Rework HAL CTS to use the C API #3934

benvanik · 2020-11-20T03:48:51Z

It currently is using C++ types directly, which reduces its usefulness with respect to testing the HAL and documenting its usage. It's also blocking porting portions of the HAL to C which is required to allow external user-provided HAL backends.

This isn't urgent but if left too long the CTS will likely need to be deleted until one using the C API can be written (it's not providing enough value in its current form relative to unblocking out-of-tree HAL drivers). A CTS against the C API would be extremely valuable for developers of out-of-tree drivers (as well as for ensuring good coverage of the C API for us), so it's definitely good to get a plan in place to fix it up.

benvanik · 2020-11-22T04:52:37Z

As we are likely to be making more changes to the HAL during the Q1 having a CTS (or using it as an excuse to flesh one out) would be useful to ensuring that we are working correctly across metal and vulkan (and webgpu and such).

Issue #3934 is tracking this.

ScottTodd · 2021-01-13T17:53:18Z

It would also be nice to rework the gtest setup so that out of tree drivers can run against the same set of tests, or the list of drivers to test against could be controlled more easily. The current

INSTANTIATE_TEST_SUITE_P(
    AllDrivers, AllocatorTest,
    ::testing::ValuesIn(testing::EnumerateAvailableDrivers()),
    GenerateTestName());

