warp-lang 0.11.0__py3-none-manylinux2014_x86_64.whl → 1.0.0__py3-none-manylinux2014_x86_64.whl

warp/examples/core/example_fluid.py

@@ -0,0 +1,267 @@
+# Copyright (c) 2022 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 Fluid
+#
+# Shows how to implement a simple 2D Stable Fluids solver using
+# multidimensional arrays and launches.
+#
+###########################################################################
+
+import math
+
+import warp as wp
+import warp.render
+
+wp.init()
+
+grid_width = wp.constant(256)
+grid_height = wp.constant(128)
+
+
+@wp.func
+def lookup_float(f: wp.array2d(dtype=float), x: int, y: int):
+    x = wp.clamp(x, 0, grid_width - 1)
+    y = wp.clamp(y, 0, grid_height - 1)
+
+    return f[x, y]
+
+
+@wp.func
+def sample_float(f: wp.array2d(dtype=float), x: float, y: float):
+    lx = int(wp.floor(x))
+    ly = int(wp.floor(y))
+
+    tx = x - float(lx)
+    ty = y - float(ly)
+
+    s0 = wp.lerp(lookup_float(f, lx, ly), lookup_float(f, lx + 1, ly), tx)
+    s1 = wp.lerp(lookup_float(f, lx, ly + 1), lookup_float(f, lx + 1, ly + 1), tx)
+
+    s = wp.lerp(s0, s1, ty)
+    return s
+
+
+@wp.func
+def lookup_vel(f: wp.array2d(dtype=wp.vec2), x: int, y: int):
+    if x < 0 or x >= grid_width:
+        return wp.vec2()
+    if y < 0 or y >= grid_height:
+        return wp.vec2()
+
+    return f[x, y]
+
+
+@wp.func
+def sample_vel(f: wp.array2d(dtype=wp.vec2), x: float, y: float):
+    lx = int(wp.floor(x))
+    ly = int(wp.floor(y))
+
+    tx = x - float(lx)
+    ty = y - float(ly)
+
+    s0 = wp.lerp(lookup_vel(f, lx, ly), lookup_vel(f, lx + 1, ly), tx)
+    s1 = wp.lerp(lookup_vel(f, lx, ly + 1), lookup_vel(f, lx + 1, ly + 1), tx)
+
+    s = wp.lerp(s0, s1, ty)
+    return s
+
+
+@wp.kernel
+def advect(
+    u0: wp.array2d(dtype=wp.vec2),
+    u1: wp.array2d(dtype=wp.vec2),
+    rho0: wp.array2d(dtype=float),
+    rho1: wp.array2d(dtype=float),
+    dt: float,
+):
+    i, j = wp.tid()
+
+    u = u0[i, j]
+
+    # trace backward
+    p = wp.vec2(float(i), float(j))
+    p = p - u * dt
+
+    # advect
+    u1[i, j] = sample_vel(u0, p[0], p[1])
+    rho1[i, j] = sample_float(rho0, p[0], p[1])
+
+
+@wp.kernel
+def divergence(u: wp.array2d(dtype=wp.vec2), div: wp.array2d(dtype=float)):
+    i, j = wp.tid()
+
+    if i == grid_width - 1:
+        return
+    if j == grid_height - 1:
+        return
+
+    dx = (u[i + 1, j][0] - u[i, j][0]) * 0.5
+    dy = (u[i, j + 1][1] - u[i, j][1]) * 0.5
+
+    div[i, j] = dx + dy
+
+
+@wp.kernel
+def pressure_solve(p0: wp.array2d(dtype=float), p1: wp.array2d(dtype=float), div: wp.array2d(dtype=float)):
+    i, j = wp.tid()
+
+    s1 = lookup_float(p0, i - 1, j)
+    s2 = lookup_float(p0, i + 1, j)
+    s3 = lookup_float(p0, i, j - 1)
+    s4 = lookup_float(p0, i, j + 1)
+
+    # Jacobi update
+    err = s1 + s2 + s3 + s4 - div[i, j]
+
+    p1[i, j] = err * 0.25
+
+
+@wp.kernel
+def pressure_apply(p: wp.array2d(dtype=float), u: wp.array2d(dtype=wp.vec2)):
+    i, j = wp.tid()
+
+    if i == 0 or i == grid_width - 1:
+        return
+    if j == 0 or j == grid_height - 1:
+        return
+
+    # pressure gradient
+    f_p = wp.vec2(p[i + 1, j] - p[i - 1, j], p[i, j + 1] - p[i, j - 1]) * 0.5
+
+    u[i, j] = u[i, j] - f_p
+
+
+@wp.kernel
+def integrate(u: wp.array2d(dtype=wp.vec2), rho: wp.array2d(dtype=float), dt: float):
+    i, j = wp.tid()
+
+    # gravity
+    f_g = wp.vec2(-90.8, 0.0) * rho[i, j]
+
+    # integrate
+    u[i, j] = u[i, j] + dt * f_g
+
+    # fade
+    rho[i, j] = rho[i, j] * (1.0 - 0.1 * dt)
+
+
+@wp.kernel
+def init(rho: wp.array2d(dtype=float), u: wp.array2d(dtype=wp.vec2), radius: int, dir: wp.vec2):
+    i, j = wp.tid()
+
+    d = wp.length(wp.vec2(float(i - grid_width / 2), float(j - grid_height / 2)))
+
+    if d < radius:
+        rho[i, j] = 1.0
