Add support for instance gas update, dynamic geometry example now working as expected

Author: keckj

examples/dynamic_geometry.py

@@ -1,20 +1,22 @@
-import os, sys, enum, logging, collections
+import os, sys, enum, copy, logging, collections
 
 import numpy as np
 import cupy as cp
 import optix as ox
 import glfw, imgui
 
-from optix.sutils.gui import init_ui, display_stats
-from optix.sutils.gl_display import GLDisplay
-from optix.sutils.trackball import Trackball, TrackballViewMode
-from optix.sutils.cuda_output_buffer import CudaOutputBuffer, CudaOutputBufferType, BufferImageFormat
+from optix.sutil.gui import init_ui, display_stats
+from optix.sutil.gl_display import GLDisplay
+from optix.sutil.trackball import Trackball, TrackballViewMode
+from optix.sutil.cuda_output_buffer import CudaOutputBuffer, CudaOutputBufferType, BufferImageFormat
 
 logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
 log = logging.getLogger()
 
 script_dir = os.path.dirname(os.path.abspath(__file__))
 
+DEBUG=False
+
 #------------------------------------------------------------------------------
 # Local types
 #------------------------------------------------------------------------------
@@ -58,7 +60,8 @@ class DynamicGeometryState:
     __slots__ = ['params', 'time', 'ctx', 'module', 'pipeline', 'pipeline_opts',
             'raygen_grp', 'miss_grp', 'hit_grp', 'sbt',
             'generate_vertices_kernel', 'd_temp_vertices', 'last_exploding_sphere_rebuild_time',
-            'static_gas', 'deforming_gas', 'exploding_gas', 'ias',
+            'gas_build_input', 'static_gas', 'deforming_gas', 'exploding_gas',
+            'ias_build_input', 'ias',
             'trackball', 'camera_changed', 'mouse_button', 'resize_dirty', 'minimized']
 
     def __init__(self):
@@ -77,7 +80,7 @@ def camera(self):
         return self.trackball.camera
 
     @property
-    def dimensions(self):
+    def launch_dimensions(self):
         return (int(self.params.width), int(self.params.height))
 
 class AnimationMode(enum.Enum):
@@ -203,7 +206,7 @@ def update_state(output_buffer, state):
 def launch_subframe(output_buffer, state):
     state.params.frame_buffer = output_buffer.map()
 
-    state.pipeline.launch(state.sbt, dimensions=state.dimensions,
+    state.pipeline.launch(state.sbt, dimensions=state.launch_dimensions,
             params=state.params.handle, stream=output_buffer.stream)
 
     output_buffer.unmap()
@@ -229,7 +232,7 @@ def init_camera_state(state):
 
 def create_context(state):
     logger = ox.Logger(log)
-    ctx = ox.DeviceContext(validation_mode=True, log_callback_function=logger, log_callback_level=4)
+    ctx = ox.DeviceContext(validation_mode=False, log_callback_function=logger, log_callback_level=4)
     ctx.cache_enabled = False
     state.ctx = ctx
 
@@ -246,8 +249,30 @@ def launch_generate_animated_vertices(state, animation_mode):
     generate_animated_vertices(state.d_temp_vertices, animation_mode, state.time, g_tessellation_resolution, g_tessellation_resolution)
 
 def update_mesh_accel(state):
+    # first sphere is static
+
+    # second sphere moves by updating its transform matrix
+    transform = state.ias_build_input.get_transform_view(1)
+    transform[1,-1] = np.sin(4*state.time)
+
+    # third sphere deforms
     launch_generate_animated_vertices(state, AnimationMode.DEFORM)
-    raise NotImplementedError
+    state.deforming_gas.update(state.gas_build_input)
+
+    # fourth sphere explodes
+    launch_generate_animated_vertices(state, AnimationMode.EXPLODE)
+
+    # we occasionally rebuild the exploding sphere to maintain AS quality
+    if state.time - state.last_exploding_sphere_rebuild_time > 1 / g_exploding_gas_rebuild_frequency:
+        state.last_exploding_sphere_rebuild_time = state.time
+        state.exploding_gas = ox.AccelerationStructure(state.ctx, state.gas_build_input,
+                compact=True, allow_update=True, random_vertex_access=True)
+        state.ias_build_input.instances[3].update_traversable(state.exploding_gas)
+        state.ias_build_input.update_instance(3)
+    else:
+        state.exploding_gas.update(state.gas_build_input)
+
+    state.ias.update(state.ias_build_input)
 
 def build_vertex_generation_kernel(state):
     cuda_source = os.path.join(script_dir, 'cuda', 'dynamic_geometry_vertex_generation.cu')
@@ -277,12 +302,15 @@ def build_mesh_accel(state):
 
