Merge branch 'ershi/1.4.2-rc.3' into 'release-1.4'

Changes for 1.4.2-rc.3

See merge request omniverse/warp!848

CHANGELOG.md

@@ -13,6 +13,8 @@
 - Fix kernel compile error when printing structs.
 - Fix an incorrect user function being sometimes resolved when multiple overloads are available with array parameters with different `dtype` values.
 - Fix error being raised when static and dynamic for-loops are written in sequence with the same iteration variable names ([GH-331](https://github.com/NVIDIA/warp/issues/331)).
+- Fix an issue with the `Texture Write` node, used in the Mandelbrot Omniverse sample, sometimes erroring out in multi-GPU environments.
+- Code generation of in-place multiplication and division operations (regression introduced in a69d061)([GH-342](https://github.com/NVIDIA/warp/issues/342)).
 
 ## [1.4.1] - 2024-10-15
 

VERSION.md

@@ -1 +1 @@
-1.4.2-rc.2
+1.4.2-rc.3

docs/modules/differentiability.rst

@@ -891,6 +891,34 @@ Warp uses a source-code transformation approach to auto-differentiation.
 In this approach, the backwards pass must keep a record of intermediate values computed during the forward pass.
 This imposes some restrictions on what kernels can do if they are to remain differentiable.
 
+In-Place Math Operations
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+In-place addition and subtraction can be used in kernels participating in the backward pass, e.g.
+
+.. code-block:: python
+
+    @wp.kernel
+    def inplace(a: wp.array(dtype=float), b: wp.array(dtype=float)):
+        i = wp.tid()
+
+        a[i] -= b[i]
+
+
+    a = wp.full(10, value=10.0, dtype=float, requires_grad=True)
+    b = wp.full(10, value=2.0, dtype=float, requires_grad=True)
+
+    with wp.Tape() as tape:
+        wp.launch(inplace, a.shape, inputs=[a, b])
+
+    tape.backward(grads={a: wp.ones_like(a)})
+
+    print(a.grad)  # [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
+    print(b.grad)  # [-1. -1. -1. -1. -1. -1. -1. -1. -1. -1.]
+
+In-place multiplication and division are *not* supported and incorrect results will be obtained in the backward pass.
+A warning will be emitted during code generation if ``wp.config.verbose = True``.
+
 Dynamic Loops
 ^^^^^^^^^^^^^
 Currently, dynamic loops are not replayed or unrolled in the backward pass, meaning intermediate values that are

exts/omni.warp.core/config/extension.toml

@@ -1,6 +1,6 @@
 [package]
 # Semantic Versioning is used: https://semver.org/
-version = "1.4.2-rc.2"
+version = "1.4.2-rc.3"
 authors = ["NVIDIA"]
 title = "Warp Core"
 description="The core Warp Python module"

exts/omni.warp.core/docs/CHANGELOG.md

@@ -1,6 +1,6 @@
 # CHANGELOG
 
-## [1.4.2-rc.2] - 2024-11-01
+## [1.4.2-rc.3] - 2024-11-01
 
 ### Changed
 
@@ -13,6 +13,8 @@
 - Fix kernel compile error when printing structs.
 - Fix an incorrect user function being sometimes resolved when multiple overloads are available with array parameters with different `dtype` values.
 - Fix error being raised when static and dynamic for-loops are written in sequence with the same iteration variable names ([GH-331](https://github.com/NVIDIA/warp/issues/331)).
+- Fix an issue with the `Texture Write` node, used in the Mandelbrot Omniverse sample, sometimes erroring out in multi-GPU environments.
+- Code generation of in-place multiplication and division operations (regression introduced in a69d061)([GH-342](https://github.com/NVIDIA/warp/issues/342)).
 
 ## [1.4.1] - 2024-10-15
 

exts/omni.warp/config/extension.toml

@@ -1,6 +1,6 @@
 [package]
 # Semantic Versioning is used: https://semver.org/
-version = "1.4.2-rc.2"
+version = "1.4.2-rc.3"
 authors = ["NVIDIA"]
 title = "Warp"
 description="Warp OmniGraph Nodes and Sample Scenes"
@@ -35,7 +35,7 @@ exclude = ["Ogn*Database.py", "*/ogn*"]
 "omni.timeline" = {}
 "omni.ui" = {optional = true}
 "omni.usd" = {}
-"omni.warp.core" = {version = "1.4.2-rc.2", exact = true}
+"omni.warp.core" = {version = "1.4.2-rc.3", exact = true}
 
 [[python.module]]
 name = "omni.warp._extension"

exts/omni.warp/docs/CHANGELOG.md

@@ -1,6 +1,6 @@
 # CHANGELOG
 
-## [1.4.2-rc.2] - 2024-11-01
+## [1.4.2-rc.3] - 2024-11-01
 
 ### Changed
 
@@ -13,6 +13,8 @@
 - Fix kernel compile error when printing structs.
 - Fix an incorrect user function being sometimes resolved when multiple overloads are available with array parameters with different `dtype` values.
 - Fix error being raised when static and dynamic for-loops are written in sequence with the same iteration variable names ([GH-331](https://github.com/NVIDIA/warp/issues/331)).
+- Fix an issue with the `Texture Write` node, used in the Mandelbrot Omniverse sample, sometimes erroring out in multi-GPU environments.
+- Code generation of in-place multiplication and division operations (regression introduced in a69d061)([GH-342](https://github.com/NVIDIA/warp/issues/342)).
 