+        u[i, j] = dir
+
+
+class Example:
+    def __init__(self, **kwargs):
+
+        self.sim_fps = 60.0
+        self.sim_substeps = 2
+        self.iterations = 100
+        self.sim_dt = (1.0 / self.sim_fps) / self.sim_substeps
+        self.sim_time = 0.0
+
+        shape = (grid_width, grid_height)
+
+        self.u0 = wp.zeros(shape, dtype=wp.vec2)
+        self.u1 = wp.zeros(shape, dtype=wp.vec2)
+
+        self.rho0 = wp.zeros(shape, dtype=float)
+        self.rho1 = wp.zeros(shape, dtype=float)
+
+        self.p0 = wp.zeros(shape, dtype=float)
+        self.p1 = wp.zeros(shape, dtype=float)
+        self.div = wp.zeros(shape, dtype=float)
+
+        # capture pressure solve as a CUDA graph
+        self.use_graph = wp.get_device().is_cuda
+        if self.use_graph:
+            with wp.ScopedCapture() as capture:
+                self.pressure_iterations()
+            self.graph = capture.graph
+
+    def step(self):
+        with wp.ScopedTimer("step"):
+            for _ in range(self.sim_substeps):
+                shape = (grid_width, grid_height)
+                dt = self.sim_dt
+
+                speed = 400.0
+                angle = math.sin(self.sim_time * 4.0) * 1.5
+                vel = wp.vec2(math.cos(angle) * speed, math.sin(angle) * speed)
+
+                # update emitters
+                wp.launch(init, dim=shape, inputs=[self.rho0, self.u0, 5, vel])
+
+                # force integrate
+                wp.launch(integrate, dim=shape, inputs=[self.u0, self.rho0, dt])
+                wp.launch(divergence, dim=shape, inputs=[self.u0, self.div])
+
+                # pressure solve
+                self.p0.zero_()
+                self.p1.zero_()
+
+                if self.use_graph:
+                    wp.capture_launch(self.graph)
+                else:
+                    self.pressure_iterations()
+
+                # velocity update
+                wp.launch(pressure_apply, dim=shape, inputs=[self.p0, self.u0])
+
+                # semi-Lagrangian advection
+                wp.launch(advect, dim=shape, inputs=[self.u0, self.u1, self.rho0, self.rho1, dt])
+
+                # swap buffers
+                (self.u0, self.u1) = (self.u1, self.u0)
+                (self.rho0, self.rho1) = (self.rho1, self.rho0)
+
+                self.sim_time += dt
+
+    def render(self):
+        pass
+
+    def pressure_iterations(self):
+        for _ in range(self.iterations):
+            wp.launch(pressure_solve, dim=self.p0.shape, inputs=[self.p0, self.p1, self.div])
+
+            # swap pressure fields
+            (self.p0, self.p1) = (self.p1, self.p0)
+
+    def step_and_render_frame(self, frame_num=None, img=None):
+        self.step()
+
+        with wp.ScopedTimer("render"):
+            if img:
+                img.set_array(self.rho0.numpy())
+
+        return (img,)
+
+
+if __name__ == "__main__":
+    import matplotlib
+    import matplotlib.animation as anim
+    import matplotlib.pyplot as plt
+
+    example = Example()
+
+    fig = plt.figure()
+
+    img = plt.imshow(example.rho0.numpy(), origin="lower", animated=True, interpolation="antialiased")
+    img.set_norm(matplotlib.colors.Normalize(0.0, 1.0))
+    seq = anim.FuncAnimation(
+        fig, example.step_and_render_frame, fargs=(img,), frames=100000, blit=True, interval=8, repeat=False
+    )
+
+    plt.show()

warp/examples/core/example_graph_capture.py

@@ -0,0 +1,129 @@
+# Copyright (c) 2022 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 Graph Capture
+#
+# Shows how to implement CUDA graph capture using Python's try-finally
+# pattern. The finally block ensures wp.capture_end() gets called, even
+# if an exception occurs during capture, which would otherwise
+# trap the stream in a capturing state.
