colab examples

Sudi Sabet

unread,

Mar 16, 2020, 11:31:24 PM3/16/20

to iree-discuss

Hi there,

We used to be able to run iree/colab examples, like edge_detection. However now, when it reaches

compiler_module = ireec.tf_load_saved_model(saved_model_path)

we get this error:

The kernel appears to have died. It will restart automatically.

Has anyone else encountered this problem?

Thanks

Sudi

Stella Laurenzo

unread,

Mar 16, 2020, 11:47:13 PM3/16/20

to Sudi Sabet, Geoffrey Martin-Noble, iree-discuss

Hi Sudi - I strongly suspect that you are hitting #1000, which had an attempted fix by me over the weekend and (I believe) a successful fix by gcmn@ 7 hours ago.

+Geoffrey Martin-Noble This is why I've been proposing that we either integrate from upstream to an unstable branch (and make master == stable) or provide an explicit stable branch that folks can rely on. Right now, "master" is basically an integration branch and can be quite unstable at times due to the high frequency of incompatible changes in the dependent projects. As we start to hit our completeness milestones, it becomes a lot more necessary to have a stable tag/branch for non-core devs/users to depend on (even if it lags behind head).

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/fdac8314-3b94-453d-bfba-3e596aa5dbff%40googlegroups.com.

Oleg Rybakov

unread,

Mar 16, 2020, 11:56:32 PM3/16/20

to Stella Laurenzo, Geoffrey Martin-Noble, iree-discuss

+1 Geoffrey we should introduce stable versions/branches.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAGS7HzaGjBpeN9encyP7Ucir3Q08phufd6y7e6bbULiSP%2Bs-ag%40mail.gmail.com.

Geoffrey Martin-Noble

unread,

Mar 17, 2020, 12:57:30 AM3/17/20

to Oleg Rybakov, Stella Laurenzo, iree-discuss

The CI for the tensorflow integration has been failing for about 2 weeks, unfortunately (since https://github.com/google/iree/commit/aed15812a515837f4aa76fa631a9d34f8751d95a). It's now "mostly" fixed but we've grown a new issue due to a missing dependency rewrite coming from upstream. I think that shouldn't affect the colab though. Please let us know if checking out master at HEAD, https://github.com/google/iree/commit/ed928f45c9d198f8e87c27293d269fec8af57b16 (my latest fix), or https://github.com/google/iree/commit/ceb125fa75de99ccbfad4a627a3e609819b4d5e8 (one before the original breakage) fixes the issue.

Regarding a stable tag/branch. Yes I think we've gotten to that point. It hasn't risen to the top of my priority list previously, as I've been focusing primarily on people doing core development, for whom stable == stale :-P Any preferences on tag vs a branch? We've had some discussion back and forth. I'm currently leaning towards branch as it's not like a release version where we're going to have a lot of them and it gives us a bit more flexibility (like making it the default checkout).

Ben Vanik

unread,

Mar 17, 2020, 12:59:32 AM3/17/20

to Geoffrey Martin-Noble, Oleg Rybakov, Stella Laurenzo, iree-discuss

+1 branch, much easier to use

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAKeN2eeqf-PGJH6-dts_STtAJ0MYhpwbyYihTZP%2BB-%2Bp9NQXGA%40mail.gmail.com.

Geoffrey Martin-Noble

unread,

Mar 17, 2020, 1:04:45 AM3/17/20

to Ben Vanik, Oleg Rybakov, Stella Laurenzo, iree-discuss

https://github.com/google/iree/tree/stable

Sudi Sabet

unread,

Mar 17, 2020, 2:57:58 PM3/17/20

to iree-discuss

Thanks for your responses everyone!

I tried the stable branch. Running:

bazel test -k iree/...

These tests failed:

/iree/hal/cts:allocator_test FAILED in 0.1s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/cts/allocator_test/test.log

//iree/hal/cts:command_buffer_test FAILED in 0.1s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/cts/command_buffer_test/test.log

//iree/hal/vulkan:dynamic_symbols_test FAILED in 0.2s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/vulkan/dynamic_symbols_test/test.log

//iree/modules/check/test:failure.mlir.test FAILED in 0.3s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/modules/check/test/failure.mlir.test/test.log

//iree/modules/check/test:success.mlir.test FAILED in 0.4s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/modules/check/test/success.mlir.test/test.log

//iree/samples/simple_embedding:simple_embedding_test FAILED in 0.2s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/samples/simple_embedding/simple_embedding_test/test.log

//iree/tools/test:multiple_args.mlir.test FAILED in 0.3s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/multiple_args.mlir.test/test.log

//iree/tools/test:scalars.mlir.test FAILED in 0.3s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/scalars.mlir.test/test.log

//iree/tools/test:simple.mlir.test FAILED in 0.3s

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/simple.mlir.test/test.log

Will appreciate your support!

