warp-lang 0.15.0__py3-none-manylinux2014_x86_64.whl → 1.0.0__py3-none-manylinux2014_x86_64.whl

warp/examples/sim/example_quadruped.py

@@ -79,12 +79,12 @@ class Example:
             floating=True,
             density=1000,
             armature=0.01,
-            stiffness=120,
+            stiffness=200,
             damping=1,
-            shape_ke=1.0e4,
-            shape_kd=1.0e2,
-            shape_kf=1.0e2,
-            shape_mu=0.0,
+            contact_ke=1.0e4,
+            contact_kd=1.0e2,
+            contact_kf=1.0e2,
+            contact_mu=1.0,
             limit_ke=1.0e4,
             limit_kd=1.0e1,
         )
@@ -106,17 +106,21 @@ class Example:
 
             builder.joint_q[-12:] = [0.2, 0.4, -0.6, -0.2, -0.4, 0.6, -0.2, 0.4, -0.6, 0.2, -0.4, 0.6]
 
-            builder.joint_target[-12:] = [0.2, 0.4, -0.6, -0.2, -0.4, 0.6, -0.2, 0.4, -0.6, 0.2, -0.4, 0.6]
+            builder.joint_axis_mode = [wp.sim.JOINT_MODE_TARGET_POSITION] * len(builder.joint_axis_mode)
+            builder.joint_act[-12:] = [0.2, 0.4, -0.6, -0.2, -0.4, 0.6, -0.2, 0.4, -0.6, 0.2, -0.4, 0.6]
 
         np.set_printoptions(suppress=True)
         # finalize model
         self.model = builder.finalize()
         self.model.ground = True
+        # self.model.gravity = 0.0
 
         self.model.joint_attach_ke = 16000.0
         self.model.joint_attach_kd = 200.0
 
-        self.integrator = wp.sim.XPBDIntegrator()
+        # self.integrator = wp.sim.XPBDIntegrator()
+        # self.integrator = wp.sim.SemiImplicitIntegrator()
+        self.integrator = wp.sim.FeatherstoneIntegrator(self.model)
 
         self.renderer = None
         if stage:
@@ -130,10 +134,11 @@ class Example:
         wp.sim.eval_fk(self.model, self.model.joint_q, self.model.joint_qd, None, self.state_0)
 
         self.use_graph = wp.get_device().is_cuda
+        self.graph = None
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def simulate(self):
         for _ in range(self.sim_substeps):
@@ -172,4 +177,3 @@ if __name__ == "__main__":
 
     if example.renderer:
         example.renderer.save()
-

warp/examples/sim/example_rigid_chain.py

@@ -45,7 +45,7 @@ class Example:
         episode_duration = 5.0  # seconds
         self.episode_frames = int(episode_duration / self.frame_dt)
 
-        self.sim_substeps = 32  # 5
+        self.sim_substeps = 10
         self.sim_dt = self.frame_dt / self.sim_substeps
 
         for c, t in enumerate(self.chain_types):
@@ -88,8 +88,8 @@ class Example:
                         limit_upper=joint_limit_upper,
                         target_ke=0.0,
                         target_kd=0.0,
-                        limit_ke=30.0,
-                        limit_kd=30.0,
+                        limit_ke=1e5,
+                        limit_kd=1.0,
                     )
 
                 elif joint_type == wp.sim.JOINT_UNIVERSAL:
@@ -128,11 +128,11 @@ class Example:
                         parent_xform=parent_joint_xform,
                         child_xform=wp.transform_identity(),
                     )
-        # finalize model
+
         self.model = builder.finalize()
         self.model.ground = False
 
-        self.integrator = wp.sim.XPBDIntegrator(iterations=5)
+        self.integrator = wp.sim.FeatherstoneIntegrator(self.model)
 
         self.renderer = None
         if stage:
@@ -145,9 +145,9 @@ class Example:
 
         self.use_graph = wp.get_device().is_cuda
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def simulate(self):
         for _ in range(self.sim_substeps):
@@ -167,6 +167,9 @@ class Example:
         if self.renderer is None:
             return
 
