Skip to content

Tags: Spiritaine/iree

Tags

candidate-20230216.432

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
[spirv] Vectorize integer extend ops in lowering to subgroup_mma (ire…

…e-org#12202)

For integer types, integer extend ops are matched against neighboring
vector.transfer_read/contract ops when lowering to mma ops. This enables
vectorizing the extend ops to cooperative matrix sizes. This also
enables support for cases with mixed signedness.

Depends on https://reviews.llvm.org/D143922

candidate-20230215.431

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
[Codegen][LLVMGPU] Add a transform op for createAsyncGroups (iree-org…

…#12171)

This patch connects the createAsyncGroups utility function in the
transform dialect.

candidate-20230214.430

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Add transform op for pipelining of shared memory copies (iree-org#12191)

Add an op to apply the gpu pipelining to a scf.for op.

candidate-20230213.429

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Add python 3.11 deps for macOS (iree-org#12166)

candidate-20230212.428

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Add "Programming Language :: Python :: 3.11" classifier (iree-org#12163)

Add setup.py classifiers for Python 3.11

candidate-20230211.427

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Supporting CUDA importable buffers in compatibility check. (iree-org#…

…12093)

Folds in the fix from iree-org#12066 in
addition to some nice textual error messages when the buffer
compatibility check fails during allocation or import.

This also fixes the aliasing of IREE_HAL_MEMORY_TYPE_HOST_LOCAL with
IREE_HAL_MEMORY_TYPE_DEVICE_LOCAL that made it impossible to
differentiate the two based solely on HOST_VISIBLE/DEVICE_VISIBLE. I
don't recall why the bits aliased and can't find the reason. It's
possible that since the compiler was never producing the values with the
bit set there was other code
compensating for this as part of the change to use OPTIMAL bits and it
went unnoticed.

In order to avoid breaking the binary format the bit is now added but
existing vmfb files should continue to work as expected.
Whenever a full breaking change is introduced we can clean up the bit
ordering in the bitfield.

Fixes iree-org#12066.

candidate-20230210.426

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Add option to compiler to Preprocessing step. (iree-org#12140)

Add an option to `--compile-to` which allows stopping the compilation after preprocessing.
Usage : `iree-compile --compile-to=preprocessing`.

candidate-20230209.425

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Plumb through tensor.pack e2e execution for llvm-cpu backend. (iree-o…

…rg#11875)

All the tensor.pack ops with static inner_tile_sizes are vectorized, which are all covered by e2e tests.

candidate-20230208.424

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Fix `IREE_FLAGS_ENABLE_FLAG_FILE` file IO support (iree-org#12098)

`IREE_FLAGS_ENABLE_FLAG_FILE` only works when file IO is enabled. For
generic platforms with `IREE_FILE_IO_ENABLE=0` it could cause build
failure. Add additional checks for the default setting.

candidate-20230207.423

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
Fix F32 `mma.sync` lit test (iree-org#12079)

Fixing the `MMASYNC` to `CHECK` in F32 `mma.sync` lit tests. We have
moved `mma.sync` lit test checking linalg-to-nvvm-pipeline in a separate
file. Thus, using the default `CHECK` macro to run the lit tests.

Also, making `mma.sync` verification more verbose. It now checks for
instruction shape, datatype, and layout.