Sudi

On Monday, March 16, 2020 at 10:04:45 PM UTC-7, Geoffrey Martin-Noble wrote:

https://github.com/google/iree/tree/stable

On Mon, Mar 16, 2020 at 9:59 PM Ben Vanik <benv...@google.com> wrote:

+1 branch, much easier to use

On Mon, Mar 16, 2020 at 9:57 PM 'Geoffrey Martin-Noble' via iree-discuss <iree-d...@googlegroups.com> wrote:

The CI for the tensorflow integration has been failing for about 2 weeks, unfortunately (since https://github.com/google/iree/commit/aed15812a515837f4aa76fa631a9d34f8751d95a). It's now "mostly" fixed but we've grown a new issue due to a missing dependency rewrite coming from upstream. I think that shouldn't affect the colab though. Please let us know if checking out master at HEAD, https://github.com/google/iree/commit/ed928f45c9d198f8e87c27293d269fec8af57b16 (my latest fix), or https://github.com/google/iree/commit/ceb125fa75de99ccbfad4a627a3e609819b4d5e8 (one before the original breakage) fixes the issue.

Regarding a stable tag/branch. Yes I think we've gotten to that point. It hasn't risen to the top of my priority list previously, as I've been focusing primarily on people doing core development, for whom stable == stale :-P Any preferences on tag vs a branch? We've had some discussion back and forth. I'm currently leaning towards branch as it's not like a release version where we're going to have a lot of them and it gives us a bit more flexibility (like making it the default checkout).

On Mon, Mar 16, 2020 at 8:56 PM Oleg Rybakov <ryb...@google.com> wrote:

+1 Geoffrey we should introduce stable versions/branches.

On Mon, Mar 16, 2020 at 8:47 PM 'Stella Laurenzo' via iree-discuss <iree-d...@googlegroups.com> wrote:

Hi Sudi - I strongly suspect that you are hitting #1000, which had an attempted fix by me over the weekend and (I believe) a successful fix by gcmn@ 7 hours ago.

+Geoffrey Martin-Noble This is why I've been proposing that we either integrate from upstream to an unstable branch (and make master == stable) or provide an explicit stable branch that folks can rely on. Right now, "master" is basically an integration branch and can be quite unstable at times due to the high frequency of incompatible changes in the dependent projects. As we start to hit our completeness milestones, it becomes a lot more necessary to have a stable tag/branch for non-core devs/users to depend on (even if it lags behind head).

On Mon, Mar 16, 2020 at 8:31 PM Sudi Sabet <su...@nod-labs.com> wrote:

Hi there,

We used to be able to run iree/colab examples, like edge_detection. However now, when it reaches

compiler_module = ireec.tf_load_saved_model(saved_model_path)

we get this error:

The kernel appears to have died. It will restart automatically.

Has anyone else encountered this problem?

Thanks
Sudi

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/fdac8314-3b94-453d-bfba-3e596aa5dbff%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAGS7HzaGjBpeN9encyP7Ucir3Q08phufd6y7e6bbULiSP%2Bs-ag%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

Sudi Sabet

unread,

Mar 17, 2020, 3:06:46 PM3/17/20

to iree-discuss

running edge_detection.ipynb, the result is:

Screenshot from 2020-03-17 12-04-07.png

On Monday, March 16, 2020 at 10:04:45 PM UTC-7, Geoffrey Martin-Noble wrote:

https://github.com/google/iree/tree/stable

On Mon, Mar 16, 2020 at 9:59 PM Ben Vanik <benv...@google.com> wrote:

+1 branch, much easier to use

On Mon, Mar 16, 2020 at 9:57 PM 'Geoffrey Martin-Noble' via iree-discuss <iree-d...@googlegroups.com> wrote:

The CI for the tensorflow integration has been failing for about 2 weeks, unfortunately (since https://github.com/google/iree/commit/aed15812a515837f4aa76fa631a9d34f8751d95a). It's now "mostly" fixed but we've grown a new issue due to a missing dependency rewrite coming from upstream. I think that shouldn't affect the colab though. Please let us know if checking out master at HEAD, https://github.com/google/iree/commit/ed928f45c9d198f8e87c27293d269fec8af57b16 (my latest fix), or https://github.com/google/iree/commit/ceb125fa75de99ccbfad4a627a3e609819b4d5e8 (one before the original breakage) fixes the issue.

Regarding a stable tag/branch. Yes I think we've gotten to that point. It hasn't risen to the top of my priority list previously, as I've been focusing primarily on people doing core development, for whom stable == stale :-P Any preferences on tag vs a branch? We've had some discussion back and forth. I'm currently leaning towards branch as it's not like a release version where we're going to have a lot of them and it gives us a bit more flexibility (like making it the default checkout).