+        if self.renderer is None:
+            return
+
         with wp.ScopedTimer("render", active=True):
             self.renderer.begin_frame(self.sim_time)
             self.renderer.render(self.state_0)
@@ -184,3 +187,5 @@ if __name__ == "__main__":
 
     if example.renderer:
         example.renderer.save()
+    if example.renderer:
+        example.renderer.save()

warp/examples/sim/example_rigid_contact.py

@@ -125,9 +125,9 @@ class Example:
 
         self.use_graph = wp.get_device().is_cuda
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def load_mesh(self, filename, path):
         asset_stage = Usd.Stage.Open(filename)
@@ -156,7 +156,7 @@ class Example:
     def render(self):
         if self.renderer is None:
             return
-        
+
         with wp.ScopedTimer("render", active=True):
             self.renderer.begin_frame(self.sim_time)
             self.renderer.render(self.state_0)

warp/examples/sim/example_rigid_force.py

@@ -25,7 +25,7 @@ wp.init()
 
 class Example:
     parser = argparse.ArgumentParser()
-    parser.add_argument("--opengl", action='store_true')
+    parser.add_argument("--opengl", action="store_true")
 
     def __init__(self, stage=None, args=None, **kwargs):
         if args is None:
@@ -33,7 +33,7 @@ class Example:
             args = argparse.Namespace(**kwargs)
             args = Example.parser.parse_args(args=[], namespace=args)
         self._args = args
-        
+
         self.sim_fps = 60.0
         self.sim_substeps = 5
         self.sim_duration = 5.0
@@ -64,9 +64,9 @@ class Example:
 
         self.use_graph = wp.get_device().is_cuda
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def simulate(self):
         for _ in range(self.sim_substeps):
@@ -77,7 +77,7 @@ class Example:
 
             self.state_0.body_f.assign(
                 [
-                    [0.0, 0.0, -10000.0, 0.0, 0.0, 0.0],
+                    [0.0, 0.0, -7000.0, 0.0, 0.0, 0.0],
                 ]
             )
 
@@ -123,4 +123,4 @@ if __name__ == "__main__":
     if example.renderer:
         example.renderer.save()
 
-    example.renderer = None
+    example.renderer = None

warp/examples/sim/example_rigid_soft_contact.py

@@ -77,9 +77,9 @@ class Example:
 
         self.use_graph = wp.get_device().is_cuda
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def simulate(self):
         for s in range(self.sim_substeps):
@@ -98,7 +98,7 @@ class Example:
             if self.use_graph:
                 wp.capture_launch(self.graph)
             else:
-                self.simulate()                    
+                self.simulate()
             self.sim_time += self.frame_dt
 
     def render(self):

warp/examples/sim/example_soft_body.py

@@ -118,9 +118,9 @@ class Example:
 
         self.use_graph = wp.get_device().is_cuda
         if self.use_graph:
-            wp.capture_begin()
-            self.simulate()
-            self.graph = wp.capture_end()
+            with wp.ScopedCapture() as capture:
+                self.simulate()
+            self.graph = capture.graph
 
     def simulate(self):
         for _ in range(self.sim_substeps):

warp/jax.py

@@ -6,6 +6,7 @@
 # license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 import warp
+from warp.context import type_str
 
 
 def device_to_jax(wp_device):
@@ -34,6 +35,50 @@ def device_from_jax(jax_device):
         raise RuntimeError(f"Unknown or unsupported Jax device platform '{jax_device.platform}'")
 
