def benchmark_nested_struct_fill_and_clear():
a = ti.var(dt=ti.f32)
N = 512
@ti.layout
def place():
ti.root.pointer(ti.ij, [N, N]).dense(ti.ij, [8, 8]).place(a)
@ti.kernel
def fill():
for i, j in ti.ndrange(N * 8, N * 8):
a[i, j] = 2.0
@ti.kernel
def clear():
for i, j in a.parent():
ti.deactivate(a.parent().parent(), [i, j])
def task():
fill()
clear()
return ti.benchmark(task, repeat=30)
def test_grouped():
ti.get_runtime().print_preprocessed = True
val = ti.var(ti.i32)
n = 4
m = 8
p = 16
@ti.layout
def values():
ti.root.dense(ti.i, n).dense(ti.j, m).dense(ti.k, p).place(val)
@ti.kernel
def test():
for I in ti.grouped(val):
val[I] = I[0] + I[1] * 2 + I[2] * 3
test()
for i in range(n):
for j in range(m):
for k in range(p):
assert val[i, j, k] == i + j * 2 + k * 3
def test_io_simple():
n = 32
x1 = ti.var(ti.f32, shape=(n, n))
t1 = torch.tensor(2 * np.ones((n, n), dtype=np.float32))
x2 = ti.Matrix(2, 3, ti.f32, shape=(n, n))
t2 = torch.tensor(2 * np.ones((n, n, 2, 3), dtype=np.float32))
x1.from_torch(t1)
for i in range(n):
for j in range(n):
assert x1[i, j] == 2
x2.from_torch(t2)
for i in range(n):
for j in range(n):
for k in range(2):
for l in range(3):
assert x2[i, j][k, l] == 2
t3 = x2.to_torch()
assert (t2 == t3).all()
def test_atomic_add_with_local_store_simplify2():
# Test for the following LocalStoreStmt simplification case:
#
# local store [$a <- ...]
# atomic add ($a, ...)
#
# Specifically, the local store should not be removed, because
# atomic_add can return its value.
x = ti.var(ti.i32)
step = 42
ti.root.dense(ti.i, n).place(x)
@ti.kernel
def func():
for i in range(n):
j = i
x[i] = ti.atomic_add(j, step)
func()
for i in range(n):
assert x[i] == i
def benchmark_nested_range_blocked():
a = ti.var(dt=ti.f32)
N = 512
@ti.layout
def place():
ti.root.dense(ti.ij, [N, N]).dense(ti.ij, [8, 8]).place(a)
@ti.kernel
def fill():
for X in range(N * N):
for Y in range(64):
a[X // N * 8 + Y // 8, X % N * 8 + Y % 8] = 2.0
fill()
ti.get_runtime().sync()
t = time.time()
for n in range(100):
fill()
elapsed = time.time() - t
ti.get_runtime().sync()
return elapsed / 100
def test_loop_arg_as_range():
# Dynamic range loops are intended to make sure global tmps work
x = ti.var(ti.i32)
n = 1000
@ti.layout
def layout():
ti.root.dense(ti.i, n).place(x)
@ti.kernel
def test(b: ti.i32, e: ti.i32):
for i in range(b, e):
x[i - b] = i
pairs = [
(0, n // 2),
(n // 2, n),
(-n // 2, -n // 3),
]
for b, e in pairs:
test(b, e)
for i in range(b, e):
assert x[i - b] == i
def test_numpy():
val = ti.var(ti.i32)
n = 4
@ti.layout
