Merge main -> google #5520

GMNGeoffrey · 2021-04-19T19:38:31Z

3c275fc Merge pull request Cleaning up HAL timeouts and tunneling iree_hal_device_submit_and_wait through the stack. #5501 from google/benvanik-hal-submit-and-wait
4d25359 Improving timing docs.
92a60a2 Fixing use of memcpy that should have been a memmove. This misuse was causing ..
81a988a Cache the submit-and-wait semaphore and use the new submit_and_wait.
4aac7b4 Plumbing for iree_hal_task_queue_submit_and_wait.
aaf92fa Adding an iree_timeout_t type to make timeouts cleaner. This avoids the duplic..
c155e7b Adding a iree_hal_device_submit_and_wait_with_deadline API.
74be346 Merge pull request Small tweaks to reduce the total size of iree_hal_dispatch_cmd_t. #5513 from google/benvanik-shrink-cmd
131f2e2 Create ops with static shapes in PadTensorOpConversion if they're static (Create ops with static shapes in PadTensorOpConversion if they're static #5516)
840f4c3 Add support for converting rank-reducing subtensor ops to Flow dialect (Add support for converting rank-reducing subtensor ops to Flow dialect #5515)
03f455b Fix rank-mismatch issue in flow.tensor.reshape folders. (Fix rank-mismatch issue in flow.tensor.reshape folders. #5475)
1bd938c Delete duplicated tests in iree/test/e2e/ (Delete duplicated tests in iree/test/e2e/ #5446)
5816c57 Removing a bunch of pointers from iree_hal_cmd_dispatch_t. They are easy to re..
4e0ff3c Removing dispatch_task pointer from shared shard state. It now lives on the sh..
b53052a Removing redundant workgroup_count[3] on dispatch shard state.
292d4d7 Combining allocations of HAL heap buffer's handle and storage. (Combining allocations of HAL heap buffer's handle and storage. #5512)
ad8afdd Merge pull request Small improvements to enable buffer aliasing/importing. #5511 from google/benvanik-buffer-improvements
7afbd4b [spirv] Migrate matmul pipeline test to linalg on tensors ([spirv] Migrate matmul pipeline test to linalg on tensors #5381)
d1b5956 Delete LoadStoreVectorizationPass which is unused in IREE (Delete LoadStoreVectorizationPass which is unused in IREE #5514)
da61e20 Adding the IREE_VM_BYTE_BUFFER_ORIGIN_* marker. This enables us to know the so..
d057b26 Adding IREE_HAL_BUFFER_COMPATIBILITY_IMPORTABLE. This allows allocators to dif..
064a416 Mention supported pip versions in getting_started_python. (Mention supported pip versions in getting_started_python. #5230)

This allows allocators to differentiate between whether it can allocate a new buffer of a specified type and whether it can import/reuse one from the system. Future handle-based import/export APIs will use this.

This enables us to know the source of a byte buffer without trusting modules about its lifetime guarantees. A byte buffer that has a lifetime of the entire module can safely be aliased as by construction it will outlive any use from the module while allocated or external buffers may not have the same guarantees.

) We have a proper flow to vectorize element wise ops after tiling, and it is not used in old pipeline. We can delete the pass and the option. The registration of `createVectorizeLinalgConvPass` was missing. The PR also adds it to init_conversions.h.

) This is a step towards iree-org#5155.

…ents Small improvements to enable buffer aliasing/importing.

…org#5512) Halves the allocation count for buffers, keeps the handle and storage memory close by to hopefully get some caching benefits on initial use, and reduces the noise in allocation traces.

It now lives on the shards themselves as their lifetime is much shorter than the dispatch (which may be recorded into a command buffer much earlier).

They are easy to recompute and this helps keep the command buffer size smaller.

Also delete unnecessary flags in iree/test/e2e and iree/compiler/Dialect/HAL/Target/LLVM/test/smoketest.mlir Part of iree-org#5442

If the ranks mismatch, we can `return false` ealier. It would cause crash when the rank of lhs is greater than the rnak of rhs.

iree-org#5515)

…tic (iree-org#5516) Currently, the mhlo.pad will be lowered to linalg.pad_tensor and then lowered to `linalg.init_tensor + linalg.fill + subtensor_insert`. The init_tensor op will produce a dynamic shape even if the shape is static. This leads a `tensor.cast` op added and relies on further patterns to fix them. This is not needed to static shape and will hit some issues related to iree-org#5385.

Small tweaks to reduce the total size of `iree_hal_dispatch_cmd_t`.

This avoids the duplication of _with_deadline/_with_timeout throughout the HAL. This change does not try to optimize any of the use-sites. Some of the Vulkan methods are going rel->abs->rel (and were before, but now it's clear and easy to fix).

This misuse was causing a 16 byte copy to take 7ms o_o

…wait Cleaning up HAL timeouts and tunneling iree_hal_device_submit_and_wait through the stack.

