warp-lang 1.7.2rc1__py3-none-macosx_10_13_universal2.whl → 1.8.0__py3-none-macosx_10_13_universal2.whl

warp/tests/tile/test_tile.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import unittest
+from typing import Any
 
 import numpy as np
 
@@ -360,13 +361,208 @@ def test_tile_operators(test, device):
     assert_np_equal(input_wp.grad.numpy(), np.ones_like(input) * 0.75)
 
 
+@wp.kernel
+def test_tile_tile_preserve_type_kernel(x: wp.array(dtype=Any), y: wp.array(dtype=Any)):
+    a = x[0]
+    t = wp.tile(a, preserve_type=True)
+    wp.tile_store(y, t)
+
+
+@wp.kernel
+def test_tile_tile_scalar_expansion_kernel(x: wp.array(dtype=float), y: wp.array(dtype=float)):
+    a = x[0]
+    t = wp.tile(a)
+    wp.tile_store(y, t)
+
+
+@wp.kernel
+def test_tile_tile_vec_expansion_kernel(x: wp.array(dtype=wp.vec3), y: wp.array2d(dtype=float)):
+    a = x[0]
+    t = wp.tile(a)
+    wp.tile_store(y, t)
+
+
+@wp.kernel
+def test_tile_tile_mat_expansion_kernel(x: wp.array(dtype=wp.mat33), y: wp.array3d(dtype=float)):
+    a = x[0]
+    t = wp.tile(a)
+    wp.tile_store(y, t)
+
+
+def test_tile_tile(test, device):
+    # preserve type
+    def test_func_preserve_type(type: Any):
+        x = wp.ones(1, dtype=type, requires_grad=True, device=device)
+        y = wp.zeros((TILE_DIM), dtype=type, requires_grad=True, device=device)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_tile_tile_preserve_type_kernel,
+                dim=[TILE_DIM],
+                inputs=[x],
+                outputs=[y],
+                block_dim=TILE_DIM,
+                device=device,
+            )
+
+        y.grad = wp.ones_like(y)
+
+        tape.backward()
+
+        assert_np_equal(y.numpy(), wp.full((TILE_DIM), type(1.0), dtype=type, device="cpu").numpy())
+        assert_np_equal(x.grad.numpy(), wp.full((1,), type(TILE_DIM), dtype=type, device="cpu").numpy())
+
+    test_func_preserve_type(float)
+    test_func_preserve_type(wp.vec3)
+    test_func_preserve_type(wp.quat)
+    test_func_preserve_type(wp.mat33)
+
+    # scalar expansion
+    x = wp.ones(1, dtype=float, requires_grad=True, device=device)
+    y = wp.zeros((TILE_DIM), dtype=float, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(
+            test_tile_tile_scalar_expansion_kernel,
+            dim=[TILE_DIM],
+            inputs=[x],
+            outputs=[y],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    y.grad = wp.ones_like(y)
+
+    tape.backward()
+
+    assert_np_equal(y.numpy(), wp.full((TILE_DIM), 1.0, dtype=float, device="cpu").numpy())
+    assert_np_equal(x.grad.numpy(), wp.full((1,), wp.float32(TILE_DIM), dtype=float, device="cpu").numpy())
+
+    # vec expansion
+    x = wp.ones(1, dtype=wp.vec3, requires_grad=True, device=device)
+    y = wp.zeros((3, TILE_DIM), dtype=float, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(
+            test_tile_tile_vec_expansion_kernel,
+            dim=[TILE_DIM],
+            inputs=[x],