 ## [1.4.1] - 2024-10-15
 

exts/omni.warp/omni/warp/nodes/_impl/OgnClothSimulate.py

@@ -351,7 +351,7 @@ def update_collider(
     db: OgnClothSimulateDatabase,
 ) -> None:
     """Updates the collider state."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     points = omni.warp.nodes.mesh_get_points(db.inputs.collider)
     xform = omni.warp.nodes.bundle_get_world_xform(db.inputs.collider)
@@ -397,7 +397,7 @@ def update_cloth(
     db: OgnClothSimulateDatabase,
 ) -> None:
     """Updates the cloth state."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     xform = omni.warp.nodes.bundle_get_world_xform(db.inputs.cloth)
 
@@ -425,7 +425,7 @@ def update_cloth(
 
 def step(db: OgnClothSimulateDatabase) -> None:
     """Steps through the simulation."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     sim_dt = state.sim_dt / db.inputs.substepCount
 
@@ -461,7 +461,7 @@ def step(db: OgnClothSimulateDatabase) -> None:
 
 def simulate(db: OgnClothSimulateDatabase) -> None:
     """Simulates the cloth at the current time."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     if USE_GRAPH:
         wp.capture_launch(state.graph)
@@ -474,7 +474,7 @@ def compute(db: OgnClothSimulateDatabase, device: wp.context.Device) -> None:
     if not db.inputs.cloth.valid or not db.outputs.cloth.valid:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     if not db.inputs.enabled:
         # Pass through the data.
@@ -633,10 +633,10 @@ def compute(db: OgnClothSimulateDatabase) -> None:
                 compute(db, device)
         except Exception:
             db.log_error(traceback.format_exc())
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
         # Fire the execution for the downstream nodes.
         db.outputs.execOut = og.ExecutionAttributeState.ENABLED

exts/omni.warp/omni/warp/nodes/_impl/OgnFixedTime.py

@@ -28,7 +28,7 @@ def compute(db) -> bool:
         """Compute the outputs from the current input"""
 
         timeline = omni.timeline.get_timeline_interface()
-        context = db.internal_state
+        context = db.per_instance_state
 
         if not context.initialized:
             context.time = db.inputs.start

exts/omni.warp/omni/warp/nodes/_impl/OgnGridCreate.py

@@ -49,7 +49,7 @@ def compute(db: OgnGridCreateDatabase) -> None:
     if not db.outputs.mesh.valid:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     if state.is_valid and not state.attr_tracking.have_attrs_changed(db):
         return
@@ -105,10 +105,10 @@ def compute(db: OgnGridCreateDatabase) -> None:
                 compute(db)
         except Exception:
             db.log_error(traceback.format_exc())
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
         # Fire the execution for the downstream nodes.
         db.outputs.execOut = og.ExecutionAttributeState.ENABLED

exts/omni.warp/omni/warp/nodes/_impl/OgnKernel.py

@@ -151,38 +151,38 @@ def compute(db: OgnKernelDatabase, device: wp.context.Device) -> None:
 
     # Ensure that our internal state is correctly initialized.
     timeline = omni.timeline.get_timeline_interface()
-    if db.internal_state.needs_initialization(db, timeline.is_stopped()):
-        if not db.internal_state.initialize(db, len(kernel_shape)):
+    if db.per_instance_state.needs_initialization(db, timeline.is_stopped()):
+        if not db.per_instance_state.initialize(db, len(kernel_shape)):
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
     # Exit early if there are no outputs defined.
-    if not db.internal_state.attr_infos[ATTR_PORT_TYPE_OUTPUT]:
+    if not db.per_instance_state.attr_infos[ATTR_PORT_TYPE_OUTPUT]:
         return
 
     # Retrieve the inputs and outputs argument values to pass to the kernel.
     inputs, outputs = get_kernel_args(
         db.inputs,
         db.outputs,
-        db.internal_state.attr_infos,
-        db.internal_state.kernel_module,
+        db.per_instance_state.attr_infos,
+        db.per_instance_state.kernel_module,
         kernel_shape,
     )
 
     # Ensure that all array input values are valid.
-    validate_input_arrays(db.node, db.internal_state.attr_infos, inputs)
+    validate_input_arrays(db.node, db.per_instance_state.attr_infos, inputs)
 
