Now that I've landed this change to Dawn to disable FXC optimizations:
https://dawn-review.googlesource.com/c/dawn/+/70700,, we can reland this
change. The Tint-into-Dawn roll was failing because FXC would miscompile
certain loops into infinite loops when not unrolled.
Also reland
test/bug/fxc/gradient_in_varying_loop/1112.wgsl.expected.hlsl
Bug: tint:1112
Bug: dawn:1203
Change-Id: I641d68864b833e0fbe3b117d397b89ae96482536
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/71000
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
This reverts commit 11d09f2fe7.
Reason for revert: Failing roll of Tint to Dawn: https://dawn-review.googlesource.com/c/dawn/+/70100
Original change's description:
> HLSL: force FXC to never unroll loops
>
> Emit the "[loop]" attribute on "for" and "while" so that FXC does not
> attempt to unroll them. This is to work around an FXC bug where it fails
> to unroll loops with gradient operations.
>
> FXC ostensibly unrolls such loops because gradient operations require
> uniform control flow, and loops that have varying iterations may
> possibly not be uniform. Tint will eventually validate that control flow
> is indeed uniform, so forcing FXC to avoid unrolling in these cases
> should be fine.
>
> Bug: tint:1112
> Change-Id: I10077f8b62fbbb230a0003f3864c75a8fe0e1d18
> Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69880
> Kokoro: Kokoro <noreply+kokoro@google.com>
> Reviewed-by: Ben Clayton <bclayton@google.com>
> Commit-Queue: Antonio Maiorano <amaiorano@google.com>
# Not skipping CQ checks because original CL landed > 1 day ago.
Bug: tint:1112
Change-Id: I8e8f3c0abfa6e6bc5d0e67af9428a46ef867d5c1
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/70540
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Emit the "[loop]" attribute on "for" and "while" so that FXC does not
attempt to unroll them. This is to work around an FXC bug where it fails
to unroll loops with gradient operations.
FXC ostensibly unrolls such loops because gradient operations require
uniform control flow, and loops that have varying iterations may
possibly not be uniform. Tint will eventually validate that control flow
is indeed uniform, so forcing FXC to avoid unrolling in these cases
should be fine.
Bug: tint:1112
Change-Id: I10077f8b62fbbb230a0003f3864c75a8fe0e1d18
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69880
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Arrays can be extremely large, and having the load and store functions unroll the elements can make the complier explode.
Fixed: chromium:1229233
Change-Id: Ieb5654254e16f5ce724a205d21d954ef9a0cd053
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/58382
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
Auto-Submit: Ben Clayton <bclayton@google.com>
CloneContext::Replace(T* what, T* with) is bug-prone, as complex transforms may want to clone `what` multiple times, or not at all. In both cases, this will likely result in an ICE as either the replacement will be reachable multiple times, or not at all.
The CTS test: webgpu:shader,execution,robust_access:linear_memory:storageClass="storage";storageMode="read_write";access="read";atomic=true;baseType="i32"
Was triggering this brokenness with DecomposeMemoryAccess's use of CloneContext::Replace(T*, T*).
Switch the usage of CloneContext::Replace(T*, T*) to the new function form.
As std::function is copyable, it cannot hold a captured std::unique_ptr.
This prevented the Replace() lambdas from capturing the necessary `BufferAccess` data, as this held a `std::unique_ptr<Offset>`.
To fix this, use a `BlockAllocator` for Offsets, and use raw pointers instead.
Because the function passed to Replace() is called just before the node is cloned, insertion of new functions will occur just before the currently evaluated module-scope entity.
This allows us to remove the "insert_after" arguments to LoadFunc(), StoreFunc(), and AtomicFunc().
We can also kill the icky InsertGlobal() and TypeDeclOf() helpers.
Bug: tint:993
Change-Id: I60972bc13a2fa819a163ee2671f61e82d0e68d2a
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/58222
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
FXC has trouble dealing with these.
This was originally added to handle returning arrays as structures.
HLSL supports typedefs, which is a much simpiler solution, and doesn't upset FXC.
Bug: tint:848
Bug: tint:904
Change-Id: Ie841c9c454461a885a35c41476fd4d05d3f34cbf
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/56774
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
For structures and arrays.
This behaves identically to the per-element zero-initialization, but can be significantly less verbose.
Change-Id: I380ef86f16c2b3f37a9de2820e707f368955b761
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/56764
Auto-Submit: Ben Clayton <bclayton@google.com>
Reviewed-by: David Neto <dneto@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Instead of a ConstantBuffer.
HLSL requires that each structure field in a UBO is 16 byte aligned.
WGSL has much looser constraints with its UBO field alignment rules.
Instead generate an array of uint4 vectors, and index into this, much
like we index into [RW]ByteAddressBuffers for SSBOs.
Extend the DecomposeStorageAccess transform to support uniforms too.
This has been renamed to DecomposeMemoryAccess.
Change-Id: I3868ff80af1ab3b3dddfbf5b969724cb87ef0744
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/55246
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: David Neto <dneto@google.com>
- When storing to sample_mask output, write to the 0th element
- Only make a return struct if it has members
- Adjust type signedness coercion when loading special builtins.
- Adapt tests
- Update expectations for end-to-end tests
- Handle sample_mask with stride
Input variables normally don't have layout. But they can have it
up through SPIR-V 1.4.
Handle this case in the SPIR-V reader, by seeing through the
intermediate alias type created for the strided array type.
Bug: tint:508
Change-Id: I0f19dc1305d3f250dbbc0698a602288c34245274
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/54743
Auto-Submit: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: James Price <jrprice@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Combined with the new PadArrayElements transform, arrays with strides
are now correctly emitted.
Fixed: tint:182
Fixed: tint:895
Change-Id: I26a1be94dee6e4c9d9747c8317a932fc1fb3c810
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/54640
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Antonio Maiorano
Ben Clayton
David Neto