style setup is pretty constrained in how it runs and the output isn't as useful as it could be when different drivers are mixed together test case by test case.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Removing unused HAL files/functions/types. Removing last use of HAL C++ from the HAL VM module. Swapping many C++ HAL types/enums for C ones. Removing iree/base/bitfield.h. The formatting functionality will reappear in iree_string_builder. Removing iree/hal/cts/ for now until it can be ported to C. Issue iree-org#3934 is tracking this. Moving HAL C++ types to iree/hal/cc/. Removing C++ command buffer validation as it'll be moved into C. Porting the HAL C++ API to C. Unfortunately the string handling code is still very dependent on abseil, but everything else is now just plain C. Fixing quoting on cmake variables that *must* be quoted. (iree-org#4404) Removing unused iree/hal/testing/ mock types. Adding a new arena type with a shared thread-safe block pool. This will be used to provide storage for command buffers and other builder types (submissions/etc) in the local HAL backend. Fixing tracing macro errors with the latest VC++. Got a new error emitted: `error C4576: a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntax` Unfortunately C99 supports paren-init and not non-paren-init and C++ supports the opposite so this error is dumb. Oh well. Porting a majority of Vulkan/Metal to C. Removing Metal backend until it can be replaced. This commit can be used to revive the configuration for the driver and view a partially-ported version of it. The actual implementation will need to be rebased onto the C type system and written in Objective C. Note that the compiler backend has not been removed - we can use that for verifying metal shader output. Issue iree-org#4370 tracks reviving this. Removing iree/base/time.h now that no C++ remains using it. Windows ptr atomics fix (needs test coverage, obviously). Ensure wait handles are zero-initialized. Removing unused status results. Adding support for cleaning up tasks when retired/discarded. Making call and dispatch task closures independent types. Adding batch task submission enqueuing. Fixing executables with multiple pipelines in Vulkan. Disabling executable linking until it can be fixed iree-org#4536. Revive the HAL CTS using the C API. (iree-org#4498) Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes iree-org#3934, part of iree-org#4369. Allowing iree_task_barrier_t to be initialized in two steps. This is required when incrementally building the barrier: you want to insert the barrier into the task DAG prior to knowing what the dependent tasks are. Improving `iree_task_post_batch_select_worker`. With the way command buffers are issued today it tended to bias toward only dispatching on the current worker and - even if worked around - would bias towards idle workers even after the batch is queuing work on them. This more evenly distributes work when many tasks are dispatched within the same batch. Working around crap _mktemp behavior. It only generates 26 unique names per thread o_o Adding a new iree/hal/local/ path as the local CPU driver. This subsumes the existing iree/hal/host/ and will allow us to share the same device for all local execution - including mixing and matching executables in different forms (AOT, vmla, JIT, etc) within the same device. It also adds the (initial, unoptimized) task-system based command buffer and semaphore implementations. Porting both the dylib and vmla drivers to use iree/hal/local. Both entire drivers are now just loader implementations and then shims that match the existing registration. In the future there will just be --driver=local and dylib/vmla will just be compiler registrations like HAL driver registrations currently are. Removing iree/hal/host/ now that it is unused. Moving VMLA VM module to iree/modules/vmla/. Fixing temp file path cleanup. Can't wait to remove this code. Making IREE_EXPECT_OK/IREE_ASSERT_OK print better messages. Removing unused line to remove warning. Tweaking iree_atomic_* macros to better check types and adding a test. Fixing a type typo in the python bindings. Fixing third_party/half bazel-to-cmake mapping. Fixing a race where threads start prior to their handle being assigned. Treating push constants offset/length as bytes in the API. IR still needs a fixup, but is fine for now. Making iree_status_t a non-void type to get better warnings on misuse. Force mapping on hal.allocator.wrap.byte_buffer (today). Cleanup of compile-time detection of sanitizers to enable TSAN/MSAN. Also adds --config=tsan and --config=msan to iree.bazelrc. Making threads join on shutdown (for sanity). Doesn't seem to change anything but there's not really a reason to detach and I'd rather never have to think about this code again. Fixing push constants not being included in executable layout caching. This would cause all kind of badness when multiple executables shared the same hal.interface bindings but had differing push constants. Adding iree_task_pool_t tests. Adding iree_task_queue_t tests. Adding iree_task_scope_t tests. Adding iree_task_list_discard test. Adding iree_task_topology_t tests and simplifying allocation. The max size on the stack is 3KB, which is fine for something transient like this that is only used for construction of the executor. Removing HAL backend implementations of fill/read/write/copy. They were all implemented using mapping anyway and this way we only have that code once and make the restriction that buffers be mappable part of their usage uniformly. Better error status when drivers are not registered. Split iree/tools/test/simple.mlir as it was a PITA all as one. Fix Vulkan GUI sample after HAL C changes. (iree-org#4603) More fixes after HAL C changes. (iree-org#4604) * Use `IREE_RESTRICT` for older MSVC versions. * Update iree-run-module-vulkan-gui-main.cc * Include <atomic> for std::atomic use in GetTempFile. Tested CMake build of `all` (vmla+vulkan only) on my Windows machine, with samples and Vulkan enabled. Properly handling nop tasks in the task system. Adding tests for the various iree_task_t types. Temporarily disabling failing task system wait test. This case is not yet exercised from our code but I'll be fixing it ASAP anyway.