+#
+###########################################################################
+
+import warp as wp
+import numpy as np
+import matplotlib.pyplot as plt
+import matplotlib.colors
+from matplotlib.animation import FuncAnimation
+
+wp.init()
+
+
+@wp.kernel
+def fbm(
+    kernel_seed: int,
+    frequency: float,
+    amplitude: float,
+    x: wp.array(dtype=float),
+    y: wp.array(dtype=float),
+    z: wp.array2d(dtype=float),
+):
+    i, j = wp.tid()
+    state = wp.rand_init(kernel_seed)
+
+    p = frequency * wp.vec2(x[j], y[i])
+    n = amplitude * wp.noise(state, p)
+
+    z[i, j] += n
+
+
+@wp.kernel
+def slide(x: wp.array(dtype=float), shift: float):
+    tid = wp.tid()
+    x[tid] += shift
+
+
+class Example:
+    def __init__(self):
+        self.W = 128
+        self.H = 128
+        min_x, max_x = 0.0, 2.0
+        min_y, max_y = 0.0, 2.0
+
+        # create a grid of pixels
+        x = np.linspace(min_x, max_x, self.W)
+        y = np.linspace(min_y, max_y, self.H)
+
+        self.x = wp.array(x, dtype=float)
+        self.y = wp.array(y, dtype=float)
+        self.pixel_values = wp.zeros((self.W, self.H), dtype=float)
+
+        self.seed = 42
+        self.shift = 2e-2
+        self.frequency = 1.0
+        self.amplitude = 1.0
+
+        # use graph capture if launching from a CUDA-capable device
+        self.use_graph = wp.get_device().is_cuda
+        if self.use_graph:
+            # record launches
+            with wp.ScopedCapture() as capture:
+                self.fbm()
+            self.graph = capture.graph
+
+    def fbm(self):
+        for _ in range(16):
+            wp.launch(
+                kernel=fbm,
+                dim=(self.H, self.W),
+                inputs=[self.seed, self.frequency, self.amplitude, self.x, self.y],
+                outputs=[self.pixel_values],
+            )
+            self.frequency *= 2.0
+            self.amplitude *= 0.5
+
+    def step(self):
+        self.pixel_values.zero_()
+        self.frequency = 1.0
+        self.amplitude = 1.0
+
+        with wp.ScopedTimer("step", active=True):
+            wp.launch(kernel=slide, dim=self.W, inputs=[self.x, self.shift])
+
+            if self.use_graph:
+                wp.capture_launch(self.graph)
+            else:  # cpu path
+                self.fbm()
+
+    def render(self):
+        pass
+
+    def step_and_render(self, frame_num=None, img=None):
+        self.step()
+
+        with wp.ScopedTimer("render"):
+            if img:
+                pixels = self.pixel_values.numpy()
+                pixels = (pixels + 1.0) / 2.0
+                img.set_array(pixels)
+
+        return (img,)
+
+
+if __name__ == "__main__":
+    example = Example()
+
+    # Create the animation
+    fig = plt.figure()
+    img = plt.imshow(example.pixel_values.numpy(), "gray", origin="lower", animated=True)
+    img.set_norm(matplotlib.colors.Normalize(0.0, 1.0))
+
+    ani = FuncAnimation(fig, example.step_and_render, fargs=(img,), frames=1000, interval=30)
+
+    # Display the animation
+    plt.show()

warp/examples/core/example_marching_cubes.py

@@ -0,0 +1,177 @@
+# Copyright (c) 2022 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 Marching Cubes
+#
+# Shows how use the built-in marching cubes functionality to extract
+# the iso-surface from a density field.
+#
+# Note: requires a CUDA-capable device
+###########################################################################
+
+
+import os
+from typing import Optional
+
+import warp as wp
+import warp.render
+
+wp.init()
+
+
+@wp.func
+def sdf_create_box(pos: wp.vec3, size: wp.vec3):
+    """Creates a SDF box primitive."""
+    # https://iquilezles.org/articles/distfunctions
+    q = wp.vec3(
+        wp.abs(pos[0]) - size[0],
+        wp.abs(pos[1]) - size[1],
+        wp.abs(pos[2]) - size[2],
+    )
+    qp = wp.vec3(wp.max(q[0], 0.0), wp.max(q[1], 0.0), wp.max(q[2], 0.0))
+    return wp.length(qp) + wp.min(wp.max(q[0], wp.max(q[1], q[2])), 0.0)
+
+
+@wp.func
+def sdf_create_torus(pos: wp.vec3, major_radius: float, minor_radius: float):
+    """Creates a SDF torus primitive."""
