test: xfail regression for #1051 (bwds crash on runtime-start diff loop) + docs workaround

docs/src/autodiff/autodiff.rst

@@ -62,6 +62,34 @@ Now we'll attach gradients to the result and set them to 1, then run back propag
 That's the lot! The call to ``bwds`` generates a kernel that calls ``bwds_diff(polynomial)`` in Slang, and automatically
 deals with passing in/out the correct data.
 
+.. note::
+
+    **Loops in differentiable functions should start at a compile-time constant.** A
+    ``[Differentiable]`` function whose body contains a loop whose *start value is a runtime
+    (non-constant) expression* — for example ``for (int dx = -radius; dx <= radius; ++dx)``
+    where ``radius`` is a runtime value — can crash during the backward pass
+    (`slangpy#1051 <https://github.com/shader-slang/slangpy/issues/1051>`_, tracked upstream as
+    `slang#12070 <https://github.com/shader-slang/slang/issues/12070>`_). It is the loop start
+    being *non-constant* that triggers the bug, not its being negative: a constant start such as
+    ``for (int dx = -2; dx <= 2; ++dx)`` differentiates fine. This originates in the Slang
+    compiler's reverse-mode auto-diff, not in SlangPy, which hands the function body unchanged to
+    ``bwd_diff``. The forward pass is unaffected. Until the upstream fix lands, rewrite such loops
+    so the start is a constant (typically zero) and offset the index, which is mathematically
+    identical and differentiates correctly:
+
+    .. code-block:: slang
+
+        // Avoid: runtime (non-constant) loop start crashes bwds()
+        for (int dx = -radius; dx <= radius; ++dx)
+            acc += src[clamp(base + dx, 0, width - 1)];
+
+        // Prefer: constant (zero) start with an offset index
+        for (int i = 0; i < 2 * radius + 1; ++i)
+        {
+            int dx = i - radius;
+            acc += src[clamp(base + dx, 0, width - 1)];
+        }
+
 It is worth noting that SlangPy currently **always accumulates** gradients, so you will need to ensure gradient buffers
 are zeroed. In the demo above, we used ``zero=True`` when creating the tensor to do so.
 

