Merge branch 'walker_update' into 'main'

Updated walker example uses tile API

See merge request omniverse/warp!967

CHANGELOG.md

@@ -11,6 +11,7 @@
 - Add per-module option to disable fused floating point operations, use `wp.set_module_options({"fuse_fp": False})`
   ([GH-379](https://github.com/NVIDIA/warp/issues/379)).
 - Add per-module option to add CUDA-C line information for profiling, use `wp.set_module_options({"lineinfo": True})`.
+- Add `example_tile_walker.py`, which reworks the existing `walker.py` to use Warp's tile API for matrix multiplication.
 
 ### Changed
 

warp/examples/optim/example_walker.py

@@ -15,6 +15,11 @@
 # simulated forward in time and then evaluated based on the center of mass
 # momentum of the mesh.
 #
+# This example uses the deprecated wp.matmul() for matrix multiplication,
+# which will be removed in a future version. See the updated version of
+# this example, example_tile_walker.py, in examples/tile for the new
+# approach to GEMMs using Warp's tile API.
+#
 ###########################################################################
 
 import math

warp/examples/tile/example_tile_walker.py

@@ -0,0 +1,319 @@
+# Copyright (c) 2025 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.
+
+###########################################################################
+# Example Tile Walker
+#
+# Trains a tetrahedral mesh quadruped to run. Feeds 8 time-varying input
+# phases as inputs into a single layer fully connected network with a tanh
+# activation function. Interprets the output of the network as tet
+# activations, which are fed into the wp.sim soft mesh model. This is
+# simulated forward in time and then evaluated based on the center of mass
+# momentum of the mesh.
+#
+# This example uses the Warp tile API, which as of Warp 1.6 is the
+# recommended way to handle matrix multiplication. example_walker.py in
+# examples/optim demonstrates the old way of doing matrix multiplication,
+# wp.matmul(), which will be deprecated in a future version.
+#
+###########################################################################
+
+import math
+import os
+
+import numpy as np
+from pxr import Gf, Usd, UsdGeom
+
+import warp as wp
+import warp.examples
+import warp.optim
+import warp.sim
+import warp.sim.render
+
+PHASE_COUNT = 8
+PHASE_STEP = wp.constant((2.0 * math.pi) / PHASE_COUNT)
+PHASE_FREQ = wp.constant(5.0)
+ACTIVATION_STRENGTH = wp.constant(0.3)
+
+TILE_TETS = wp.constant(8)
+TILE_THREADS = 64
+
+
+@wp.kernel
+def loss_kernel(com: wp.array(dtype=wp.vec3), loss: wp.array(dtype=float)):
+    tid = wp.tid()
+    vx = com[tid][0]
+    vy = com[tid][1]
+    vz = com[tid][2]
+    delta = wp.sqrt(vx * vx) + wp.sqrt(vy * vy) - vz
+
+    wp.atomic_add(loss, 0, delta)
+
+
+@wp.kernel
+def com_kernel(velocities: wp.array(dtype=wp.vec3), n: int, com: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+    v = velocities[tid]
+    a = v / wp.float32(n)
+    wp.atomic_add(com, 0, a)
+
+
+@wp.kernel
+def compute_phases(phases: wp.array(dtype=float), sim_time: float):
+    tid = wp.tid()
+    phases[tid] = wp.sin(PHASE_FREQ * sim_time + wp.float32(tid) * PHASE_STEP)
+
+
+@wp.func
+def tanh(x: float):
+    return wp.tanh(x) * ACTIVATION_STRENGTH
+
+
+@wp.kernel
+def network(
+    phases: wp.array2d(dtype=float), weights: wp.array2d(dtype=float), tet_activations: wp.array2d(dtype=float)
+):
+    # output tile index
+    i = wp.tid()
+
+    # GEMM
+    p = wp.tile_load(phases, 0, 0, m=PHASE_COUNT, n=1)
+    w = wp.tile_load(weights, i, 0, m=TILE_TETS, n=PHASE_COUNT)
+    out = wp.tile_matmul(w, p)
+
+    # activation
+    activations = wp.tile_map(tanh, out)
+    wp.tile_store(tet_activations, i, 0, activations)
+
+
+class Example:
+    def __init__(self, stage_path="example_tile_walker.usd", verbose=False, num_frames=300):
+        self.verbose = verbose
+
+        fps = 60
+        self.frame_dt = 1.0 / fps
+        self.num_frames = num_frames
+
+        self.sim_substeps = 80