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Removing unused HAL files/functions/types. Removing last use of HAL C++ from the HAL VM module. Swapping many C++ HAL types/enums for C ones. Removing iree/base/bitfield.h. The formatting functionality will reappear in iree_string_builder. Removing iree/hal/cts/ for now until it can be ported to C. Issue iree-org#3934 is tracking this. Moving HAL C++ types to iree/hal/cc/. Removing C++ command buffer validation as it'll be moved into C. Porting the HAL C++ API to C. Unfortunately the string handling code is still very dependent on abseil, but everything else is now just plain C. Fixing quoting on cmake variables that *must* be quoted. (iree-org#4404) Removing unused iree/hal/testing/ mock types. Adding a new arena type with a shared thread-safe block pool. This will be used to provide storage for command buffers and other builder types (submissions/etc) in the local HAL backend. Fixing tracing macro errors with the latest VC++. Got a new error emitted: `error C4576: a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntax` Unfortunately C99 supports paren-init and not non-paren-init and C++ supports the opposite so this error is dumb. Oh well. Porting a majority of Vulkan/Metal to C. Removing Metal backend until it can be replaced. This commit can be used to revive the configuration for the driver and view a partially-ported version of it. The actual implementation will need to be rebased onto the C type system and written in Objective C. Note that the compiler backend has not been removed - we can use that for verifying metal shader output. Issue iree-org#4370 tracks reviving this. Removing iree/base/time.h now that no C++ remains using it. Windows ptr atomics fix (needs test coverage, obviously). Ensure wait handles are zero-initialized. Removing unused status results. Adding support for cleaning up tasks when retired/discarded. Making call and dispatch task closures independent types. Adding batch task submission enqueuing. Fixing executables with multiple pipelines in Vulkan. Disabling executable linking until it can be fixed iree-org#4536. Revive the HAL CTS using the C API. (iree-org#4498) Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes iree-org#3934, part of iree-org#4369. Allowing iree_task_barrier_t to be initialized in two steps. This is required when incrementally building the barrier: you want to insert the barrier into the task DAG prior to knowing what the dependent tasks are. Improving `iree_task_post_batch_select_worker`. With the way command buffers are issued today it tended to bias toward only dispatching on the current worker and - even if worked around - would bias towards idle workers even after the batch is queuing work on them. This more evenly distributes work when many tasks are dispatched within the same batch. Working around crap _mktemp behavior. It only generates 26 unique names per thread o_o Adding a new iree/hal/local/ path as the local CPU driver. This subsumes the existing iree/hal/host/ and will allow us to share the same device for all local execution - including mixing and matching executables in different forms (AOT, vmla, JIT, etc) within the same device. It also adds the (initial, unoptimized) task-system based command buffer and semaphore implementations. Porting both the dylib and vmla drivers to use iree/hal/local. Both entire drivers are now just loader implementations and then shims that match the existing registration. In the future there will just be --driver=local and dylib/vmla will just be compiler registrations like HAL driver registrations currently are. Removing iree/hal/host/ now that it is unused. Moving VMLA VM module to iree/modules/vmla/. Fixing temp file path cleanup. Can't wait to remove this code. Making IREE_EXPECT_OK/IREE_ASSERT_OK print better messages. Removing unused line to remove warning. Tweaking iree_atomic_* macros to better check types and adding a test. Fixing a type typo in the python bindings. Fixing third_party/half bazel-to-cmake mapping. Fixing a race where threads start prior to their handle being assigned. Treating push constants offset/length as bytes in the API. IR still needs a fixup, but is fine for now. Making iree_status_t a non-void type to get better warnings on misuse. Force mapping on hal.allocator.wrap.byte_buffer (today). Cleanup of compile-time detection of sanitizers to enable TSAN/MSAN. Also adds --config=tsan and --config=msan to iree.bazelrc. Making threads join on shutdown (for sanity). Doesn't seem to change anything but there's not really a reason to detach and I'd rather never have to think about this code again. Fixing push constants not being included in executable layout caching. This would cause all kind of badness when multiple executables shared the same hal.interface bindings but had differing push constants. Adding iree_task_pool_t tests. Adding iree_task_queue_t tests. Adding iree_task_scope_t tests. Adding iree_task_list_discard test. Adding iree_task_topology_t tests and simplifying allocation. The max size on the stack is 3KB, which is fine for something transient like this that is only used for construction of the executor. Removing HAL backend implementations of fill/read/write/copy. They were all implemented using mapping anyway and this way we only have that code once and make the restriction that buffers be mappable part of their usage uniformly. Better error status when drivers are not registered. Split iree/tools/test/simple.mlir as it was a PITA all as one. Fix Vulkan GUI sample after HAL C changes. (iree-org#4603) More fixes after HAL C changes. (iree-org#4604) * Use `IREE_RESTRICT` for older MSVC versions. * Update iree-run-module-vulkan-gui-main.cc * Include <atomic> for std::atomic use in GetTempFile. Tested CMake build of `all` (vmla+vulkan only) on my Windows machine, with samples and Vulkan enabled. Properly handling nop tasks in the task system. Adding tests for the various iree_task_t types. Temporarily disabling failing task system wait test. This case is not yet exercised from our code but I'll be fixing it ASAP anyway. Fixing off-by-one in minilcg128.