+    # https://iquilezles.org/articles/distfunctions
+    q = wp.vec2(wp.length(wp.vec2(pos[0], pos[2])) - major_radius, pos[1])
+    return wp.length(q) - minor_radius
+
+
+@wp.func
+def sdf_translate(pos: wp.vec3, offset: wp.vec3):
+    """Translates a SDF position vector with an offset."""
+    return pos - offset
+
+
+@wp.func
+def sdf_rotate(pos: wp.vec3, angles: wp.vec3):
+    """Rotates a SDF position vector using Euler angles."""
+    rot = wp.quat_rpy(
+        wp.radians(angles[0]),
+        wp.radians(angles[1]),
+        wp.radians(angles[2]),
+    )
+    return wp.quat_rotate_inv(rot, pos)
+
+
+@wp.func
+def sdf_smooth_min(a: float, b: float, radius: float):
+    """Creates a SDF torus primitive."""
+    # https://iquilezles.org/articles/smin
+    h = wp.max(radius - wp.abs(a - b), 0.0) / radius
+    return wp.min(a, b) - h * h * h * radius * (1.0 / 6.0)
+
+
+@wp.kernel(enable_backward=False)
+def make_field(
+    torus_altitude: float,
+    torus_major_radius: float,
+    torus_minor_radius: float,
+    smooth_min_radius: float,
+    dim: int,
+    time: float,
+    out_data: wp.array3d(dtype=float),
+):
+    """Kernel to generate a SDF volume based on primitives."""
+    i, j, k = wp.tid()
+
+    # Retrieve the position of the current cell in a normalized [-1, 1] range
+    # for each dimension.
+    pos = wp.vec3(
+        2.0 * ((float(i) + 0.5) / float(dim)) - 1.0,
+        2.0 * ((float(j) + 0.5) / float(dim)) - 1.0,
+        2.0 * ((float(k) + 0.5) / float(dim)) - 1.0,
+    )
+
+    box = sdf_create_box(
+        sdf_translate(pos, wp.vec3(0.0, -0.7, 0.0)),
+        wp.vec3(0.9, 0.3, 0.9),
+    )
+    torus = sdf_create_torus(
+        sdf_rotate(
+            sdf_translate(pos, wp.vec3(0.0, torus_altitude, 0.0)),
+            wp.vec3(wp.sin(time) * 90.0, wp.cos(time) * 45.0, 0.0),
+        ),
+        torus_major_radius,
+        torus_minor_radius,
+    )
+    out_data[i, j, k] = sdf_smooth_min(box, torus, smooth_min_radius)
+
+
+class Example:
+    def __init__(self, stage: Optional[str]):
+        self.dim = 64
+        self.max_verts = int(1e6)
+        self.max_tris = int(1e6)
+
+        self.torus_altitude = -0.5
+        self.torus_major_radius = 0.5
+        self.torus_minor_radius = 0.1
+        self.smooth_min_radius = 0.5
+
+        self.fps = 60.0
+        self.frame = 0
+
+        self.field = wp.zeros((self.dim, self.dim, self.dim), dtype=float)
+        self.mc = wp.MarchingCubes(self.dim, self.dim, self.dim, self.max_verts, self.max_tris)
+
+        self.renderer = None
+        if stage is not None:
+            self.renderer = wp.render.UsdRenderer(stage)
+
+    def step(self):
+        with wp.ScopedTimer("Update Field"):
+            wp.launch(
+                make_field,
+                dim=self.field.shape,
+                inputs=(
+                    self.torus_altitude,
+                    self.torus_major_radius,
+                    self.torus_minor_radius,
+                    self.smooth_min_radius,
+                    self.dim,
+                    self.frame / self.fps,
+                ),
+                outputs=(self.field,),
+            )
+
+        with wp.ScopedTimer("Surface Extraction"):
+            self.mc.surface(self.field, 0.0)
+
+    def render(self):
+        if self.renderer is None:
+            return
+
+        with wp.ScopedTimer("Render"):
+            self.renderer.begin_frame(self.frame / self.fps)
+            self.renderer.render_mesh(
+                "surface",
+                self.mc.verts.numpy(),
+                self.mc.indices.numpy(),
+                colors=((0.35, 0.55, 0.9),) * len(self.mc.verts),
+                update_topology=True,
+            )
+            self.renderer.end_frame()
+
+
+if __name__ == "__main__":
+    this_dir = os.path.realpath(os.path.dirname(__file__))
+    this_file = os.path.basename(__file__).split(".")[0]
+    stage_path = os.path.join(this_dir, f"{this_file}.usd")
+
+    example = Example(stage_path)
+    for _ in range(240):
+        example.step()
+        example.render()
+        example.frame += 1
+
+    if example.renderer is not None:
+        example.renderer.save()