+        self.sim_dt = self.frame_dt / self.sim_substeps
+        self.sim_time = 0.0
+
+        self.iter = 0
+        self.train_rate = 0.025
+
+        self.phase_count = PHASE_COUNT
+
+        self.render_time = 0.0
+
+        # bear
+        asset_stage = Usd.Stage.Open(os.path.join(warp.examples.get_asset_directory(), "bear.usd"))
+
+        geom = UsdGeom.Mesh(asset_stage.GetPrimAtPath("/root/bear"))
+        points = geom.GetPointsAttr().Get()
+
+        xform = Gf.Matrix4f(geom.ComputeLocalToWorldTransform(0.0))
+        for i in range(len(points)):
+            points[i] = xform.Transform(points[i])
+
+        self.points = [wp.vec3(point) for point in points]
+        self.tet_indices = geom.GetPrim().GetAttribute("tetraIndices").Get()
+
+        # sim model
+        builder = wp.sim.ModelBuilder()
+        builder.add_soft_mesh(
+            pos=wp.vec3(0.0, 0.5, 0.0),
+            rot=wp.quat_identity(),
+            scale=1.0,
+            vel=wp.vec3(0.0, 0.0, 0.0),
+            vertices=self.points,
+            indices=self.tet_indices,
+            density=1.0,
+            k_mu=2000.0,
+            k_lambda=2000.0,
+            k_damp=2.0,
+            tri_ke=0.0,
+            tri_ka=1e-8,
+            tri_kd=0.0,
+            tri_drag=0.0,
+            tri_lift=0.0,
+        )
+
+        # finalize model
+        self.model = builder.finalize(requires_grad=True)
+        self.control = self.model.control()
+
+        self.model.soft_contact_ke = 2.0e3
+        self.model.soft_contact_kd = 0.1
+        self.model.soft_contact_kf = 10.0
+        self.model.soft_contact_mu = 0.7
+
+        radii = wp.zeros(self.model.particle_count, dtype=float)
+        radii.fill_(0.05)
+        self.model.particle_radius = radii
+        self.model.ground = True
+
+        # allocate sim states
+        self.states = []
+        for _i in range(self.num_frames * self.sim_substeps + 1):
+            self.states.append(self.model.state(requires_grad=True))
+
+        # initialize the integrator.
+        self.integrator = wp.sim.SemiImplicitIntegrator()
+
+        # model input
+        self.phases = []
+        for _i in range(self.num_frames):
+            self.phases.append(wp.zeros(self.phase_count, dtype=float, requires_grad=True))
+
+        # weights matrix for linear network
+        rng = np.random.default_rng(42)
+        k = 1.0 / self.phase_count
+        weights = rng.uniform(-np.sqrt(k), np.sqrt(k), (self.model.tet_count, self.phase_count))
+        self.weights = wp.array(weights, dtype=float, requires_grad=True)
+
+        # tanh activation layer array
+        self.tet_activations = []
+        for _i in range(self.num_frames):
+            self.tet_activations.append(wp.zeros(self.model.tet_count, dtype=float, requires_grad=True))
+
+        # optimization
+        self.loss = wp.zeros(1, dtype=float, requires_grad=True)
+        self.coms = []
+        for _i in range(self.num_frames):
+            self.coms.append(wp.zeros(1, dtype=wp.vec3, requires_grad=True))
+        self.optimizer = warp.optim.Adam([self.weights.flatten()], lr=self.train_rate)
+
+        # rendering
+        if stage_path:
+            self.renderer = wp.sim.render.SimRenderer(self.model, stage_path)
+        else:
+            self.renderer = None
+
+        # capture forward/backward passes
+        self.use_cuda_graph = wp.get_device().is_cuda
+        if self.use_cuda_graph:
+            with wp.ScopedCapture() as capture:
+                self.tape = wp.Tape()
+                with self.tape:
+                    for i in range(self.num_frames):
+                        self.forward(i)
+                self.tape.backward(self.loss)
+            self.graph = capture.graph
+
+    def forward(self, frame):
+        with wp.ScopedTimer("network", active=self.verbose):
+            # build sinusoidal input phases
+            wp.launch(kernel=compute_phases, dim=self.phase_count, inputs=[self.phases[frame], self.sim_time])
+
+            # apply linear network with tanh activation