Removing unused HAL files/functions/types. Removing last use of HAL C++ from the HAL VM module. Swapping many C++ HAL types/enums for C ones. Removing iree/base/bitfield.h. The formatting functionality will reappear in iree_string_builder. Removing iree/hal/cts/ for now until it can be ported to C. Issue iree-org#3934 is tracking this. Moving HAL C++ types to iree/hal/cc/. Removing C++ command buffer validation as it'll be moved into C. Porting the HAL C++ API to C. Unfortunately the string handling code is still very dependent on abseil, but everything else is now just plain C. Fixing quoting on cmake variables that *must* be quoted. (iree-org#4404) Removing unused iree/hal/testing/ mock types. Adding a new arena type with a shared thread-safe block pool. This will be used to provide storage for command buffers and other builder types (submissions/etc) in the local HAL backend. Fixing tracing macro errors with the latest VC++. Got a new error emitted: `error C4576: a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntax` Unfortunately C99 supports paren-init and not non-paren-init and C++ supports the opposite so this error is dumb. Oh well. Porting a majority of Vulkan/Metal to C. Removing Metal backend until it can be replaced. This commit can be used to revive the configuration for the driver and view a partially-ported version of it. The actual implementation will need to be rebased onto the C type system and written in Objective C. Note that the compiler backend has not been removed - we can use that for verifying metal shader output. Issue iree-org#4370 tracks reviving this. Removing iree/base/time.h now that no C++ remains using it. Windows ptr atomics fix (needs test coverage, obviously). Ensure wait handles are zero-initialized. Removing unused status results. Adding support for cleaning up tasks when retired/discarded. Making call and dispatch task closures independent types. Adding batch task submission enqueuing. Fixing executables with multiple pipelines in Vulkan. Disabling executable linking until it can be fixed iree-org#4536. Revive the HAL CTS using the C API. (iree-org#4498) Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes iree-org#3934, part of iree-org#4369. Allowing iree_task_barrier_t to be initialized in two steps. This is required when incrementally building the barrier: you want to insert the barrier into the task DAG prior to knowing what the dependent tasks are. Improving `iree_task_post_batch_select_worker`. With the way command buffers are issued today it tended to bias toward only dispatching on the current worker and - even if worked around - would bias towards idle workers even after the batch is queuing work on them. This more evenly distributes work when many tasks are dispatched within the same batch. Working around crap _mktemp behavior. It only generates 26 unique names per thread o_o Adding a new iree/hal/local/ path as the local CPU driver. This subsumes the existing iree/hal/host/ and will allow us to share the same device for all local execution - including mixing and matching executables in different forms (AOT, vmla, JIT, etc) within the same device. It also adds the (initial, unoptimized) task-system based command buffer and semaphore implementations. Porting both the dylib and vmla drivers to use iree/hal/local. Both entire drivers are now just loader implementations and then shims that match the existing registration. In the future there will just be --driver=local and dylib/vmla will just be compiler registrations like HAL driver registrations currently are. Removing iree/hal/host/ now that it is unused. Moving VMLA VM module to iree/modules/vmla/. Fixing temp file path cleanup. Can't wait to remove this code. Making IREE_EXPECT_OK/IREE_ASSERT_OK print better messages. Removing unused line to remove warning. Tweaking iree_atomic_* macros to better check types and adding a test. Fixing a type typo in the python bindings. Fixing third_party/half bazel-to-cmake mapping. Fixing a race where threads start prior to their handle being assigned. Treating push constants offset/length as bytes in the API. IR still needs a fixup, but is fine for now. Making iree_status_t a non-void type to get better warnings on misuse. Force mapping on hal.allocator.wrap.byte_buffer (today). Cleanup of compile-time detection of sanitizers to enable TSAN/MSAN. Also adds --config=tsan and --config=msan to iree.bazelrc. Making threads join on shutdown (for sanity). Doesn't seem to change anything but there's not really a reason to detach and I'd rather never have to think about this code again. Fixing push constants not being included in executable layout caching. This would cause all kind of badness when multiple executables shared the same hal.interface bindings but had differing push constants. Adding iree_task_pool_t tests. Adding iree_task_queue_t tests. Adding iree_task_scope_t tests. Adding iree_task_list_discard test. Adding iree_task_topology_t tests and simplifying allocation. The max size on the stack is 3KB, which is fine for something transient like this that is only used for construction of the executor. Removing HAL backend implementations of fill/read/write/copy. They were all implemented using mapping anyway and this way we only have that code once and make the restriction that buffers be mappable part of their usage uniformly. Better error status when drivers are not registered. Split iree/tools/test/simple.mlir as it was a PITA all as one. Fix Vulkan GUI sample after HAL C changes. (iree-org#4603) More fixes after HAL C changes. (iree-org#4604) * Use `IREE_RESTRICT` for older MSVC versions. * Update iree-run-module-vulkan-gui-main.cc * Include <atomic> for std::atomic use in GetTempFile. Tested CMake build of `all` (vmla+vulkan only) on my Windows machine, with samples and Vulkan enabled. Properly handling nop tasks in the task system. Adding tests for the various iree_task_t types. Temporarily disabling failing task system wait test. This case is not yet exercised from our code but I'll be fixing it ASAP anyway. Fixing off-by-one in minilcg128. Fixing scope name length truncation. Fixing ubsan-flagged potential undefined shift right behavior. Android build fixes for HAL rewrite branch. (iree-org#4622) Unfortunately not caught by any of our open source builds. --- ``` typedef size_t iree_host_size_t; typedef uint64_t iree_device_size_t; ``` `iree_device_size_t` should use `% " PRIu64 "` instead of `%zu` --- `ppoll` is available since Android API 21 (or 28 for `ppoll64`? seeing conflicting docs) --- `iree_hal_buffer_calculate_range` outputs into a `iree_device_size_t`, which isn't directly compatible with ``` typedef struct { uint8_t* data; iree_host_size_t data_length; } iree_byte_span_t; ```