On Mon, Mar 16, 2020 at 8:56 PM Oleg Rybakov <ryb...@google.com> wrote:

+1 Geoffrey we should introduce stable versions/branches.

On Mon, Mar 16, 2020 at 8:47 PM 'Stella Laurenzo' via iree-discuss <iree-d...@googlegroups.com> wrote:

Hi Sudi - I strongly suspect that you are hitting #1000, which had an attempted fix by me over the weekend and (I believe) a successful fix by gcmn@ 7 hours ago.

+Geoffrey Martin-Noble This is why I've been proposing that we either integrate from upstream to an unstable branch (and make master == stable) or provide an explicit stable branch that folks can rely on. Right now, "master" is basically an integration branch and can be quite unstable at times due to the high frequency of incompatible changes in the dependent projects. As we start to hit our completeness milestones, it becomes a lot more necessary to have a stable tag/branch for non-core devs/users to depend on (even if it lags behind head).

On Mon, Mar 16, 2020 at 8:31 PM Sudi Sabet <su...@nod-labs.com> wrote:

Hi there,

We used to be able to run iree/colab examples, like edge_detection. However now, when it reaches

compiler_module = ireec.tf_load_saved_model(saved_model_path)

we get this error:

The kernel appears to have died. It will restart automatically.

Has anyone else encountered this problem?

Thanks
Sudi

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/fdac8314-3b94-453d-bfba-3e596aa5dbff%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAGS7HzaGjBpeN9encyP7Ucir3Q08phufd6y7e6bbULiSP%2Bs-ag%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-d...@googlegroups.com.

Geoffrey Martin-Noble

unread,

Mar 17, 2020, 3:07:38 PM3/17/20

to Sudi Sabet, iree-discuss

Can you attach the logs? Just based on which tests are failing, I'd guess vulkan isn't working

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/9e9a6b7d-718b-46f3-b442-f79432047099%40googlegroups.com.

Stella Laurenzo

unread,

Mar 17, 2020, 3:28:32 PM3/17/20

to Sudi Sabet, iree-discuss

This does seem like a new failure mode. Would you mind including one of the log files? The one for the last test is likely the simplest and indicative.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/9e9a6b7d-718b-46f3-b442-f79432047099%40googlegroups.com.

Sudi Sabet

unread,

Mar 17, 2020, 3:34:56 PM3/17/20

to iree-discuss

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/cts/allocator_test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/hal/cts:allocator_test

-----------------------------------------------------------------------------

[==========] Running 4 tests from 1 test suite.

[----------] Global test environment set-up.

[----------] 4 tests from AllDrivers/AllocatorTest

[ RUN ] AllDrivers/AllocatorTest.CanAllocate/vmla

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vmla'...

I ./iree/hal/cts/cts_test_base.h:44] Creating default device...

I ./iree/hal/cts/cts_test_base.h:46] Created device 'vmla'

[ OK ] AllDrivers/AllocatorTest.CanAllocate/vmla (0 ms)

[ RUN ] AllDrivers/AllocatorTest.CanAllocate/vulkan

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vulkan'...

./iree/hal/cts/cts_test_base.h:43: Failure

Value of: _status_or_value0.status()

Expected: is OK

Actual: INTERNAL: VK_ERROR_INITIALIZATION_FAILED (of type iree::Status), which is not OK

INTERNAL: VK_ERROR_INITIALIZATION_FAILED

[ FAILED ] AllDrivers/AllocatorTest.CanAllocate/vulkan, where GetParam() = "vulkan" (2 ms)

[ RUN ] AllDrivers/AllocatorTest.Allocate/vmla

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vmla'...

I ./iree/hal/cts/cts_test_base.h:44] Creating default device...

I ./iree/hal/cts/cts_test_base.h:46] Created device 'vmla'

[ OK ] AllDrivers/AllocatorTest.Allocate/vmla (0 ms)

[ RUN ] AllDrivers/AllocatorTest.Allocate/vulkan

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vulkan'...

./iree/hal/cts/cts_test_base.h:43: Failure

Value of: _status_or_value0.status()

Expected: is OK

Actual: INTERNAL: VK_ERROR_INITIALIZATION_FAILED (of type iree::Status), which is not OK

INTERNAL: VK_ERROR_INITIALIZATION_FAILED

[ FAILED ] AllDrivers/AllocatorTest.Allocate/vulkan, where GetParam() = "vulkan" (1 ms)

[----------] 4 tests from AllDrivers/AllocatorTest (3 ms total)

[----------] Global test environment tear-down

[==========] 4 tests from 1 test suite ran. (3 ms total)

[ PASSED ] 2 tests.