     # Launch the kernel.
     wp.launch(
-        db.internal_state.kernel_module.compute,
+        db.per_instance_state.kernel_module.compute,
         dim=kernel_shape,
         inputs=[inputs],
         outputs=[outputs],
     )
 
     # Write the output values to the node's attributes.
-    write_output_attrs(db.outputs, db.internal_state.attr_infos, outputs)
+    write_output_attrs(db.outputs, db.per_instance_state.attr_infos, outputs)
 
 
 #   Node Entry Point
@@ -222,7 +222,7 @@ def compute(db: OgnKernelDatabase) -> None:
             with wp.ScopedDevice(device):
                 compute(db, device)
         except Exception:
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             db.log_error(traceback.format_exc())
             wp.config.quiet = True
             return

exts/omni.warp/omni/warp/nodes/_impl/OgnMeshFromVolume.py

@@ -101,7 +101,7 @@ def compute(db: OgnMeshFromVolumeDatabase) -> None:
     if not db.inputs.data.memory or db.inputs.data.shape[0] == 0:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     # Initialize the internal state if it hasn't been already.
     if state.needs_initialization(db):
@@ -210,10 +210,10 @@ def compute(db: OgnMeshFromVolumeDatabase) -> None:
                 compute(db)
         except Exception:
             db.log_error(traceback.format_exc())
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
         # Fire the execution for the downstream nodes.
         db.outputs.execOut = og.ExecutionAttributeState.ENABLED

exts/omni.warp/omni/warp/nodes/_impl/OgnParticlesFromMesh.py

@@ -196,7 +196,7 @@ def spawn_particles(db: OgnParticlesFromMeshDatabase) -> Tuple[wp.array, int]:
         kernel=sample_mesh_kernel,
         dim=dims,
         inputs=[
-            db.internal_state.mesh.id,
+            db.per_instance_state.mesh.id,
             extent[0],
             db.inputs.maxPoints,
             db.inputs.minSdf,
@@ -224,7 +224,7 @@ def compute(db: OgnParticlesFromMeshDatabase) -> None:
     if not db.inputs.mesh.valid or not db.outputs.particles.valid:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     # Initialize the internal state if it hasn't been already.
     if state.needs_initialization(db):
@@ -307,10 +307,10 @@ def compute(db: OgnParticlesFromMeshDatabase) -> None:
                 compute(db)
         except Exception:
             db.log_error(traceback.format_exc())
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
         # Fire the execution for the downstream nodes.
         db.outputs.execOut = og.ExecutionAttributeState.ENABLED

exts/omni.warp/omni/warp/nodes/_impl/OgnParticlesSimulate.py

@@ -352,7 +352,7 @@ def update_collider(
     db: OgnParticlesSimulateDatabase,
 ) -> None:
     """Updates the collider state."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     points = omni.warp.nodes.mesh_get_points(db.inputs.collider)
     xform = omni.warp.nodes.bundle_get_world_xform(db.inputs.collider)
@@ -398,7 +398,7 @@ def update_particles(
     db: OgnParticlesSimulateDatabase,
 ) -> None:
     """Updates the particles state."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     xform = omni.warp.nodes.bundle_get_world_xform(db.inputs.particles)
 
@@ -426,7 +426,7 @@ def update_particles(
 
 def step(db: OgnParticlesSimulateDatabase) -> None:
     """Steps through the simulation."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     sim_dt = state.sim_dt / db.inputs.substepCount
 
@@ -448,7 +448,7 @@ def step(db: OgnParticlesSimulateDatabase) -> None:
 
 def simulate(db: OgnParticlesSimulateDatabase) -> None:
     """Simulates the particles at the current time."""
-    state = db.internal_state
+    state = db.per_instance_state
 
     state.model.particle_grid.build(
         state.state_0.particle_q,
@@ -466,7 +466,7 @@ def compute(db: OgnParticlesSimulateDatabase, device: wp.context.Device) -> None
     if not db.inputs.particles.valid or not db.outputs.particles.valid:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     if not db.inputs.enabled:
         # Pass through the data.
@@ -556,10 +556,10 @@ def compute(db: OgnParticlesSimulateDatabase) -> None:
                 compute(db, device)
         except Exception:
             db.log_error(traceback.format_exc())
-            db.internal_state.is_valid = False
+            db.per_instance_state.is_valid = False
             return
 
-        db.internal_state.is_valid = True
+        db.per_instance_state.is_valid = True
 
         # Fire the execution for the downstream nodes.
         db.outputs.execOut = og.ExecutionAttributeState.ENABLED

exts/omni.warp/omni/warp/nodes/_impl/OgnSampleOceanDeform.py

@@ -210,7 +210,7 @@ def compute(db: OgnSampleOceanDeformDatabase) -> None:
     if not db.inputs.mesh.valid or not db.outputs.mesh.valid:
         return
 
-    state = db.internal_state
+    state = db.per_instance_state
 
     amplitude = max(0.0001, min(1000.0, db.inputs.amplitude))
     direction = db.inputs.direction % 6.28318530718