Issue #3934 is tracking this.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

Removing unused HAL files/functions/types. Removing last use of HAL C++ from the HAL VM module. Swapping many C++ HAL types/enums for C ones. Removing iree/base/bitfield.h. The formatting functionality will reappear in iree_string_builder. Removing iree/hal/cts/ for now until it can be ported to C. Issue iree-org#3934 is tracking this. Moving HAL C++ types to iree/hal/cc/. Removing C++ command buffer validation as it'll be moved into C. Porting the HAL C++ API to C. Unfortunately the string handling code is still very dependent on abseil, but everything else is now just plain C. Fixing quoting on cmake variables that *must* be quoted. (iree-org#4404) Removing unused iree/hal/testing/ mock types. Adding a new arena type with a shared thread-safe block pool. This will be used to provide storage for command buffers and other builder types (submissions/etc) in the local HAL backend. Fixing tracing macro errors with the latest VC++. Got a new error emitted: `error C4576: a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntax` Unfortunately C99 supports paren-init and not non-paren-init and C++ supports the opposite so this error is dumb. Oh well. Porting a majority of Vulkan/Metal to C. Removing Metal backend until it can be replaced. This commit can be used to revive the configuration for the driver and view a partially-ported version of it. The actual implementation will need to be rebased onto the C type system and written in Objective C. Note that the compiler backend has not been removed - we can use that for verifying metal shader output. Issue iree-org#4370 tracks reviving this. Removing iree/base/time.h now that no C++ remains using it. Windows ptr atomics fix (needs test coverage, obviously). Ensure wait handles are zero-initialized. Removing unused status results. Adding support for cleaning up tasks when retired/discarded. Making call and dispatch task closures independent types. Adding batch task submission enqueuing. Fixing executables with multiple pipelines in Vulkan. Disabling executable linking until it can be fixed iree-org#4536. Revive the HAL CTS using the C API. (iree-org#4498) Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes iree-org#3934, part of iree-org#4369. Allowing iree_task_barrier_t to be initialized in two steps. This is required when incrementally building the barrier: you want to insert the barrier into the task DAG prior to knowing what the dependent tasks are. Improving `iree_task_post_batch_select_worker`. With the way command buffers are issued today it tended to bias toward only dispatching on the current worker and - even if worked around - would bias towards idle workers even after the batch is queuing work on them. This more evenly distributes work when many tasks are dispatched within the same batch. Working around crap _mktemp behavior. It only generates 26 unique names per thread o_o Adding a new iree/hal/local/ path as the local CPU driver. This subsumes the existing iree/hal/host/ and will allow us to share the same device for all local execution - including mixing and matching executables in different forms (AOT, vmla, JIT, etc) within the same device. It also adds the (initial, unoptimized) task-system based command buffer and semaphore implementations. Porting both the dylib and vmla drivers to use iree/hal/local. Both entire drivers are now just loader implementations and then shims that match the existing registration. In the future there will just be --driver=local and dylib/vmla will just be compiler registrations