[ FAILED ] 2 tests, listed below:

[ FAILED ] AllDrivers/AllocatorTest.CanAllocate/vulkan, where GetParam() = "vulkan"

[ FAILED ] AllDrivers/AllocatorTest.Allocate/vulkan, where GetParam() = "vulkan"

2 FAILED TESTS

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/cts/command_buffer_test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/hal/cts:command_buffer_test

-----------------------------------------------------------------------------

[==========] Running 2 tests from 1 test suite.

[----------] Global test environment set-up.

[----------] 2 tests from AllDrivers/CommandBufferTest

[ RUN ] AllDrivers/CommandBufferTest.CreateCommandBuffer/vmla

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vmla'...

I ./iree/hal/cts/cts_test_base.h:44] Creating default device...

I ./iree/hal/cts/cts_test_base.h:46] Created device 'vmla'

[ OK ] AllDrivers/CommandBufferTest.CreateCommandBuffer/vmla (0 ms)

[ RUN ] AllDrivers/CommandBufferTest.CreateCommandBuffer/vulkan

I ./iree/hal/cts/cts_test_base.h:35] Creating driver 'vulkan'...

./iree/hal/cts/cts_test_base.h:43: Failure

Value of: _status_or_value0.status()

Expected: is OK

Actual: INTERNAL: VK_ERROR_INITIALIZATION_FAILED (of type iree::Status), which is not OK

INTERNAL: VK_ERROR_INITIALIZATION_FAILED

[ FAILED ] AllDrivers/CommandBufferTest.CreateCommandBuffer/vulkan, where GetParam() = "vulkan" (2 ms)

[----------] 2 tests from AllDrivers/CommandBufferTest (2 ms total)

[----------] Global test environment tear-down

[==========] 2 tests from 1 test suite ran. (2 ms total)

[ PASSED ] 1 test.

[ FAILED ] 1 test, listed below:

[ FAILED ] AllDrivers/CommandBufferTest.CreateCommandBuffer/vulkan, where GetParam() = "vulkan"

1 FAILED TEST

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/hal/vulkan/dynamic_symbols_test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/hal/vulkan:dynamic_symbols_test

-----------------------------------------------------------------------------

[==========] Running 1 test from 1 test suite.

[----------] Global test environment set-up.

[----------] 1 test from DynamicSymbolsTest

[ RUN ] DynamicSymbolsTest.CreateFromSystemLoader

F iree/hal/vulkan/dynamic_symbols_test.cc:62] Check failed: ::iree::OkStatus() == (::iree::hal::vulkan::VkResultToStatus(syms->vkCreateInstance(&create_info, nullptr, &instance))) (OK vs. INTERNAL: VK_ERROR_INITIALIZATION_FAILED)

external/bazel_tools/tools/test/test-setup.sh: line 310: 14 Aborted (core dumped) "${TEST_PATH}" "$@" 2>&1

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/modules/check/test/failure.mlir.test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/modules/check/test:failure.mlir.test

-----------------------------------------------------------------------------

run_lit.sh: iree/modules/check/test/failure.mlir

PWD=/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7930/execroot/iree_core/bazel-out/k8-opt/bin/iree/modules/check/test/failure.mlir.test.runfiles/iree_core

RUNLINE: // RUN: check-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module %s | iree-check-module --expect_failure | IreeFileCheck %s

RUNNING TEST: check-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module iree/modules/check/test/failure.mlir | iree-check-module --expect_failure | IreeFileCheck iree/modules/check/test/failure.mlir

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vmla'...

--- COMPLETE ---

RUNLINE: // RUN: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module %s | iree-check-module --driver=vulkan --expect_failure | IreeFileCheck %s)

RUNNING TEST: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/modules/check/test/failure.mlir | iree-check-module --driver=vulkan --expect_failure | IreeFileCheck iree/modules/check/test/failure.mlir)

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

F iree/base/internal/statusor.cc:32] Attempting to fetch value instead of handling error INTERNAL: Creating driver 'vulkan'

FileCheck error: '<stdin>' is empty.

FileCheck command line: FileCheck --enable-var-scope --dump-input=fail iree/modules/check/test/failure.mlir

!!! ERROR EVALUATING: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/modules/check/test/failure.mlir | iree-check-module --driver=vulkan --expect_failure | IreeFileCheck iree/modules/check/test/failure.mlir)

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/modules/check/test/success.mlir.test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/modules/check/test:success.mlir.test

-----------------------------------------------------------------------------

run_lit.sh: iree/modules/check/test/success.mlir

PWD=/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7928/execroot/iree_core/bazel-out/k8-opt/bin/iree/modules/check/test/success.mlir.test.runfiles/iree_core

RUNLINE: // RUN: check-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module %s | iree-check-module --driver=vmla

RUNNING TEST: check-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module iree/modules/check/test/success.mlir | iree-check-module --driver=vmla

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vmla'...

