPECIALIZED_SHADE;
+ break;
+ }
+ }
+
if (auto envstr = getenv("CYCLES_METAL_SPECIALIZATION_LEVEL")) {
kernel_specialization_level = (MetalPipelineType)atoi(envstr);
}
@@ -444,7 +456,7 @@ void MetalDevice::compile_and_load(int device_id, MetalPipelineType pso_type)
source);
}
- const double starttime = time_dt();
+ double starttime = time_dt();
NSError *error = NULL;
id<MTLLibrary> mtlLibrary = [mtlDevice newLibraryWithSource:@(source.c_str())
@@ -457,6 +469,12 @@ void MetalDevice::compile_and_load(int device_id, MetalPipelineType pso_type)
[options release];
+ bool blocking_pso_build = (getenv("CYCLES_METAL_PROFILING") || MetalDeviceKernels::is_benchmark_warmup());
+ if (blocking_pso_build) {
+ MetalDeviceKernels::wait_for_all();
+ starttime = 0.0;
+ }
+
/* Save the compiled MTLLibrary and trigger the AIR->PSO builds (if the MetalDevice still
* exists). */
{
@@ -464,6 +482,8 @@ void MetalDevice::compile_and_load(int device_id, MetalPipelineType pso_type)
if (MetalDevice *instance = get_device_by_ID(device_id, lock)) {
if (mtlLibrary) {
instance->mtlLibrary[pso_type] = mtlLibrary;
+
+ starttime = time_dt();
MetalDeviceKernels::load(instance, pso_type);
}
else {
@@ -472,6 +492,14 @@ void MetalDevice::compile_and_load(int device_id, MetalPipelineType pso_type)
}
}
}
+
+ if (starttime && blocking_pso_build) {
+ MetalDeviceKernels::wait_for_all();
+
+ metal_printf("Back-end compilation finished in %.1f seconds (%s)\n",
+ time_dt() - starttime,
+ kernel_type_as_string(pso_type));
+ }
}
void MetalDevice::load_texture_info()
@@ -832,10 +860,8 @@ void MetalDevice
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list