like HAL driver registrations currently are. Removing iree/hal/host/ now that it is unused. Moving VMLA VM module to iree/modules/vmla/. Fixing temp file path cleanup. Can't wait to remove this code. Making IREE_EXPECT_OK/IREE_ASSERT_OK print better messages. Removing unused line to remove warning. Tweaking iree_atomic_* macros to better check types and adding a test. Fixing a type typo in the python bindings. Fixing third_party/half bazel-to-cmake mapping. Fixing a race where threads start prior to their handle being assigned. Treating push constants offset/length as bytes in the API. IR still needs a fixup, but is fine for now. Making iree_status_t a non-void type to get better warnings on misuse. Force mapping on hal.allocator.wrap.byte_buffer (today). Cleanup of compile-time detection of sanitizers to enable TSAN/MSAN. Also adds --config=tsan and --config=msan to iree.bazelrc. Making threads join on shutdown (for sanity). Doesn't seem to change anything but there's not really a reason to detach and I'd rather never have to think about this code again. Fixing push constants not being included in executable layout caching. This would cause all kind of badness when multiple executables shared the same hal.interface bindings but had differing push constants. Adding iree_task_pool_t tests. Adding iree_task_queue_t tests. Adding iree_task_scope_t tests. Adding iree_task_list_discard test. Adding iree_task_topology_t tests and simplifying allocation. The max size on the stack is 3KB, which is fine for something transient like this that is only used for construction of the executor. Removing HAL backend implementations of fill/read/write/copy. They were all implemented using mapping anyway and this way we only have that code once and make the restriction that buffers be mappable part of their usage uniformly. Better error status when drivers are not registered. Split iree/tools/test/simple.mlir as it was a PITA all as one. Fix Vulkan GUI sample after HAL C changes. (iree-org#4603) More fixes after HAL C changes. (iree-org#4604) * Use `IREE_RESTRICT` for older MSVC versions. * Update iree-run-module-vulkan-gui-main.cc * Include <atomic> for std::atomic use in GetTempFile. Tested CMake build of `all` (vmla+vulkan only) on my Windows machine, with samples and Vulkan enabled. Properly handling nop tasks in the task system. Adding tests for the various iree_task_t types. Temporarily disabling failing task system wait test. This case is not yet exercised from our code but I'll be fixing it ASAP anyway. Fixing off-by-one in minilcg128. Fixing scope name length truncation. Fixing ubsan-flagged potential undefined shift right behavior. Android build fixes for HAL rewrite branch. (iree-org#4622) Unfortunately not caught by any of our open source builds. --- ``` typedef size_t iree_host_size_t; typedef uint64_t iree_device_size_t; ``` `iree_device_size_t` should use `% " PRIu64 "` instead of `%zu` --- `ppoll` is available since Android API 21 (or 28 for `ppoll64`? seeing conflicting docs) --- `iree_hal_buffer_calculate_range` outputs into a `iree_device_size_t`, which isn't directly compatible with ``` typedef struct { uint8_t* data; iree_host_size_t data_length; } iree_byte_span_t; ``` Fixing ubsan discovery in topology tests on large machines. Possibly: hard to test without 128 cores. I did try turning the topology mask into a uint8_t on my 64 core machine and it passed --config=ubsan. Add missing device release in CTS. Disabling TaskWaitTest IssueSignaled for now. The wait functionality isn't done yet. Fallback to 1 thread when cpuinfo_get_cores_count() == 0.