[==========] Running 8 tests from 1 test suite.

[----------] Global test environment set-up.

[----------] 8 tests from module

[ RUN ] module.expect_true

[ OK ] module.expect_true (25 ms)

[ RUN ] module.expect_false

[ OK ] module.expect_false (16 ms)

[ RUN ] module.expect_all_true

[ OK ] module.expect_all_true (17 ms)

[ RUN ] module.expect_all_true_tensor

[ OK ] module.expect_all_true_tensor (16 ms)

[ RUN ] module.expect_eq

[ OK ] module.expect_eq (14 ms)

[ RUN ] module.expect_almost_eq

[ OK ] module.expect_almost_eq (15 ms)

[ RUN ] module.abs

[ OK ] module.abs (44 ms)

[ RUN ] module.floats

[ OK ] module.floats (46 ms)

[----------] 8 tests from module (193 ms total)

[----------] Global test environment tear-down

[==========] 8 tests from 1 test suite ran. (193 ms total)

[ PASSED ] 8 tests.

IMPORTANT NOTICE - DO NOT IGNORE:

This test program did NOT call testing::InitGoogleTest() before calling RUN_ALL_TESTS(). This is INVALID. Soon Google Test will start to enforce the valid usage. Please fix it ASAP, or IT WILL START TO FAIL.

--- COMPLETE ---

RUNLINE: // RUN: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module %s | iree-check-module --driver=vulkan)

RUNNING TEST: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/modules/check/test/success.mlir | iree-check-module --driver=vulkan)

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

F iree/base/internal/statusor.cc:32] Attempting to fetch value instead of handling error INTERNAL: Creating driver 'vulkan'

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7928/execroot/iree_core/bazel-out/k8-opt/bin/iree/modules/check/test/success.mlir.test.runfiles/iree_core/iree/modules/check/test/success.mlir.test: line 86: 100 Done check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/modules/check/test/success.mlir

101 Aborted (core dumped) | iree-check-module --driver=vulkan

!!! ERROR EVALUATING: [[ $IREE_VULKAN_DISABLE == 1 ]] || (check-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/modules/check/test/success.mlir | iree-check-module --driver=vulkan)

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/samples/simple_embedding/simple_embedding_test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/samples/simple_embedding:simple_embedding_test

-----------------------------------------------------------------------------

[==========] Running 2 tests from 1 test suite.

[----------] Global test environment set-up.

[----------] 2 tests from AllDrivers/SimpleEmbeddingTest

[ RUN ] AllDrivers/SimpleEmbeddingTest.RunOnce/vmla

I iree/samples/simple_embedding/simple_embedding_test.cc:76] Creating driver 'vmla'...

I iree/samples/simple_embedding/simple_embedding_test.cc:90] Loading simple_module_test.mlir...

I iree/samples/simple_embedding/simple_embedding_test.cc:105] Module loaded and context is ready for use

I iree/samples/simple_embedding/simple_embedding_test.cc:120] Creating I/O buffers...

I iree/samples/simple_embedding/simple_embedding_test.cc:160] Calling module.simple_mul...

I iree/samples/simple_embedding/simple_embedding_test.cc:167] Retreiving results...

I iree/samples/simple_embedding/simple_embedding_test.cc:173] Reading back results...

I iree/samples/simple_embedding/simple_embedding_test.cc:182] Results match!

[ OK ] AllDrivers/SimpleEmbeddingTest.RunOnce/vmla (58 ms)

[ RUN ] AllDrivers/SimpleEmbeddingTest.RunOnce/vulkan

I iree/samples/simple_embedding/simple_embedding_test.cc:76] Creating driver 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

iree/samples/simple_embedding/simple_embedding_test.cc:80: Failure

Expected equality of these values:

IREE_STATUS_OK

Which is: 0

((iree_status_code_t)((iree_hal_driver_registry_create_driver( iree_string_view_t{driver_name.data(), driver_name.size()}, { 0, iree_allocator_system_allocate, iree_allocator_system_free }, &driver)) & IREE_STATUS_CODE_MASK))

Which is: 13

[ FAILED ] AllDrivers/SimpleEmbeddingTest.RunOnce/vulkan, where GetParam() = vulkan (2 ms)

[----------] 2 tests from AllDrivers/SimpleEmbeddingTest (60 ms total)

[----------] Global test environment tear-down

[==========] 2 tests from 1 test suite ran. (60 ms total)

[ PASSED ] 1 test.