 
+def dtype_to_jax(wp_dtype):
+    import jax.numpy as jp
+
+    warp_to_jax_dict = {
+        warp.float16: jp.float16,
+        warp.float32: jp.float32,
+        warp.float64: jp.float64,
+        warp.int8: jp.int8,
+        warp.int16: jp.int16,
+        warp.int32: jp.int32,
+        warp.int64: jp.int64,
+        warp.uint8: jp.uint8,
+        warp.uint16: jp.uint16,
+        warp.uint32: jp.uint32,
+        warp.uint64: jp.uint64,
+    }
+    jax_dtype = warp_to_jax_dict.get(wp_dtype)
+    if jax_dtype is None:
+        raise TypeError(f"Invalid or unsupported data type: {type_str(wp_dtype)}")
+    return jax_dtype
+
+
+def dtype_from_jax(jax_dtype):
+    import jax.numpy as jp
+
+    jax_to_warp_dict = {
+        jp.float16: warp.float16,
+        jp.float32: warp.float32,
+        jp.float64: warp.float64,
+        jp.int8: warp.int8,
+        jp.int16: warp.int16,
+        jp.int32: warp.int32,
+        jp.int64: warp.int64,
+        jp.uint8: warp.uint8,
+        jp.uint16: warp.uint16,
+        jp.uint32: warp.uint32,
+        jp.uint64: warp.uint64,
+    }
+    wp_dtype = jax_to_warp_dict.get(jax_dtype)
+    if wp_dtype is None:
+        raise TypeError(f"Invalid or unsupported data type: {jax_dtype}")
+    return wp_dtype
+
+
 def to_jax(wp_array):
     import jax.dlpack
 