def values():
ti.root.dense(ti.i, n).place(val)
@ti.kernel
def test_numpy(arr: np.ndarray):
for i in range(n):
arr[i] = arr[i] ** 2
a = np.array([4, 8, 1, 24], dtype=np.float32)
for i in range(n):
a[i] = i * 2
test_numpy(a)
for i in range(n):
assert a[i] == i * i * 4
def with_data_type(dt):
val = ti.var(ti.i32)
n = 4
@ti.layout
def values():
ti.root.dense(ti.i, n).place(val)
@ti.kernel
def test_numpy(arr: ti.ext_arr()):
for i in range(n):
arr[i] = arr[i]**2
a = np.array([4, 8, 1, 24], dtype=dt)
for i in range(n):
a[i] = i * 2
test_numpy(a)
for i in range(n):
assert a[i] == i * i * 4
def __init__(
self,
nx, # domain size
ny,
niu, # viscosity of fluid
bc_type, # [left,top,right,bottom] boundary conditions: 0 -> Dirichlet ; 1 -> Neumann
bc_value, # if bc_type = 0, we need to specify the velocity in bc_value
cy=0, # whether to place a cylindrical obstacle
cy_para=[0.0, 0.0, 0.0], # location and radius of the cylinder
steps=60000): # total steps to run
self.nx = nx # by convention, dx = dy = dt = 1.0 (lattice units)
self.ny = ny
self.niu = niu
self.tau = 3.0 * niu + 0.5
self.inv_tau = 1.0 / self.tau
self.rho = ti.var(dt=ti.f32, shape=(nx, ny))
self.vel = ti.Vector(2, dt=ti.f32, shape=(nx, ny))
self.mask = ti.var(dt=ti.f32, shape=(nx, ny))
self.f_old = ti.Vector(9, dt=ti.f32, shape=(nx, ny))
self.f_new = ti.Vector(9, dt=ti.f32, shape=(nx, ny))
self.w = ti.var(dt=ti.f32, shape=9)
self.e = ti.var(dt=ti.i32, shape=(9, 2))
self.bc_type = ti.var(dt=ti.i32, shape=4)
self.bc_value = ti.var(dt=ti.f32, shape=(4, 2))
self.cy = cy
self.cy_para = ti.var(dt=ti.f32, shape=3)
self.bc_type.from_numpy(np.array(bc_type, dtype=np.int32))
self.bc_value.from_numpy(np.array(bc_value, dtype=np.float32))
self.cy_para.from_numpy(np.array(cy_para, dtype=np.float32))
self.steps = steps
arr = np.array([
4.0 / 9.0, 1.0 / 9.0, 1.0 / 9.0, 1.0 / 9.0, 1.0 / 9.0, 1.0 / 36.0,
1.0 / 36.0, 1.0 / 36.0, 1.0 / 36.0
],
dtype=np.float32)
self.w.from_numpy(arr)
arr = np.array([[0, 0], [1, 0], [0, 1], [-1, 0], [0, -1], [1, 1],
[-1, 1], [-1, -1], [1, -1]],
dtype=np.int32)
self.e.from_numpy(arr)
def test_argument_error():
x = ti.var(ti.i32)
@ti.layout
def layout():
ti.root.place(x)
try:
@ti.kernel
def set_i32_notype(v):
pass
except ti.KernelDefError:
pass
try:
@ti.kernel
def set_i32_args(*args):
pass
except ti.KernelDefError:
pass
try:
@ti.kernel
def set_i32_kwargs(**kwargs):
pass
except ti.KernelDefError:
pass
@ti.kernel
def set_i32(v: ti.i32):
x[None] = v
set_i32(123)
assert x[None] == 123
def test_local_store_in_nested_for_and_if():