[ FAILED ] 1 test, listed below:

[ FAILED ] AllDrivers/SimpleEmbeddingTest.RunOnce/vulkan, where GetParam() = vulkan

1 FAILED TEST

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/multiple_args.mlir.test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/tools/test:multiple_args.mlir.test

-----------------------------------------------------------------------------

run_lit.sh: iree/tools/test/multiple_args.mlir

PWD=/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7929/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/multiple_args.mlir.test.runfiles/iree_core

RUNLINE: // RUN: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module %s | iree-run-module --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]") | IreeFileCheck %s

RUNNING TEST: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module iree/tools/test/multiple_args.mlir | iree-run-module --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]") | IreeFileCheck iree/tools/test/multiple_args.mlir

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vmla'...

I iree/tools/vm_util.cc:178] result[0]: Buffer<sint32[2]>

I iree/tools/vm_util.cc:178] result[1]: Buffer<sint32[2]>

--- COMPLETE ---

RUNLINE: // RUN: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module %s | iree-run-module --driver=vulkan --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]") | IreeFileCheck %s)

RUNNING TEST: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/multiple_args.mlir | iree-run-module --driver=vulkan --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]") | IreeFileCheck iree/tools/test/multiple_args.mlir)

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

F iree/tools/run_module_main.cc:144] Check failed: ::iree::OkStatus() == (Run()) (OK vs. INTERNAL: Creating driver 'vulkan')

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7929/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/multiple_args.mlir.test.runfiles/iree_core/iree/tools/test/multiple_args.mlir.test: line 86: 43 Done iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/multiple_args.mlir

44 Aborted (core dumped) | iree-run-module --driver=vulkan --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]"

FileCheck error: '<stdin>' is empty.

FileCheck command line: FileCheck --enable-var-scope --dump-input=fail iree/tools/test/multiple_args.mlir

!!! ERROR EVALUATING: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/multiple_args.mlir | iree-run-module --driver=vulkan --entry_function=multi_input --inputs="2xi32=[1 2], 2xi32=[3 4]") | IreeFileCheck iree/tools/test/multiple_args.mlir)

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/scalars.mlir.test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/tools/test:scalars.mlir.test

-----------------------------------------------------------------------------

run_lit.sh: iree/tools/test/scalars.mlir

PWD=/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7924/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/scalars.mlir.test.runfiles/iree_core

RUNLINE: // RUN: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module %s | iree-run-module --entry_function=scalar --inputs="i32=42") | IreeFileCheck %s

RUNNING TEST: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module iree/tools/test/scalars.mlir | iree-run-module --entry_function=scalar --inputs="i32=42") | IreeFileCheck iree/tools/test/scalars.mlir

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vmla'...

I iree/tools/vm_util.cc:178] result[0]: sint32

--- COMPLETE ---

RUNLINE: // RUN: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module %s | iree-run-module --driver=vulkan --entry_function=scalar --inputs="i32=42") | IreeFileCheck %s)

RUNNING TEST: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/scalars.mlir | iree-run-module --driver=vulkan --entry_function=scalar --inputs="i32=42") | IreeFileCheck iree/tools/test/scalars.mlir)

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

F iree/tools/run_module_main.cc:144] Check failed: ::iree::OkStatus() == (Run()) (OK vs. INTERNAL: Creating driver 'vulkan')

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7924/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/scalars.mlir.test.runfiles/iree_core/iree/tools/test/scalars.mlir.test: line 86: 43 Done iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/scalars.mlir

44 Aborted (core dumped) | iree-run-module --driver=vulkan --entry_function=scalar --inputs="i32=42"

FileCheck error: '<stdin>' is empty.

FileCheck command line: FileCheck --enable-var-scope --dump-input=fail iree/tools/test/scalars.mlir

!!! ERROR EVALUATING: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/scalars.mlir | iree-run-module --driver=vulkan --entry_function=scalar --inputs="i32=42") | IreeFileCheck iree/tools/test/scalars.mlir)

cat /home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/execroot/iree_core/bazel-out/k8-opt/testlogs/iree/tools/test/simple.mlir.test/test.log

exec ${PAGER:-/usr/bin/less} "$0" || exit 1

Executing tests from //iree/tools/test:simple.mlir.test

-----------------------------------------------------------------------------

run_lit.sh: iree/tools/test/simple.mlir

PWD=/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7931/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/simple.mlir.test.runfiles/iree_core

RUNLINE: // RUN: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module %s | iree-run-module --driver=vmla --entry_function=abs --inputs="i32=-2") | IreeFileCheck %s

RUNNING TEST: (iree-translate --iree-hal-target-backends=vmla -iree-mlir-to-vm-bytecode-module iree/tools/test/simple.mlir | iree-run-module --driver=vmla --entry_function=abs --inputs="i32=-2") | IreeFileCheck iree/tools/test/simple.mlir

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vmla'...