warp/jax_experimental.py

@@ -0,0 +1,339 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.  All rights reserved.
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import ctypes
+import warp as wp
+from warp.types import array_t, launch_bounds_t, strides_from_shape
+from warp.context import type_str
+import jax
+import jax.numpy as jp
+
+_jax_warp_p = None
+
+# Holder for the custom callback to keep it alive.
+_cc_callback = None
+_registered_kernels = [None]
+_registered_kernel_to_id = {}
+
+
+def jax_kernel(wp_kernel):
+    """Create a Jax primitive from a Warp kernel.
+
+    NOTE: This is an experimental feature under development.
+
+    Current limitations:
+    - All kernel arguments must be arrays.
+    - Kernel launch dimensions are inferred from the shape of the first argument.
+    - Input arguments are followed by output arguments in the Warp kernel definition.
+    - There must be at least one input argument and at least one output argument.
+    - Output shapes must match the launch dimensions (i.e., output shapes must match the shape of the first argument).
+    - All arrays must be contiguous.
+    - Only the CUDA backend is supported.
+    """
+
+    if _jax_warp_p == None:
+        # Create and register the primitive
+        _create_jax_warp_primitive()
+    if not wp_kernel in _registered_kernel_to_id:
+        id = len(_registered_kernels)
+        _registered_kernels.append(wp_kernel)
+        _registered_kernel_to_id[wp_kernel] = id
+    else:
+        id = _registered_kernel_to_id[wp_kernel]
+
+    def bind(*args):
+        return _jax_warp_p.bind(*args, kernel=id)
+
+    return bind
+
+
+def _warp_custom_callback(stream, buffers, opaque, opaque_len):
+    # The descriptor is the form
+    # <kernel-id>|<launch-dims>|<arg-dims-list>
+    # Example:  42|16,32|16,32;100;16,32
+    kernel_id_str, dim_str, args_str = opaque.decode().split("|")
+
+    # Get the kernel from the registry.
+    kernel_id = int(kernel_id_str)
+    kernel = _registered_kernels[kernel_id]
+
+    # Parse launch dimensions.
+    dims = [int(d) for d in dim_str.split(",")]
+    bounds = launch_bounds_t(dims)
+
+    # Parse arguments.
+    arg_strings = args_str.split(";")
+    num_args = len(arg_strings)
+    assert num_args == len(kernel.adj.args), "Incorrect number of arguments"
+
+    # First param is the launch bounds.
+    kernel_params = (ctypes.c_void_p * (1 + num_args))()
+    kernel_params[0] = ctypes.addressof(bounds)
+
+    # Parse array descriptors.
+    args = []
+    for i in range(num_args):
+        dtype = kernel.adj.args[i].type.dtype
+        shape = [int(d) for d in arg_strings[i].split(",")]
+        strides = strides_from_shape(shape, dtype)
+
+        arr = array_t(buffers[i], 0, len(shape), shape, strides)
+        args.append(arr)  # keep a reference
+        arg_ptr = ctypes.addressof(arr)
+
+        kernel_params[i + 1] = arg_ptr
+
+    # Get current device.
+    device = wp.device_from_jax(_get_jax_device())
+
+    # Get kernel hooks.
+    # Note: module was loaded during jit lowering.
+    hooks = kernel.module.get_kernel_hooks(kernel, device)
+    assert hooks.forward, "Failed to find kernel entry point"
+
+    # Launch the kernel.
+    wp.context.runtime.core.cuda_launch_kernel(
+        device.context, hooks.forward, bounds.size, 0, kernel_params, stream
+    )
+
+
+# TODO: is there a simpler way of getting the Jax "current" device?
+def _get_jax_device():
+    # check if jax.default_device() context manager is active
+    device = jax.config.jax_default_device
+    # if default device is not set, use first device
+    if device is None:
+        device = jax.devices()[0]
+    return device
+
+
+def _create_jax_warp_primitive():
+    from functools import reduce
+    import jax
+    from jax._src.interpreters import batching
+    from jax.interpreters import mlir
+    from jax.interpreters.mlir import ir
+    from jaxlib.hlo_helpers import custom_call
+
+    global _jax_warp_p
+    global _cc_callback
+
+    # Create and register the primitive.
+    # TODO add default implementation that calls the kernel via warp.
+    _jax_warp_p = jax.core.Primitive("jax_warp")
+    _jax_warp_p.multiple_results = True
+
+    # TODO Just launch the kernel directly, but make sure the argument
+    # shapes are massaged the same way as below so that vmap works.
+    def impl(*args):
+        raise Exception("Not implemented")
+
+    _jax_warp_p.def_impl(impl)
+
+    # Auto-batching. Make sure all the arguments are fully broadcasted