# See https://github.com/taichi-dev/taichi/pull/862.
val = ti.var(ti.i32, shape=(3, 3, 3))
@ti.kernel
def func():
ti.serialize()
for i, j, k in val:
if i < 2 and j < 2 and k < 2:
a = 0
for di, dj, dk in ti.ndrange((0, 2), (0, 2), (0, 2)):
if val[i + di, j + dj, k + dk] == 1:
a = val[i + di, j + dj, k + dk]
for di, dj, dk in ti.ndrange((0, 2), (0, 2), (0, 2)):
val[i + di, j + dj, k + dk] = a
val[1, 1, 1] = 1
func()
for i in range(3):
for j in range(3):
for k in range(3):
assert (val[i, j, k] == 1)
def test_numpy_2d_transpose():
val = ti.var(ti.i32)
n = 8
m = 8
ti.root.dense(ti.ij, (n, m)).place(val)
@ti.kernel
def test_numpy(arr: ti.ext_arr()):
for i in ti.grouped(val):
val[i] = arr[i]
a = np.empty(shape=(n, m), dtype=np.int32)
for i in range(n):
for j in range(m):
a[i, j] = i * j + i * 4
test_numpy(a.transpose())
for i in range(n):
for j in range(m):
assert val[i, j] == i * j + j * 4
def test_pointer2():
x = ti.var(ti.f32)
n = 16
@ti.layout
def place():
ti.root.dense(ti.i, n).pointer().dense(ti.i, n).place(x)
@ti.kernel
def func():
for i in range(n * n):
x[i] = 1.0
@ti.kernel
def set10():
x[10] = 10.0
@ti.kernel
def clear():
for i in x.parent().parent():
ti.deactivate(x.parent().parent(), i)
func()
clear()
for i in range(n * n):
assert x[i] == 0.0
set10()
for i in range(n * n):
if i != 10:
assert x[i] == 0.0
else:
assert x[i] == 10.0
def test_scope():
# In the future the following code should throw an exception at the python front end
# instead of crashing the compiler
return
ti.runtime.print_preprocessed = True
for arch in [ti.x86_64, ti.cuda]:
# ti.reset()
ti.cfg.arch = arch
x = ti.var(ti.f32)
N = 1
@ti.layout
def place():
ti.root.dense(ti.i, N).place(x)
@ti.kernel
def func():
if 1 > 0:
val = 1
ti.print(val)
func()
def test_static_grouped_ndrange():
val = ti.var(ti.i32)
n = 4
m = 8
ti.root.dense(ti.ij, (n, m)).place(val)
x0 = 2
y0 = 3
x1 = 1
y1 = 6
@ti.kernel
def test():
for I in ti.static(ti.grouped(ti.ndrange((x0, y0), (x1, y1)))):
val[I] = I[0] + I[1] * 2
test()
for i in range(n):
for j in range(m):
assert val[i, j] == (i +
j * 2 if x0 <= i < y0 and x1 <= j < y1 else 0)
def test_bitmasked_offset_child():
x = ti.var(ti.i32)
x2 = ti.var(ti.i32)
y = ti.var(ti.i32)
y2 = ti.var(ti.i32)
y3 = ti.var(ti.i32)
z = ti.var(ti.i32)
s = ti.var(ti.i32, shape=())
n = 16
# Offset children:
# * In |bm|'s cell: |bm2| has a non-zero offset
# * In |bm2|'s cell: |z| has a non-zero offset
# * We iterate over |z| to test the listgen handles offsets correctly
bm = ti.root.bitmasked(ti.i, n)
bm.dense(ti.i, 16).place(x, x2)
bm2 = bm.bitmasked(ti.i, 4)
bm2.dense(ti.i, 4).place(y, y2, y3)
bm2.bitmasked(ti.i, 4).place(z)
@ti.kernel
def func():
for _ in z:
s[None] += 1
z[0] = 1
z[7] = 1
z[42] = 1
z[53] = 1
z[88] = 1
z[101] = 1
z[233] = 1
func()
assert s[None] == 7
def __init__(self):