I iree/tools/vm_util.cc:178] result[0]: Buffer<sint32[]>

--- COMPLETE ---

RUNLINE: // RUN: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module %s | iree-run-module --driver=vulkan --entry_function=abs --inputs="i32=-2") | IreeFileCheck %s)

RUNNING TEST: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/simple.mlir | iree-run-module --driver=vulkan --entry_function=abs --inputs="i32=-2") | IreeFileCheck iree/tools/test/simple.mlir)

----------------

I iree/tools/vm_util.cc:266] Creating driver and device for 'vulkan'...

E ./iree/base/api_util.h:30] INTERNAL: VK_ERROR_INITIALIZATION_FAILED

F iree/tools/run_module_main.cc:144] Check failed: ::iree::OkStatus() == (Run()) (OK vs. INTERNAL: Creating driver 'vulkan')

/home/sudi/.cache/bazel/_bazel_sudi/fe8923d58b1cc237c81293bac6bc7c01/sandbox/linux-sandbox/7931/execroot/iree_core/bazel-out/k8-opt/bin/iree/tools/test/simple.mlir.test.runfiles/iree_core/iree/tools/test/simple.mlir.test: line 86: 107 Done iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/simple.mlir

108 Aborted (core dumped) | iree-run-module --driver=vulkan --entry_function=abs --inputs="i32=-2"

FileCheck error: '<stdin>' is empty.

FileCheck command line: FileCheck --enable-var-scope --dump-input=fail iree/tools/test/simple.mlir

!!! ERROR EVALUATING: [[ $IREE_VULKAN_DISABLE == 1 ]] || ((iree-translate --iree-hal-target-backends=vulkan-spirv -iree-mlir-to-vm-bytecode-module iree/tools/test/simple.mlir | iree-run-module --driver=vulkan --entry_function=abs --inputs="i32=-2") | IreeFileCheck iree/tools/test/simple.mlir)

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/9e9a6b7d-718b-46f3-b442-f79432047099%40googlegroups.com.

Scott Todd

unread,

Mar 17, 2020, 3:44:16 PM3/17/20

to Sudi Sabet, iree-discuss

Looks like Vulkan isn't getting set up correctly. There are some troubleshooting suggestions at https://github.com/google/iree/blob/master/docs/vulkan_and_spirv.md.

There are also notes in https://github.com/google/iree/blob/master/docs/getting_started_on_linux.md regarding configuring SwiftShader on Linux, if you don't have a hardware GPU / Vulkan driver.

I'm also noticing that the Windows builds for some Python targets (including Colab) are unstable / broken. I'll see what can be done to get those working again.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/61c46105-8491-457c-b671-2deeea7a1b02%40googlegroups.com.

Brehler, Marius

unread,

Mar 17, 2020, 5:16:59 PM3/17/20

to iree-discuss

Yes, this are exactly the tests failing when no GPU is available. This also the case using one of my machines for the CMake build. We should also add some notes to the docs regarding the CMake build.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAF-j_jfgoE19%2Bcj6SVT1e%3D_aq12%2B9ONM8rU6FzJK8dmApGUuzg%40mail.gmail.com.

Sudi Sabet

unread,

Mar 17, 2020, 7:41:55 PM3/17/20

to iree-discuss

I cloned the stable branch on a machine where all the tests used to pass. Any idea why GPU will become unavailable for this new build?

Thanks,

Sudi

To unsubscribe from this group and stop receiving emails from it, send an email to iree-...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/61c46105-8491-457c-b671-2deeea7a1b02%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-...@googlegroups.com.

Scott Todd

unread,

Mar 17, 2020, 8:29:42 PM3/17/20

to Sudi Sabet, iree-discuss

Vulkan/GPU support should be there, provided your system is configured in a way that the Vulkan loader is happy with. You could try updating your GPU drivers. There's also the option of installing the Vulkan SDK (https://www.lunarg.com/vulkan-sdk/), but I don't think that should be necessary here.

Running with the `VK_LOADER_DEBUG=all` environment variable will print more logs from the loader:

$ bazel build //iree/hal/vulkan:dynamic_symbols_test && VK_LOADER_DEBUG=all ./bazel-bin/iree/hal/vulkan/dynamic_symbols_test

...
INFO: Found ICD manifest file /usr/share/vulkan/icd.d/nvidia_icd.json, version "1.0.0"
DEBUG: Searching for ICD drivers named libGLX_nvidia.so.0
...

$ bazel build //iree/hal/cts:allocator_test && VK_LOADER_DEBUG=all ./bazel-bin/iree/hal/cts/allocator_test

I found a few things to tweak with the Windows Python builds to get them working again: https://github.com/google/iree/pull/1115 . Interestingly, the edge detection notebook isn't producing the expected output for me either, though Vulkan is set up and the notebook doesn't crash. I can continue looking into that and might compare against Linux and/or an older build tomorrow.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/4e29e4a5-dfd3-4a31-9aea-1073cec19190%40googlegroups.com.