     # Build an AS over the triangles.
     # We use un-indexed triangles so we can explode the sphere per triangle.
-    build_input = ox.BuildInputTriangleArray(state.d_temp_vertices, flags=[ox.GeometryFlags.NONE])
-    state.static_gas = ox.AccelerationStructure(state.ctx, build_input,
+    state.gas_build_input = ox.BuildInputTriangleArray(state.d_temp_vertices, flags=[ox.GeometryFlags.NONE])
+    state.static_gas = ox.AccelerationStructure(state.ctx, state.gas_build_input,
+            compact=True, allow_update=False, random_vertex_access=True)
+
+    state.deforming_gas = ox.AccelerationStructure(state.ctx, state.gas_build_input,
             compact=True, allow_update=True, random_vertex_access=True)
 
-    state.deforming_gas = state.static_gas
-    state.exploding_gas = state.static_gas
+    state.exploding_gas = ox.AccelerationStructure(state.ctx, state.gas_build_input,
+            compact=True, allow_update=True, random_vertex_access=True)
 
     traversables = [state.static_gas, state.static_gas,
                     state.deforming_gas, state.exploding_gas]
@@ -292,20 +320,25 @@ def build_mesh_accel(state):
                 sbt_offset=i, transform=g_instances[i])
         instances.append(instance)
 
-    build_input = ox.BuildInputInstanceArray(instances)
-    state.ias = ox.AccelerationStructure(context=state.ctx, build_inputs=build_input,
-            compact=True, allow_update=True)
+    state.ias_build_input = ox.BuildInputInstanceArray(instances)
+    state.ias = ox.AccelerationStructure(context=state.ctx,
+            build_inputs=state.ias_build_input, compact=True, allow_update=True)
     state.params.trav_handle = state.ias.handle
 
 
 def create_module(state):
+    if DEBUG:
+        exception_flags=ox.ExceptionFlags.DEBUG | ox.ExceptionFlags.TRACE_DEPTH | ox.ExceptionFlags.STACK_OVERFLOW,
+    else:
+        exception_flags=ox.ExceptionFlags.NONE
+
     pipeline_opts = ox.PipelineCompileOptions(
             uses_motion_blur=False,
             uses_primitive_type_flags = ox.PrimitiveTypeFlags.TRIANGLE,
             traversable_graph_flags=ox.TraversableGraphFlags.ALLOW_SINGLE_LEVEL_INSTANCING,
+            exception_flags=exception_flags,
             num_payload_values=3,
             num_attribute_values=2,
-            exception_flags=ox.ExceptionFlags.DEBUG | ox.ExceptionFlags.TRACE_DEPTH | ox.ExceptionFlags.STACK_OVERFLOW,
             pipeline_launch_params_variable_name="params")
 
     compile_opts = ox.ModuleCompileOptions(
@@ -328,7 +361,7 @@ def create_pipeline(state):
     program_grps = [state.raygen_grp, state.miss_grp, state.hit_grp]
 
     link_opts = ox.PipelineLinkOptions(max_trace_depth=1,
-                                       debug_level=ox.CompileDebugLevel.FULL)
+                                       debug_level=ox.CompileDebugLevel.LINEINFO)
 
     pipeline = ox.Pipeline(state.ctx,
                            compile_options=state.pipeline_opts,
@@ -410,7 +443,7 @@ def create_sbt(state):
 
         state.time = glfw.get_time() - tstart
 
-        #update_mesh_accel(state)
+        update_mesh_accel(state)
 
         update_state(output_buffer, state)
 

optix/build.pxd

@@ -183,7 +183,6 @@ cdef extern from "optix.h" nogil:
         unsigned int flags
         OptixTraversableHandle traversableHandle
 
-
     OptixResult optixAccelComputeMemoryUsage(OptixDeviceContext context,
                                  const OptixAccelBuildOptions * accelOptions,
                                  const OptixBuildInput * buildInputs,
@@ -282,7 +281,7 @@ cdef class Instance(OptixObject):
 
 cdef class BuildInputInstanceArray(BuildInputArray):
     cdef OptixBuildInputInstanceArray build_input
-    cdef object instances
+    cdef public object instances
     cdef object _d_instances
 
 

optix/build.pyx

@@ -436,6 +436,10 @@ cdef class Instance(OptixObject):
             raise ValueError(f"Too many entries in visibility mask. Got {visibility_mask.bit_length()} but supported are only {max_visibility_mask_bits}")
         self.instance.visibilityMask = visibility_mask
 