self.dim = 2
self.inf = 1e10
self.epsilon = 1e-5
self.on = 100
self.vn = 1000
self.en = 1000
self.node = ti.Vector(self.dim,
dt=ti.f32,
shape=self.vn,
needs_grad=True)
self.prev_node = ti.Vector(self.dim, dt=ti.f32, shape=self.vn)
self.prev_t_node = ti.Vector(self.dim, dt=ti.f32, shape=self.vn)
self.bar_node = ti.Vector(self.dim, dt=ti.f32, shape=self.vn)
self.p = ti.Vector(self.dim, dt=ti.f32, shape=self.vn)
self.element = ti.Vector(self.dim + 1, dt=ti.i32, shape=self.en)
# the end index of i's object
self.vn_object_index = ti.var(dt=ti.i32, shape=self.on)
self.en_object_index = ti.var(dt=ti.i32, shape=self.on)
self.count = ti.var(dt=ti.i32, shape=())
# the inverse obj id of each node and ele
self.node_obj_idx = ti.var(dt=ti.i32, shape=self.vn)
self.element_obj_idx = ti.var(dt=ti.i32, shape=self.en)
## for simulation
self.E = 6000 # Young modulus
self.nu = 0.4 # Poisson's ratio: nu \in [0, 0.5)
self.mu = self.E / (2 * (1 + self.nu))
self.la = self.E * self.nu / ((1 + self.nu) * (1 - 2 * self.nu))
self.dt = 5e-3
self.bar_d = 0.1
self.k = 1 # contact stiffness
# self.velocity = ti.Vector(self.dim, dt=ti.f32, shape=self.vn)
self.node_mass = ti.var(dt=ti.f32, shape=self.vn)
self.element_mass = ti.var(dt=ti.f32, shape=self.en)
self.element_volume = ti.var(dt=ti.f32, shape=self.en)
self.energy = ti.var(dt=ti.f32, shape=(), needs_grad=True)
self.prev_energy = ti.var(dt=ti.f32, shape=())
self.B = ti.Matrix(self.dim, self.dim, dt=ti.f32, shape=self.en)
self.neighbor_element_count = ti.var(dt=ti.i32, shape=self.vn)
## for rendering
self.begin_point = ti.Vector(self.dim, ti.f32, shape=(self.en * 3))
self.end_point = ti.Vector(self.dim, ti.f32, shape=(self.en * 3))
self.node_energy = ti.var(dt=ti.f32, shape=self.vn)
self.edge_energy = ti.var(dt=ti.f32, shape=(self.en * 3))
self.score = 0
self.rendering_u0 = ti.var(dt=ti.f32, shape=())
self.rendering_u1 = ti.var(dt=ti.f32, shape=())
self.rendering_u2 = ti.var(dt=ti.f32, shape=())
self.rendering_u3 = ti.var(dt=ti.f32, shape=())
self.game_over = ti.var(dt=ti.i32, shape=())
## the controlled object
self.ctrl_obj = ti.var(dt=ti.i32, shape=())
self.move_d = ti.var(dt=ti.f32, shape=())
self.ctrl_obj[None] = -1
self.move_d[None] = 1e-3
import taichi as ti
quality = 2 # Use a larger value for higher-res simulations
n_particles, n_grid = 9000 * quality ** 2, 128 * quality
dx, inv_dx = 1 / n_grid, float(n_grid)
dt = 1e-4 / quality
p_vol, p_rho = (dx * 0.5)**2, 1
p_mass = p_vol * p_rho
E, nu = 0.1e4, 0.2 # Young's modulus and Poisson's ratio
mu_0, lambda_0 = E / (2 * (1 + nu)), E * nu / ((1+nu) * (1 - 2 * nu)) # Lame parameters
x = ti.Vector(2, dt=ti.f32, shape=n_particles) # position
v = ti.Vector(2, dt=ti.f32, shape=n_particles) # velocity
C = ti.Matrix(2, 2, dt=ti.f32, shape=n_particles) # affine velocity field