+            outputs=[y],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    y.grad = wp.ones_like(y)
+
+    tape.backward()
+
+    assert_np_equal(y.numpy(), wp.full((3, TILE_DIM), 1.0, dtype=float, device="cpu").numpy())
+    assert_np_equal(x.grad.numpy(), wp.full((1,), wp.float32(TILE_DIM), dtype=wp.vec3, device="cpu").numpy())
+
+    # mat expansion
+    x = wp.ones(1, dtype=wp.mat33, requires_grad=True, device=device)
+    y = wp.zeros((3, 3, TILE_DIM), dtype=float, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(
+            test_tile_tile_mat_expansion_kernel,
+            dim=[TILE_DIM],
+            inputs=[x],
+            outputs=[y],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    y.grad = wp.ones_like(y)
+
+    tape.backward()
+
+    assert_np_equal(y.numpy(), wp.full((3, 3, TILE_DIM), 1.0, dtype=float, device="cpu").numpy())
+    assert_np_equal(x.grad.numpy(), wp.full((1,), wp.float32(TILE_DIM), dtype=wp.mat33, device="cpu").numpy())
+
+
+@wp.kernel
+def test_tile_untile_preserve_type_kernel(x: wp.array(dtype=Any), y: wp.array(dtype=Any)):
+    i = wp.tid()
+    a = x[i]
+    t = wp.tile(a, preserve_type=True)
+    b = wp.untile(t)
+    y[i] = b
+
+
+@wp.kernel
+def test_tile_untile_kernel(x: wp.array(dtype=Any), y: wp.array(dtype=Any)):
+    i = wp.tid()
+    a = x[i]
+    t = wp.tile(a)
+    b = wp.untile(t)
+    y[i] = b
+
+
+def test_tile_untile(test, device):
+    def test_func_preserve_type(type: Any):
+        x = wp.ones(TILE_DIM, dtype=type, requires_grad=True, device=device)
+        y = wp.zeros_like(x)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_tile_untile_preserve_type_kernel,
+                dim=TILE_DIM,
+                inputs=[x],
+                outputs=[y],
+                block_dim=TILE_DIM,
+                device=device,
+            )
+
+        y.grad = wp.ones_like(y)
+
+        tape.backward()
+
+        assert_np_equal(y.numpy(), x.numpy())
+        assert_np_equal(x.grad.numpy(), y.grad.numpy())
+
+    test_func_preserve_type(float)
+    test_func_preserve_type(wp.vec3)
+    test_func_preserve_type(wp.quat)
+    test_func_preserve_type(wp.mat33)
+
+    def test_func(type: Any):
+        x = wp.ones(TILE_DIM, dtype=type, requires_grad=True, device=device)
+        y = wp.zeros_like(x)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(test_tile_untile_kernel, dim=TILE_DIM, inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device)
+
+        y.grad = wp.ones_like(y)
+
+        tape.backward()
+
+        assert_np_equal(y.numpy(), x.numpy())
+        assert_np_equal(x.grad.numpy(), y.grad.numpy())
+
+    test_func(float)
+    test_func(wp.vec3)
+    test_func(wp.mat33)
+
+
+@wp.func
+def tile_sum_func(a: wp.tile(dtype=float, shape=(TILE_M, TILE_N))):
+    return wp.tile_sum(a) * 0.5
+
+
 @wp.kernel
 def tile_sum_kernel(input: wp.array3d(dtype=float), output: wp.array(dtype=float)):
     # output tile index
     i = wp.tid()
 