+    # so that Warp is not confused about dimensions.
+    def vectorized_multi_batcher(args, dims, **params):
+        # Figure out the number of outputs.
+        wp_kernel = _registered_kernels[params["kernel"]]
+        output_count = len(wp_kernel.adj.args) - len(args)
+        shape, dim = next((a.shape, d) for a, d in zip(args, dims) if d is not None)
+        size = shape[dim]
+        args = [batching.bdim_at_front(a, d, size) if len(a.shape) else a for a, d in zip(args, dims)]
+        # Create the batched primitive.
+        return _jax_warp_p.bind(*args, **params), [dims[0]] * output_count
+
+    batching.primitive_batchers[_jax_warp_p] = vectorized_multi_batcher
+
+    def get_vecmat_shape(warp_type):
+        if hasattr(warp_type.dtype, "_shape_"):
+            return warp_type.dtype._shape_
+        return []
+
+    def strip_vecmat_dimensions(warp_arg, actual_shape):
+        shape = get_vecmat_shape(warp_arg.type)
+        for i, s in enumerate(reversed(shape)):
+            item = actual_shape[-i - 1]
+            if s != item:
+                raise Exception(f"The vector/matrix shape for argument {warp_arg.label} does not match")
+        return actual_shape[: len(actual_shape) - len(shape)]
+
+    def collapse_into_leading_dimension(warp_arg, actual_shape):
+        if len(actual_shape) < warp_arg.type.ndim:
+            raise Exception(f"Argument {warp_arg.label} has too few non-matrix/vector dimensions")
+        index_rest = len(actual_shape) - warp_arg.type.ndim + 1
+        leading_size = reduce(lambda x, y: x * y, actual_shape[:index_rest])
+        return [leading_size] + actual_shape[index_rest:]
+
+    # Infer array dimensions from input type.
+    def infer_dimensions(warp_arg, actual_shape):
+        actual_shape = strip_vecmat_dimensions(warp_arg, actual_shape)
+        return collapse_into_leading_dimension(warp_arg, actual_shape)
+
+    def base_type_to_jax(warp_dtype):
+        if hasattr(warp_dtype, "_wp_scalar_type_"):
+            return wp.jax.dtype_to_jax(warp_dtype._wp_scalar_type_)
+        return wp.jax.dtype_to_jax(warp_dtype)
+
+    def base_type_to_jax_ir(warp_dtype):
+        warp_to_jax_dict = {
+            wp.float16: ir.F16Type.get(),
+            wp.float32: ir.F32Type.get(),
+            wp.float64: ir.F64Type.get(),
+            wp.int8: ir.IntegerType.get_signless(8),
+            wp.int16: ir.IntegerType.get_signless(16),
+            wp.int32: ir.IntegerType.get_signless(32),
+            wp.int64: ir.IntegerType.get_signless(64),
+            wp.uint8: ir.IntegerType.get_unsigned(8),
+            wp.uint16: ir.IntegerType.get_unsigned(16),
+            wp.uint32: ir.IntegerType.get_unsigned(32),
+            wp.uint64: ir.IntegerType.get_unsigned(64),
+        }
+        if hasattr(warp_dtype, "_wp_scalar_type_"):
+            warp_dtype = warp_dtype._wp_scalar_type_
+        jax_dtype = warp_to_jax_dict.get(warp_dtype)
+        if jax_dtype is None:
+            raise TypeError(f"Invalid or unsupported data type: {warp_dtype}")
+        return jax_dtype
+
+    def base_type_is_compatible(warp_type, jax_ir_type):
+        jax_ir_to_warp = {
+            "f16": wp.float16,
+            "f32": wp.float32,
+            "f64": wp.float64,
+            "i8": wp.int8,
+            "i16": wp.int16,
+            "i32": wp.int32,
+            "i64": wp.int64,
+            "ui8": wp.uint8,
+            "ui16": wp.uint16,
+            "ui32": wp.uint32,
+            "ui64": wp.uint64,
+        }
+        expected_warp_type = jax_ir_to_warp.get(str(jax_ir_type))
+        if expected_warp_type is not None:
+            if hasattr(warp_type, "_wp_scalar_type_"):
+                return warp_type._wp_scalar_type_ == expected_warp_type
+            else:
+                return warp_type == expected_warp_type
+        else:
+            raise TypeError(f"Invalid or unsupported data type: {jax_ir_type}")
+
+    # Abstract evaluation.
+    def jax_warp_abstract(*args, kernel=None):
+        wp_kernel = _registered_kernels[kernel]
+        # All the extra arguments to the warp kernel are outputs.
+        warp_outputs = [o.type for o in wp_kernel.adj.args[len(args) :]]
+        # TODO. Let's just use the first input dimension to infer the output's dimensions.
+        dims = strip_vecmat_dimensions(wp_kernel.adj.args[0], list(args[0].shape))
+        jax_outputs = []
+        for o in warp_outputs:
+            shape = list(dims) + list(get_vecmat_shape(o))