F = ti.Matrix(2, 2, dt=ti.f32, shape=n_particles) # deformation gradient
material = ti.var(dt=ti.i32, shape=n_particles) # material id
Jp = ti.var(dt=ti.f32, shape=n_particles) # plastic deformation
grid_v = ti.Vector(2, dt=ti.f32, shape=(n_grid, n_grid)) # grid node momemtum/velocity
grid_m = ti.var(dt=ti.f32, shape=(n_grid, n_grid)) # grid node mass
ti.cfg.arch = ti.cuda # Try to run on GPU
@ti.kernel
def substep():
for i, j in ti.ndrange(n_grid, n_grid):
grid_v[i, j] = [0, 0]
grid_m[i, j] = 0
for p in range(n_particles): # Particle state update and scatter to grid (P2G)
base = (x[p] * inv_dx - 0.5).cast(int)
fx = x[p] * inv_dx - base.cast(float)
# Quadratic kernels [http://mpm.graphics Eqn. 123, with x=fx, fx-1,fx-2]
w = [0.5 * ti.sqr(1.5 - fx), 0.75 - ti.sqr(fx - 1), 0.5 * ti.sqr(fx - 0.5)]
F[p] = (ti.Matrix.identity(ti.f32, 2) + dt * C[p]) @ F[p] # deformation gradient update
def on_init(self, n=512):
self.n = n
self.title = 'Julia Set'
self.img = ti.var(ti.f32, (self.n * 2, self.n))
self.colormap = cm.get_cmap('magma')
self.define_input()
import taichi as ti
ti.init()
x = ti.var(ti.i32)
y = ti.var(ti.i32)
ti.root.pointer(ti.ij, 4).dense(ti.ij, 8).place(x, y)
@ti.kernel
def copy():
for i, j in y:
x[i, j] = y[i, j]
copy()
import time from matplotlib.pyplot import cm import taichi as tc real = ti.f32 ti.set_default_fp(real) max_steps = 4096 vis_interval = 4 output_vis_interval = 16 steps = 204 assert steps * 2 <= max_steps vis_resolution = 1024 scalar = lambda: ti.var(dt=real) vec = lambda: ti.Vector(2, dt=real) loss = scalar() x = vec() v = vec() goal = [0.9, 0.15] n_objects = 1 ground_height = 0.1 @ti.layout def place():
import taichi as ti
ti.init(arch=ti.cpu)
n = 320
pixels = ti.var(dt=ti.f32, shape=(n * 2, n))
@ti.func
def complex_sqr(z):
return ti.Vector([z[0] ** 2 - z[1] ** 2, z[1] * z[0] * 2])
@ti.kernel
S
for i, j in pixels: # 对于所有像素，并行执行
c = ti.Vector([-0.8, ti.sin(t) * 0.2])
z = ti.Vector([float(i) / n - 1, float(j) / n - 0.5]) * 2
iterations = 0
while z.norm() < 20 and iterations < 50:
z = complex_sqr(z) + c
iterations += 1
pixels[i, j] = 1 - iterations * 0.02
gui = ti.GUI("Fractal", (n * 2, n))
for i in range(1000000):
paint(i * 0.03)
gui.set_image(pixels)
gui.show()
dim = 2
n_particles = 8192
n_grid = 128
dx = 1 / n_grid
inv_dx = 1 / dx
dt = 2.0e-4
p_vol = (dx * 0.5)**2
p_rho = 1
p_mass = p_vol * p_rho
E = 400
x = ti.Vector(dim, dt=ti.f32, shape=n_particles)
v = ti.Vector(dim, dt=ti.f32, shape=n_particles)
C = ti.Matrix(dim, dim, dt=ti.f32, shape=n_particles)
J = ti.var(dt=ti.f32, shape=n_particles)
grid_v = ti.Vector(dim, dt=ti.f32, shape=(n_grid, n_grid))
grid_m = ti.var(dt=ti.f32, shape=(n_grid, n_grid))
ti.cfg.arch = ti.cuda
@ti.kernel
def substep():
for p in x:
base = (x[p] * inv_dx - 0.5).cast(int)
fx = x[p] * inv_dx - base.cast(float)