Adding back a Conformance Test Suite of tests for HAL drivers. These tests were forked from the prior [iree/hal/cts/ tests](https://github.com/google/iree/tree/6f1afa5ebf82d31dafc0c3e835c8baf3eea2ceda/iree/hal/cts), rebased on the new HAL C APIs, then expanded upon. Tests pass with no validation layer warnings on the new Vulkan HAL on Windows with an NVIDIA GPU using latest drivers. Fixes #3934, part of #4369.

benvanik added runtime Relating to the IREE runtime library hal/vulkan Runtime Vulkan GPU HAL backend labels Nov 20, 2020

benvanik added this to the 2020Q4 Core milestone Nov 20, 2020

benvanik assigned antiagainst and ScottTodd Nov 20, 2020

benvanik removed this from the 2020Q4 Core milestone Nov 22, 2020

benvanik added a commit that referenced this issue Dec 24, 2020

Removing iree/hal/cts/ for now until it can be ported to C.

b5790ec

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Dec 30, 2020

Removing iree/hal/cts/ for now until it can be ported to C.

8b2989f

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 4, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

82697cd

Issue #3934 is tracking this.

benvanik mentioned this issue Jan 4, 2021

Rewrite the HAL in C to enable external HAL drivers #4369

Closed

25 tasks

benvanik added a commit that referenced this issue Jan 5, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

83882aa

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 5, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

1d82705

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 5, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

de1a4a1

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 5, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

7744902

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 5, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

7194531

Issue #3934 is tracking this.

ScottTodd mentioned this issue Jan 13, 2021

Revive the HAL CTS using the C API. #4498

Merged

ScottTodd closed this as completed Jan 20, 2021

benvanik added a commit that referenced this issue Jan 21, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

20ef5c9

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 22, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

29fc8af

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 22, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

9d9a4fd

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 26, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

2a9de05

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 27, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

9ee485e

Issue #3934 is tracking this.

benvanik added a commit that referenced this issue Jan 28, 2021

Removing iree/hal/cts/ for now until it can be ported to C.

075db0a

Issue #3934 is tracking this.

ScottTodd mentioned this issue Jan 29, 2021

Merge main -> google #4660

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Rework HAL CTS to use the C API #3934

Rework HAL CTS to use the C API #3934

benvanik commented Nov 20, 2020 •

edited

Loading

benvanik commented Nov 22, 2020

ScottTodd commented Jan 13, 2021

Rework HAL CTS to use the C API #3934

Rework HAL CTS to use the C API #3934

Comments

benvanik commented Nov 20, 2020 • edited Loading

benvanik commented Nov 22, 2020

ScottTodd commented Jan 13, 2021

benvanik commented Nov 20, 2020 •

edited

Loading