+            wp.launch_tiled(
+                kernel=network,
+                dim=math.ceil(self.model.tet_count / TILE_TETS),
+                inputs=[self.phases[frame].reshape((self.phase_count, 1)), self.weights],
+                outputs=[self.tet_activations[frame].reshape((self.model.tet_count, 1))],
+                block_dim=TILE_THREADS,
+            )
+            self.control.tet_activations = self.tet_activations[frame]
+
+        with wp.ScopedTimer("simulate", active=self.verbose):
+            # run simulation loop
+            for i in range(self.sim_substeps):
+                self.states[frame * self.sim_substeps + i].clear_forces()
+                self.integrator.simulate(
+                    self.model,
+                    self.states[frame * self.sim_substeps + i],
+                    self.states[frame * self.sim_substeps + i + 1],
+                    self.sim_dt,
+                    self.control,
+                )
+                self.sim_time += self.sim_dt
+
+        with wp.ScopedTimer("loss", active=self.verbose):
+            # compute center of mass velocity
+            wp.launch(
+                com_kernel,
+                dim=self.model.particle_count,
+                inputs=[
+                    self.states[(frame + 1) * self.sim_substeps].particle_qd,
+                    self.model.particle_count,
+                    self.coms[frame],
+                ],
+                outputs=[],
+            )
+            # compute loss
+            wp.launch(loss_kernel, dim=1, inputs=[self.coms[frame], self.loss], outputs=[])
+
+    def step(self):
+        with wp.ScopedTimer("step"):
+            if self.use_cuda_graph:
+                wp.capture_launch(self.graph)
+            else:
+                self.tape = wp.Tape()
+                with self.tape:
+                    for i in range(self.num_frames):
+                        self.forward(i)
+                self.tape.backward(self.loss)
+
+            # optimization
+            x = self.weights.grad.flatten()
+            self.optimizer.step([x])
+
+        loss = self.loss.numpy()
+        if self.verbose:
+            print(f"Iteration {self.iter}: {loss}")
+
+        # reset sim
+        self.sim_time = 0.0
+        self.states[0] = self.model.state(requires_grad=True)
+
+        # clear grads and zero arrays for next iteration
+        self.tape.zero()
+        self.loss.zero_()
+        for i in range(self.num_frames):
+            self.coms[i].zero_()
+
+        self.iter += 1
+
+    def render(self):
+        if self.renderer is None:
+            return
+
+        with wp.ScopedTimer("render"):
+            for i in range(self.num_frames + 1):
+                self.renderer.begin_frame(self.render_time)
+                self.renderer.render(self.states[i * self.sim_substeps])
+                self.renderer.end_frame()
+
+                self.render_time += self.frame_dt
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("--device", type=str, default=None, help="Override the default Warp device.")
+    parser.add_argument(
+        "--stage_path",
+        type=lambda x: None if x == "None" else str(x),
+        default="example_tile_walker.usd",
+        help="Path to the output USD file.",
+    )
+    parser.add_argument("--num_frames", type=int, default=300, help="Total number of frames per training iteration.")
+    parser.add_argument("--train_iters", type=int, default=30, help="Total number of training iterations.")
+    parser.add_argument("--verbose", action="store_true", help="Print out additional status messages during execution.")
+
+    args = parser.parse_known_args()[0]
+
+    with wp.ScopedDevice(args.device):
+        example = Example(stage_path=args.stage_path, verbose=args.verbose, num_frames=args.num_frames)
+
+        for _ in range(args.train_iters):
+            example.step()
+            example.render()
+
+        if example.renderer:
+            example.renderer.save()