+    def update_traversable(self, AccelerationStructure traversable):
+        self.traversable = traversable
+        self.instance.traversableHandle = self.traversable.handle
+
     def __deepcopy__(self, memodict={}):
         from copy import deepcopy
         cls = self.__class__
@@ -479,7 +483,15 @@ cdef class BuildInputInstanceArray(BuildInputArray):
 
     cdef size_t num_elements(self):
         return self.build_input.numInstances
+    
+    def update_instance(self, index):
+        src_ptr = <size_t>&((<Instance>(self.instances[index])).instance)
+        dst_ptr = self._d_instances.ptr + index*sizeof(OptixInstance)
+        cp.cuda.runtime.memcpy(dst_ptr, src_ptr, sizeof(OptixInstance), cp.cuda.runtime.memcpyHostToDevice)
 
+    def get_transform_view(self, index):
+        device_ptr = cp.cuda.MemoryPointer(mem=self._d_instances.mem, offset=<int>index*sizeof(OptixInstance))
+        return cp.ndarray(shape=(3,4), dtype=np.float32, memptr=device_ptr)
 
 
 cdef class AccelerationStructure(OptixContextObject):
@@ -740,7 +752,7 @@ cdef class AccelerationStructure(OptixContextObject):
         result._build_flags = self._build_flags
         result._buffer_sizes = self._buffer_sizes
         result._instances = deepcopy(self._instances) # copy all instances and their AccelerationStructures first
-
+    
         buffer_size = round_up(self._buffer_sizes.outputSizeInBytes, 8) + 8
         result._gas_buffer = cp.cuda.alloc(buffer_size)
         cp.cuda.runtime.memcpy(result._gas_buffer.ptr, self._gas_buffer.ptr, buffer_size, cp.cuda.runtime.memcpyDeviceToDevice)

optix/sutils/__init__.py optix/sutil/__init__.pyoptix/sutils/__init__.py renamed to optix/sutil/__init__.py

@@ -1,7 +1,7 @@
 import numpy as np
 
-from optix.sutils.vecmath import length, normalize, cross
-from optix.sutils.properties import get_member, set_float, set_float3
+from optix.sutil.vecmath import length, normalize, cross
+from optix.sutil.properties import get_member, set_float, set_float3
 
 class Camera:
     """Implements a perspective camera."""

optix/sutils/camera.py optix/sutil/camera.pyoptix/sutils/camera.py renamed to optix/sutil/camera.py

@@ -5,7 +5,7 @@
 
 import OpenGL.GL as gl
 
-from optix.sutils.vecmath import vtype_to_dtype
+from optix.sutil.vecmath import vtype_to_dtype
 
 class BufferImageFormat(enum.Enum):
     UCHAR4=0

optix/sutils/cuda_output_buffer.py optix/sutil/cuda_output_buffer.pyoptix/sutils/cuda_output_buffer.py renamed to optix/sutil/cuda_output_buffer.py

@@ -4,7 +4,7 @@
 import OpenGL.GL as gl
 import OpenGL.GL.shaders
 
-from optix.sutils.cuda_output_buffer import BufferImageFormat
+from optix.sutil.cuda_output_buffer import BufferImageFormat
 
 class GLDisplay:
     vert_source = \

optix/sutils/gl_display.py optix/sutil/gl_display.pyoptix/sutils/gl_display.py renamed to optix/sutil/gl_display.py

@@ -65,7 +65,7 @@ def display_stats(state_update_time, render_time, display_time):
     cur_time = glfw.get_time()
 
     display_stats.last_update_frames += 1
-    last_update_time = display_stats.last_update_time or cur_time - 1e-7
+    last_update_time = display_stats.last_update_time or cur_time - 0.5
     last_update_frames = display_stats.last_update_frames
     total_subframe_count = display_stats.total_subframe_count
 

optix/sutils/gui.py optix/sutil/gui.pyoptix/sutils/gui.py renamed to optix/sutil/gui.py

@@ -2,9 +2,9 @@
 
 import numpy as np
 
-from optix.sutils.properties import get_member, set_bool, set_float, set_float3
-from optix.sutils.vecmath import dot, length, normalize
-from optix.sutils.camera import Camera
+from optix.sutil.properties import get_member, set_bool, set_float, set_float3
+from optix.sutil.vecmath import dot, length, normalize
+from optix.sutil.camera import Camera
 
 class TrackballViewMode(enum.Enum):
     EyeFixed = 0