     a = wp.tile_load(input[i], shape=(TILE_M, TILE_N))
-    s = wp.tile_sum(a) * 0.5
+    s = tile_sum_func(a)
 
     wp.tile_store(output, s, offset=i)
 
@@ -728,6 +924,116 @@ def test_tile_broadcast_grad(test, device):
     assert_np_equal(a.grad.numpy(), np.ones(5) * 5.0)
 
 
+@wp.kernel
+def test_tile_squeeze_kernel(x: wp.array3d(dtype=float), y: wp.array(dtype=float)):
+    a = wp.tile_load(x, shape=(1, TILE_M, 1), offset=(0, 0, 0))
+    b = wp.tile_squeeze(a, axis=(2,))
+    c = wp.tile_squeeze(b)
+
+    wp.tile_store(y, c, offset=(0,))
+
+
+def test_tile_squeeze(test, device):
+    x = wp.ones((1, TILE_M, 1), dtype=float, device=device, requires_grad=True)
+    y = wp.zeros((TILE_M,), dtype=float, device=device, requires_grad=True)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch_tiled(test_tile_squeeze_kernel, dim=1, inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.ones((TILE_M,), dtype=np.float32))
+    assert_np_equal(x.grad.numpy(), np.ones((1, TILE_M, 1), dtype=np.float32))
+
+
+@wp.kernel
+def test_tile_reshape_kernel(x: wp.array2d(dtype=float), y: wp.array2d(dtype=float)):
+    a = wp.tile_load(x, shape=(TILE_M, TILE_N), offset=(0, 0))
+    b = wp.tile_reshape(a, shape=(wp.static(TILE_M * TILE_N), 1))
+    c = wp.tile_reshape(b, shape=(-1, 1))
+
+    wp.tile_store(y, c, offset=(0, 0))
+
+
+def test_tile_reshape(test, device):
+    x = wp.ones((TILE_M, TILE_N), dtype=float, device=device, requires_grad=True)
+    y = wp.zeros((TILE_M * TILE_N, 1), dtype=float, device=device, requires_grad=True)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch_tiled(test_tile_reshape_kernel, dim=1, inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.ones((TILE_M * TILE_N, 1), dtype=np.float32))
+    assert_np_equal(x.grad.numpy(), np.ones((TILE_M, TILE_N), dtype=np.float32))
+
+
+@wp.kernel
+def test_tile_astype_kernel(x: wp.array2d(dtype=Any), y: wp.array2d(dtype=wp.float32)):
+    a = wp.tile_load(x, shape=(TILE_M, TILE_N))
+    b = wp.tile_astype(a, dtype=wp.float32)
+    wp.tile_store(y, b)
+
+
+def test_tile_astype(test, device):
+    x_np = np.arange(TILE_M * TILE_N, dtype=np.int32).reshape((TILE_M, TILE_N))
+    x = wp.array(x_np, dtype=wp.int32, device=device)
+    y = wp.zeros((TILE_M, TILE_N), dtype=wp.float32, device=device)
+
+    wp.launch_tiled(test_tile_astype_kernel, dim=1, inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device)
+
+    assert_np_equal(y.numpy(), x_np.astype(np.float32))
+
+    x_np = np.arange(TILE_M * TILE_N, dtype=np.float64).reshape((TILE_M, TILE_N))
+    x = wp.array(x_np, dtype=wp.float64, requires_grad=True, device=device)
+    y = wp.zeros((TILE_M, TILE_N), dtype=wp.float32, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch_tiled(test_tile_astype_kernel, dim=1, inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device)
+
+    y.grad = wp.ones_like(y)
+
+    tape.backward()
+
+    assert_np_equal(y.numpy(), x_np.astype(np.float32))
+    assert_np_equal(x.grad.numpy(), np.ones_like(x_np))
+
+
+@wp.func
+def test_tile_func_return_func(tile: Any):
+    return tile
+
+
+@wp.kernel
+def test_tile_func_return_kernel(x: wp.array2d(dtype=wp.float32), y: wp.array2d(dtype=wp.float32)):
+    a = wp.tile_load(x, shape=(TILE_M, 1))
+    b = wp.tile_broadcast(a, shape=(TILE_M, TILE_K))
+    c = test_tile_func_return_func(b)
+    wp.tile_store(y, c)
+
+
+def test_tile_func_return(test, device):
+    x = wp.ones(shape=(TILE_M, 1), dtype=wp.float32, requires_grad=True, device=device)
+    y = wp.zeros(shape=(TILE_M, TILE_K), dtype=wp.float32, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch_tiled(
+            test_tile_func_return_kernel, dim=[1, 1], inputs=[x], outputs=[y], block_dim=TILE_DIM, device=device
+        )
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.ones((TILE_M, TILE_K), dtype=np.float32))
+    assert_np_equal(x.grad.numpy(), np.ones((TILE_M, 1), dtype=np.float32) * TILE_K)
+
+
 @wp.kernel
 def tile_len_kernel(
     a: wp.array(dtype=float, ndim=2),
@@ -771,6 +1077,111 @@ def test_tile_print(test, device):
     wp.synchronize()
 