* 0adee0f Merge pull request #5520 from GMNGeoffrey:main-to-google * e302cb3 Synchronize submodules with LLVM at llvm/llvm-project@80e166f81abd * 6c94570 Integrate LLVM at llvm/llvm-project@80e166f81abd * ce364ae Integrate LLVM at llvm/llvm-project@517c3aee4de5 * 32792f7 Merge pull request #5500 from rsuderman:main-to-google

* 97a7c8e Synchronize submodules with LLVM at llvm/llvm-project@2f69975683f5 * 3c09f24 Integrate LLVM at llvm/llvm-project@2f69975683f5 * 44d1063 Merge pull request #5551 from GMNGeoffrey:main-to-google * 1243070 Integrate LLVM at llvm/llvm-project@1a3f88658a02 * 361e47b Integrate LLVM at llvm/llvm-project@4bb60c285cb3 * eaeb09d Integrate LLVM at llvm/llvm-project@9430efa18b02 * dc0af0c Synchronize submodules with LLVM at llvm/llvm-project@e0adf7e06a9e * 678b1e1 Integrate LLVM at llvm/llvm-project@e0adf7e06a9e * 98cf087 Integrate LLVM at llvm/llvm-project@01ace074fcb6 * 307cef0 Integrate LLVM at llvm/llvm-project@aa80ea8a617b * 297f730 Integrate LLVM at llvm/llvm-project@fb69b92c7b33 * 5f1f4fe Integrate LLVM at llvm/llvm-project@f549176ad976 * bda62f0 Merge main -> google * 0adee0f Merge pull request #5520 from GMNGeoffrey:main-to-google * e302cb3 Synchronize submodules with LLVM at llvm/llvm-project@80e166f81abd * 6c94570 Integrate LLVM at llvm/llvm-project@80e166f81abd * ce364ae Integrate LLVM at llvm/llvm-project@517c3aee4de5 * 32792f7 Merge pull request #5500 from rsuderman:main-to-google

ScottTodd and others added 22 commits April 17, 2021 10:59

Mention supported pip versions in getting_started_python. (iree-org#5230

064a416

) Fixes iree-org#5092

Adding IREE_HAL_BUFFER_COMPATIBILITY_IMPORTABLE.

d057b26

This allows allocators to differentiate between whether it can allocate a new buffer of a specified type and whether it can import/reuse one from the system. Future handle-based import/export APIs will use this.

[spirv] Migrate matmul pipeline test to linalg on tensors (iree-org#5381

7afbd4b

) This is a step towards iree-org#5155.

Merge pull request iree-org#5511 from google/benvanik-buffer-improvem…

ad8afdd

…ents Small improvements to enable buffer aliasing/importing.

Combining allocations of HAL heap buffer's handle and storage. (iree-…

292d4d7

…org#5512) Halves the allocation count for buffers, keeps the handle and storage memory close by to hopefully get some caching benefits on initial use, and reduces the noise in allocation traces.

Removing redundant workgroup_count[3] on dispatch shard state.

b53052a

Removing dispatch_task pointer from shared shard state.

4e0ff3c

It now lives on the shards themselves as their lifetime is much shorter than the dispatch (which may be recorded into a command buffer much earlier).

Removing a bunch of pointers from iree_hal_cmd_dispatch_t.

5816c57

They are easy to recompute and this helps keep the command buffer size smaller.

Delete duplicated tests in iree/test/e2e/ (iree-org#5446)

1bd938c

Also delete unnecessary flags in iree/test/e2e and iree/compiler/Dialect/HAL/Target/LLVM/test/smoketest.mlir Part of iree-org#5442

Fix rank-mismatch issue in flow.tensor.reshape folders. (iree-org#5475)

03f455b

If the ranks mismatch, we can `return false` ealier. It would cause crash when the rank of lhs is greater than the rnak of rhs.

Add support for converting rank-reducing subtensor ops to Flow dialect (

840f4c3

iree-org#5515)

Merge pull request iree-org#5513 from google/benvanik-shrink-cmd

74be346

Small tweaks to reduce the total size of `iree_hal_dispatch_cmd_t`.

Adding a iree_hal_device_submit_and_wait_with_deadline API.

c155e7b

Plumbing for iree_hal_task_queue_submit_and_wait.

4aac7b4

Cache the submit-and-wait semaphore and use the new submit_and_wait.

81a988a

Fixing use of memcpy that should have been a memmove.

92a60a2

This misuse was causing a 16 byte copy to take 7ms o_o

Improving timing docs.

4d25359

Merge pull request iree-org#5501 from google/benvanik-hal-submit-and-…

3c275fc

…wait Cleaning up HAL timeouts and tunneling iree_hal_device_submit_and_wait through the stack.

google-cla bot added the cla: yes label Apr 19, 2021

copybara-service bot merged commit 0adee0f into iree-org:google Apr 19, 2021

GMNGeoffrey mentioned this pull request Apr 19, 2021

Merge google -> main #5526

Merged

This was referenced Apr 20, 2021

Merge google -> main #5529

Closed

Merge google -> main #5574

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Merge main -> google #5520

Merge main -> google #5520

GMNGeoffrey commented Apr 19, 2021

Merge main -> google #5520

Merge main -> google #5520

Conversation

GMNGeoffrey commented Apr 19, 2021