w = [
0.5 * ti.sqr(1.5 - fx), 0.75 - ti.sqr(fx - 1),
0.5 * ti.sqr(fx - 0.5)
]
import taichi as ti
import taichi_glsl as ts
import taichi_three as t3
ti.init(ti.opengl, kernel_profiler=True)
scene = t3.Scene()
model = t3.Model()
scene.add_model(model)
N = 2**12
faces = t3.Face.var()
vertices = t3.Vertex.var()
ti.root.dense(ti.i, N * 3).place(vertices)
ti.root.dense(ti.i, N).place(faces)
vertices_len = ti.var(ti.i32, ())
faces_len = ti.var(ti.i32, ())
model.set_vertices(vertices)
model.add_geometry(faces)
@ti.func
def glVertex(pos):
l = ti.atomic_add(vertices_len[None], 1)
vertices.pos[l] = pos
return l
@ti.func
def glFace(idx):
l = ti.atomic_add(faces_len[None], 1)
import taichi as ti
import random
n = 8
x = ti.var(dt=ti.f32)
y = ti.var(dt=ti.f32)
L = ti.var(dt=ti.f32)
@ti.layout
def data():
ti.root.dense(ti.i, n).place(x, y, x.grad, y.grad) # place gradient tensors
ti.root.place(L, L.grad)
@ti.kernel
def reduce():
global L
for i in range(n):
ti.atomic_add(L, 0.5 * (x[i] - y[i]) ** 2)
# Initialize vectors
for i in range(n):
x[i] = random.random()
y[i] = random.random()
@ti.kernel
def update():
for i in x:
x[i] -= x.grad[i] * 0.1
# Optimize with 100 gradient descent iterations
for k in range(100):
lambda_epsilon = 100.0
pbf_num_iters = 5
corr_deltaQ_coeff = 0.3
corrK = 0.001
# Need ti.pow()
# corrN = 4.0
neighbor_radius = h * 1.05
poly6_factor = 315.0 / 64.0 / np.pi
spiky_grad_factor = -45.0 / np.pi
old_positions = ti.Vector(dim, dt=ti.f32)
positions = ti.Vector(dim, dt=ti.f32)
velocities = ti.Vector(dim, dt=ti.f32)
# Once taichi supports clear(), we can get rid of grid_num_particles
grid_num_particles = ti.var(ti.i32)
grid2particles = ti.var(ti.i32)
particle_num_neighbors = ti.var(ti.i32)
particle_neighbors = ti.var(ti.i32)
lambdas = ti.var(ti.f32)
position_deltas = ti.Vector(dim, dt=ti.f32)
# 0: x-pos, 1: timestep in sin()
board_states = ti.Vector(2, dt=ti.f32)
@ti.layout
def layout():
ti.root.dense(ti.i, num_particles).place(old_positions, positions,
velocities)
grid_snode = ti.root.dense(ti.ij, grid_size)
grid_snode.place(grid_num_particles)
import taichi as ti ti.init() a = ti.var(dt=ti.f32, shape=(42, 63))
import cv2 import os real = ti.f32 ti.set_default_fp(real) # ti.cfg.print_ir = True max_steps = 4096 vis_interval = 256 output_vis_interval = 8 steps = 2048 // 2 assert steps * 2 <= max_steps vis_resolution = 1024 scalar = lambda: ti.var(dt=real) vec = lambda: ti.Vector(2, dt=real) loss = scalar() x = vec() v = vec() v_inc = vec() head_id = 10 goal = vec() n_objects = 0 # target_ball = 0 elasticity = 0.0 ground_height = 0.1
def __init__(self, n, m):
self.n = n
self.m = m
self.val = ti.var(ti.f32, shape=(n, m))
def __init__(self, n, m, increment):
self.n = n
self.m = m
self.val = ti.var(ti.f32)
self.total = ti.var(ti.f32)
self.increment = increment