slangpy/tests/slangpy_tests/test_diff_loop_runtime_start.py

@@ -0,0 +1,116 @@
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+"""Regression test for shader-slang/slangpy#1051.
+
+Root cause: shader-slang/slang#12070 (Slang reverse-mode autodiff). Fixed upstream by
+shader-slang/slang#12072.
+
+Backward dispatch (``.bwds()``) of a ``[Differentiable]`` function crashes when the function
+body contains a loop whose *start* value is a **runtime (non-constant) induction start**, e.g.
+``for (int dx = -radius; dx <= radius; ++dx)`` where ``radius`` is a ``no_diff`` runtime param.
+The forward pass is always correct, and a mathematically identical zero-based rewrite
+(``for (int i = 0; i < 2*radius+1; ++i) { int dx = i - radius; ... }``) -- whose start is the
+compile-time constant ``0`` -- produces correct gradients through the same call path.
+
+The trigger is the loop start being non-constant, *not* its being negative: a constant negative
+start such as ``for (int dx = -2; dx <= 2; ++dx)`` differentiates fine. The reporter's
+``-radius`` was incidentally both runtime and negative. The root cause is upstream in the Slang
+compiler's reverse-mode autodiff, not in SlangPy: SlangPy hands the user function verbatim to
+``bwd_diff`` and does no loop/induction rewriting.
+
+This test covers the two *constant-start* forms that must differentiate correctly -- a constant
+negative start (``const_neg_start``) and the zero-based workaround (``zero_start``) -- which pin
+the trigger to runtime-ness and guard the workaround. The crashing runtime-start form itself is
+NOT exercised here: it is a hard native crash (SIGSEGV) that, on the CI GPU runners, lands in an
+unkillable process (SlangPy's Crashpad handler / the GPU-driver fault path) and wedges the whole
+"Unit Tests (Python)" step for hours rather than failing cleanly. The crash is documented in the
+issues above and in ``docs/src/autodiff/autodiff.rst``; a live crash-reproducing test is not
+worth hanging CI. Once the upstream fix (shader-slang/slang#12072) ships in a released wheel,
+``runtime_start`` can be added here as an ordinary passing case with no subprocess scaffolding.
+"""
+
+import numpy as np
+import pytest
+
+import slangpy as spy
+from slangpy import DeviceType
+from slangpy.testing import helpers
+
+# Two [Differentiable] functions with identical math and read footprint, both with a
+# compile-time-constant loop start (so neither triggers #1051):
+#   const_neg_start -- constant negative start (-2): shows negative is not the trigger
+#   zero_start      -- zero-based workaround with an offset index
+SHADER = r"""
+[Differentiable]
+float const_neg_start(no_diff int2 pix, IDiffTensor<float, 2> src,
+                      no_diff int radius, no_diff int width)
+{
+    float acc = 0.0;
+    [MaxIters(17)]
+    for (int dx = -2; dx <= 2; ++dx)
+        acc += src[uint2(clamp(pix.x + dx, 0, width - 1), pix.y)];
+    return acc;
+}
+
+[Differentiable]
+float zero_start(no_diff int2 pix, IDiffTensor<float, 2> src,
+                 no_diff int radius, no_diff int width)
+{
+    float acc = 0.0;
+    [MaxIters(17)]
+    for (int i = 0; i < 2 * radius + 1; ++i)
+    {
+        int dx = i - radius;
+        acc += src[uint2(clamp(pix.x + dx, 0, width - 1), pix.y)];
+    }
+    return acc;
+}
+"""
+
+# Use a non-square shape so the row/column axes cannot be confused by the oracle below.
+H = 6  # rows       (pix.y, array axis 0)
+W = 8  # columns    (pix.x, clamped against `width`, array axis 1)
+R = 2  # radius; chosen so both functions cover the same [-2, 2] window
+
+
+def _expected_grad() -> np.ndarray:
+    """Independent numpy oracle for the input gradient of both loop variants.
+
+    Each output element ``(y, x)`` sums ``src[y, clamp(x + dx, 0, W-1)]`` over ``dx`` in
+    ``[-R, R]``; with the incoming result grad set to 1 everywhere, the gradient at a source
+    element is the number of output elements that read it (a clamped box-filter footprint).
+    """
+    grad = np.zeros((H, W), dtype=np.float64)
+    for y in range(H):
+        for x in range(W):
+            for dx in range(-R, R + 1):
+                grad[y, min(max(x + dx, 0), W - 1)] += 1.0
+    return grad
+
+
+def _bwds_grad(device_type: DeviceType, func_name: str, src_np: np.ndarray) -> np.ndarray:
+    """Run forward + ``.bwds()`` for ``func_name`` and return the input gradient as numpy."""
+    device = helpers.get_device(device_type)
+    function = helpers.create_function_from_module(device, func_name, SHADER)
+
+    src = spy.Tensor.from_numpy(device, src_np).with_grads()
+    out = function(spy.grid((H, W)), src, R, W)
+
+    out = out.with_grads()
+    out.grad.copy_from_numpy(np.ones((H, W), dtype=np.float32))
+    function.bwds(spy.grid((H, W)), src, R, W, _result=out)
+
+    return src.grad.to_numpy()
+
+
+@pytest.mark.parametrize("func_name", ["const_neg_start", "zero_start"])
+@pytest.mark.parametrize("device_type", helpers.DEFAULT_DEVICE_TYPES)
+def test_diff_loop_constant_start(device_type: DeviceType, func_name: str):
+    """Constant loop starts differentiate correctly (#1051 does not fire on a constant start).
+
+    ``const_neg_start`` (constant -2) shows a negative start is fine; ``zero_start`` is the
+    documented workaround. Both isolate the #1051 trigger to a *runtime* (non-constant) start.
+    """
+    if helpers.should_skip_test_for_device(device_type):
+        pytest.skip("device filtered out")
+    grad = _bwds_grad(device_type, func_name, np.ones((H, W), dtype=np.float32))
+    np.testing.assert_allclose(grad, _expected_grad(), rtol=1e-5, atol=1e-5)