+            dtype = base_type_to_jax(o.dtype)
+            jax_outputs.append(jax.core.ShapedArray(shape, dtype))
+        return jax_outputs
+
+    _jax_warp_p.def_abstract_eval(jax_warp_abstract)
+
+    # Lowering to MLIR.
+
+    # Create python-land custom call target.
+    CCALLFUNC = ctypes.CFUNCTYPE(
+        ctypes.c_voidp, ctypes.c_void_p, ctypes.POINTER(ctypes.c_void_p), ctypes.c_char_p, ctypes.c_size_t
+    )
+    _cc_callback = CCALLFUNC(_warp_custom_callback)
+    ccall_address = ctypes.cast(_cc_callback, ctypes.c_void_p)
+
+    # Put the custom call into a capsule, as required by XLA.
+    PyCapsule_Destructor = ctypes.CFUNCTYPE(None, ctypes.py_object)
+    PyCapsule_New = ctypes.pythonapi.PyCapsule_New
+    PyCapsule_New.restype = ctypes.py_object
+    PyCapsule_New.argtypes = (ctypes.c_void_p, ctypes.c_char_p, PyCapsule_Destructor)
+    capsule = PyCapsule_New(ccall_address.value, b"xla._CUSTOM_CALL_TARGET", PyCapsule_Destructor(0))
+
+    # Register the callback in XLA.
+    jax.lib.xla_client.register_custom_call_target("warp_call", capsule, platform="gpu")
+
+    def default_layout(shape):
+        return range(len(shape) - 1, -1, -1)
+
+    def warp_call_lowering(ctx, *args, kernel=None):
+        if not kernel:
+            raise Exception("Unknown kernel id " + str(kernel))
+        wp_kernel = _registered_kernels[kernel]
+
+        # TODO This may not be necessary, but it is perhaps better not to be
+        # mucking with kernel loading while already running the workload.
+        module = wp_kernel.module
+        device = wp.device_from_jax(_get_jax_device())
+        if not module.load(device):
+            raise Exception("Could not load kernel on device")
+
+        # Infer dimensions from the first input.
+        warp_arg0 = wp_kernel.adj.args[0]
+        actual_shape0 = ir.RankedTensorType(args[0].type).shape
+        dims = strip_vecmat_dimensions(warp_arg0, actual_shape0)
+        warp_dims = collapse_into_leading_dimension(warp_arg0, dims)
+
+        # Figure out the types and shapes of the input arrays.
+        arg_strings = []
+        operand_layouts = []
+        for actual, warg in zip(args, wp_kernel.adj.args):
+            wtype = warg.type
+            rtt = ir.RankedTensorType(actual.type)
+
+            if not isinstance(wtype, wp.array):
+                raise Exception("Only contiguous arrays are supported for Jax kernel arguments")
+
+            if not base_type_is_compatible(wtype.dtype, rtt.element_type):
+                raise TypeError(f"Incompatible data type for argument '{warg.label}', expected {type_str(wtype.dtype)}, got {rtt.element_type}")
+
+            # Infer array dimension (by removing the vector/matrix dimensions and
+            # collapsing the initial dimensions).
+            shape = infer_dimensions(warg, rtt.shape)
+
+            if len(shape) != wtype.ndim:
+                raise TypeError(f"Incompatible array dimensionality for argument '{warg.label}'")
+
+            arg_strings.append(",".join([str(d) for d in shape]))
+            operand_layouts.append(default_layout(rtt.shape))
+
+        # Figure out the types and shapes of the output arrays.
+        result_types = []
+        result_layouts = []
+        for warg in wp_kernel.adj.args[len(args) :]:
+            wtype = warg.type
+
+            if not isinstance(wtype, wp.array):
+                raise Exception("Only contiguous arrays are supported for Jax kernel arguments")
+
+            # Infer dimensions from the first input.
+            arg_strings.append(",".join([str(d) for d in warp_dims]))
+
+            result_shape = list(dims) + list(get_vecmat_shape(wtype))
+            result_types.append(ir.RankedTensorType.get(result_shape, base_type_to_jax_ir(wtype.dtype)))
+            result_layouts.append(default_layout(result_shape))
+
+        # Build opaque descriptor for callback.
+        shape_str = ",".join([str(d) for d in warp_dims])
+        args_str = ";".join(arg_strings)
+        descriptor = f"{kernel}|{shape_str}|{args_str}"
+
+        out = custom_call(
+            b"warp_call",
+            result_types=result_types,
+            operands=args,
+            backend_config=descriptor.encode("utf-8"),
+            operand_layouts=operand_layouts,
+            result_layouts=result_layouts,
+        ).results
+        return out
+
+    mlir.register_lowering(
+        _jax_warp_p,
+        warp_call_lowering,
+        platform="gpu",
+    )