 
+@wp.kernel
+def test_tile_add_inplace_kernel(
+    input_a: wp.array2d(dtype=float),
+    input_b: wp.array2d(dtype=float),
+    output_reg: wp.array2d(dtype=float),
+    output_shared: wp.array2d(dtype=float),
+):
+    i, j = wp.tid()
+
+    a_reg = wp.tile_load(input_a, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="register")
+    b_reg = wp.tile_load(input_b, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="register")
+    a_shared = wp.tile_load(input_a, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="shared")
+    b_shared = wp.tile_load(input_b, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="shared")
+
+    a_reg += b_reg
+    a_reg += b_shared
+    a_shared += b_reg
+    a_shared += b_shared
+
+    wp.tile_store(output_reg, a_reg, offset=(i * TILE_M, j * TILE_N))
+    wp.tile_store(output_shared, a_shared, offset=(i * TILE_M, j * TILE_N))
+
+
+@wp.kernel
+def test_tile_sub_inplace_kernel(
+    input_a: wp.array2d(dtype=float),
+    input_b: wp.array2d(dtype=float),
+    output_reg: wp.array2d(dtype=float),
+    output_shared: wp.array2d(dtype=float),
+):
+    i, j = wp.tid()
+
+    a_reg = wp.tile_load(input_a, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="register")
+    b_reg = wp.tile_load(input_b, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="register")
+    a_shared = wp.tile_load(input_a, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="shared")
+    b_shared = wp.tile_load(input_b, shape=(TILE_M, TILE_N), offset=(i * TILE_M, j * TILE_N), storage="shared")
+
+    a_reg -= b_reg
+    a_reg -= b_shared
+    a_shared -= b_reg
+    a_shared -= b_shared
+
+    wp.tile_store(output_reg, a_reg, offset=(i * TILE_M, j * TILE_N))
+    wp.tile_store(output_shared, a_shared, offset=(i * TILE_M, j * TILE_N))
+
+
+def test_tile_inplace(test, device):
+    M = TILE_M * 2
+    N = TILE_N * 2
+
+    a = wp.zeros((M, N), requires_grad=True, device=device)
+    b = wp.ones_like(a, requires_grad=True, device=device)
+    c = wp.zeros_like(a, requires_grad=True, device=device)
+    d = wp.zeros_like(a, requires_grad=True, device=device)
+
+    with wp.Tape() as tape:
+        wp.launch_tiled(
+            test_tile_add_inplace_kernel,
+            dim=[int(M / TILE_M), int(N / TILE_N)],
+            inputs=[a, b, c, d],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    assert_np_equal(a.numpy(), np.zeros((M, N)))
+    assert_np_equal(b.numpy(), np.ones((M, N)))
+    assert_np_equal(c.numpy(), 2.0 * np.ones((M, N)))
+    assert_np_equal(d.numpy(), 2.0 * np.ones((M, N)))
+
+    c.grad = wp.ones_like(c, device=device)
+    d.grad = wp.ones_like(d, device=device)
+    tape.backward()
+
+    assert_np_equal(a.grad.numpy(), 2.0 * np.ones((M, N)))
+    assert_np_equal(b.grad.numpy(), 4.0 * np.ones((M, N)))
+
+    tape.zero()
+
+    a.zero_()
+    b.fill_(1.0)
+    c.zero_()
+    d.zero_()
+
+    with wp.Tape() as tape:
+        wp.launch_tiled(
+            test_tile_sub_inplace_kernel,
+            dim=[int(M / TILE_M), int(N / TILE_N)],
+            inputs=[a, b, c, d],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    assert_np_equal(a.numpy(), np.zeros((M, N)))
+    assert_np_equal(b.numpy(), np.ones((M, N)))
+    assert_np_equal(c.numpy(), -2.0 * np.ones((M, N)))
+    assert_np_equal(d.numpy(), -2.0 * np.ones((M, N)))
+
+    c.grad = wp.ones_like(c, device=device)
+    d.grad = wp.ones_like(d, device=device)
+    tape.backward()
+
+    assert_np_equal(a.grad.numpy(), 2.0 * np.ones((M, N)))
+    assert_np_equal(b.grad.numpy(), -4.0 * np.ones((M, N)))
+
+
 devices = get_test_devices()
 
 
@@ -789,6 +1200,8 @@ add_function_test(TestTile, "test_tile_gemm_fp64", test_tile_gemm(wp.float64), d
 add_function_test(TestTile, "test_tile_transpose", test_tile_transpose, devices=devices)
 add_function_test(TestTile, "test_tile_transpose_matmul", test_tile_transpose_matmul, devices=devices)
 add_function_test(TestTile, "test_tile_operators", test_tile_operators, devices=devices)
+add_function_test(TestTile, "test_tile_tile", test_tile_tile, devices=get_cuda_test_devices())
+add_function_test(TestTile, "test_tile_untile", test_tile_untile, devices=devices)
 add_function_test(TestTile, "test_tile_sum", test_tile_sum, devices=devices, check_output=False)
 add_function_test(TestTile, "test_tile_sum_launch", test_tile_sum_launch, devices=devices)
 add_function_test(TestTile, "test_tile_extract", test_tile_extract, devices=devices)
@@ -799,8 +1212,14 @@ add_function_test(TestTile, "test_tile_broadcast_add_2d", test_tile_broadcast_ad
 add_function_test(TestTile, "test_tile_broadcast_add_3d", test_tile_broadcast_add_3d, devices=devices)
 add_function_test(TestTile, "test_tile_broadcast_add_4d", test_tile_broadcast_add_4d, devices=devices)
 add_function_test(TestTile, "test_tile_broadcast_grad", test_tile_broadcast_grad, devices=devices)
+add_function_test(TestTile, "test_tile_squeeze", test_tile_squeeze, devices=devices)
+add_function_test(TestTile, "test_tile_reshape", test_tile_reshape, devices=devices)
 add_function_test(TestTile, "test_tile_len", test_tile_len, devices=devices)
 add_function_test(TestTile, "test_tile_print", test_tile_print, devices=devices, check_output=False)
+add_function_test(TestTile, "test_tile_inplace", test_tile_inplace, devices=devices)
+add_function_test(TestTile, "test_tile_astype", test_tile_astype, devices=devices)
+add_function_test(TestTile, "test_tile_func_return", test_tile_func_return, devices=devices)
+
 
 if __name__ == "__main__":
     wp.clear_kernel_cache()