Skip to content

[AMD][BACKEND][TDM] Handle Negative Offsets as Fully Out-of-Bounds Tiles #10310

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
antiagainst merged 1 commit into from
May 15, 2026
+32 −1
Merged

[AMD][BACKEND][TDM] Handle Negative Offsets as Fully Out-of-Bounds Tiles #10310

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion third_party/amd/lib/TritonAMDGPUToLLVM/TDMUtility.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -788,7 +788,9 @@ void fillTDMDescriptor(RewriterBase &rewriter, Location loc,

// Update tensor shapes based on offset
for (size_t i = 0; i < numDims; ++i) {
tensorShape[i] = b.smax(b.i32_val(0), b.sub(tensorShape[i], offset[i]));
auto diff = b.sub(tensorShape[i], offset[i]);
Value inBounds = b.icmp_ule(diff, tensorShape[i]);
tensorShape[i] = b.select(inBounds, diff, b.i32_val(0));
}

// TDM store does not support padding in general. However, if the padding
Expand Down
29 changes: 29 additions & 0 deletions third_party/amd/python/test/test_gluon_gfx1250.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2944,6 +2944,35 @@ def kernel(a_ptr, b_ptr):
assert torch.equal(a[:, 32:], b[:, 32:]) and not torch.equal(a[:, :32], b[:, :32])


# Check that negative TDM offsets are treated as unsigned so they will mask (zero-fill) out the whole tile.
@pytest.mark.skipif(not is_hip_gfx1250(), reason="Requires GFX1250")
def test_tdm_load_negative_offset():

@gluon.jit
def tdm_load_negative_offset_kernel(a_ptr, b_ptr):
shared_layout: ttgl.constexpr = ttgl.SwizzledSharedLayout(vec=1, per_phase=1, max_phase=1, order=[1, 0])
reg_layout: ttgl.constexpr = ttgl.BlockedLayout([1, 4], [4, 8], [4, 1], [1, 0])

desc = ttgl.amd.gfx1250.tdm.make_tensor_descriptor(base=a_ptr, shape=(16, 64), strides=(64, 1),
block_shape=(16, 64), layout=shared_layout)
smem = ttgl.allocate_shared_memory(desc.dtype, shape=desc.block_shape, layout=desc.layout)

ttgl.amd.gfx1250.tdm.async_load(desc, [-1, -1], smem)
ttgl.amd.gfx1250.tdm.async_wait(0)

b_offs_m = ttgl.arange(0, 16, layout=ttgl.SliceLayout(1, reg_layout))
b_offs_n = ttgl.arange(0, 64, layout=ttgl.SliceLayout(0, reg_layout))
b_ptrs = b_ptr + b_offs_m[:, None] * 64 + b_offs_n[None, :]
tile = smem.load(reg_layout)
ttgl.store(b_ptrs, tile)

a = torch.randint(0x0, 0xFFFF, (16, 64), dtype=torch.uint16)
b_device = torch.randint(0x0, 0xFFFF, (16, 64), dtype=torch.uint16).cuda()
tdm_load_negative_offset_kernel[(1, )](a.cuda(), b_device)

assert torch.all(b_device.cpu() == 0)


@pytest.mark.skipif(not is_hip_gfx1250(), reason="Requires GFX1250")
@pytest.mark.parametrize("XBLOCK", [128])
def test_ws_store_wait_load(XBLOCK):
Expand Down
Loading