Sudi Sabet

unread,

Mar 17, 2020, 8:49:20 PM3/17/20

to iree-discuss

Scott,

These both passed.

$ bazel build //iree/hal/vulkan:dynamic_symbols_test && VK_LOADER_DEBUG=all ./bazel-bin/iree/hal/vulkan/dynamic_symbols_test

$ bazel build //iree/hal/cts:allocator_test && VK_LOADER_DEBUG=all ./bazel-bin/iree/hal/cts/allocator_test

The only print out that stood out was this:

DEBUG: Device Extension: VK_NV_ray_tracing (libGLX_nvidia.so.0) version 0.0.3

I ./iree/hal/cts/cts_test_base.h:46] Created device 'GeForce RTX 2080 Ti'

E iree/hal/vulkan/debug_reporter.cc:40] vkCreateBuffer: value of pCreateInfo->usage must not be 0. The Vulkan spec states: usage must not be 0 (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-VkBufferCreateInfo-usage-requiredbitmask)

DEBUG: Unloading layer library libVkLayer_khronos_validation.so

[ OK ] AllDrivers/AllocatorTest.Allocate/vulkan (272 ms)

[----------] 4 tests from AllDrivers/AllocatorTest (628 ms total)

[----------] Global test environment tear-down

[==========] 4 tests from 1 test suite ran. (628 ms total)

[ PASSED ] 4 tests.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/4e29e4a5-dfd3-4a31-9aea-1073cec19190%40googlegroups.com.

Scott Todd

unread,

Mar 18, 2020, 12:19:10 PM3/18/20

to Sudi Sabet, iree-discuss

Ah, our Vulkan implementation of 'conv', which is needed for the edge detection notebook has also been broken since https://github.com/google/iree/commit/d7ae43966cb6573ee467374cea9076fcdc8675ea, and that commit also disabled its test. That notebook works correctly at https://github.com/google/iree/commit/ceb125fa75de99ccbfad4a627a3e609819b4d5e8 for me, on Linux.

Geoffrey, this is why I'm concerned about updating the 'stable' branch too frequently - our automated test coverage is too sparse and the project is still changing quickly.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/dcb1239a-719c-4689-99b1-88cf82b66567%40googlegroups.com.

Stella Laurenzo

unread,

Mar 18, 2020, 12:31:43 PM3/18/20

to Scott Todd, Sudi Sabet, iree-discuss

Apologies for that -- the decision to disable the hand coded vulkan conv kennel was deliberate last Friday to resolve a pending set of changes to enable the real, general version. The hand coded original version required hard coding across the compiler which was making it hard do it right for real. While we would never do such a thing in a supported project, this does reflect the early stage disclaimer for the project right now -- and I apologise for the confusion.

I believe landing the real version for cpu+gpu needed a coordinated set of llvm and iree changes that Hanhan was working on, and he will enable the test as part of them.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAF-j_jeSgAaOjC8eK7zaVoNvexAdfQtoY1Yu8P%3Dz%2B_%3Dyu8_8sQ%40mail.gmail.com.

Stella Laurenzo

unread,

Mar 18, 2020, 12:35:07 PM3/18/20

to Scott Todd, Sudi Sabet, iree-discuss

And yes to your meta point: while I expect that over time we will get to the point where we don't make such changes, for the phase we are in, we should expect that we enter periods where we lock stable for ~days deliberately.

Mahesh Ravishankar

unread,

Mar 18, 2020, 12:37:23 PM3/18/20

to iree-discuss

We are still trying to plumb through the convolution through the codegen paths. We are 90% there, but there is a correctness issue that we are looking into. I cannot gaurantee it, but we are looking to fix this soon (Hanhan and me are both looking into this). At the very least we must be able to run the example on the llvmir backend since the issue with convolution is happening in the SPIR-V codegen path (but we havent tried this yet).

The earlier manual implementation of conv was deprecated. We had to move away from that to make it easier for us to plumb through dynamic shapes in the IREE compilation path.

If the vulkan path takes longer, and the llvmir path is also showing the same issues we can try to land conv on the vmla path to fix the regression.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss+unsubscribe@googlegroups.com.

Stella Laurenzo

unread,

Mar 18, 2020, 12:51:41 PM3/18/20

to Mahesh Ravishankar, iree-discuss

Scott/Sudi - I'm sorry for the lack of coordination on this and apologize for any time wasted. With all of us working from home, I'm really missing the ability to pop up and have a quick sync about things like this :(

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/dcb1239a-719c-4689-99b1-88cf82b66567%40googlegroups.com.

--

You received this message because you are subscribed to the Google Groups "iree-discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/6757767f-778e-4187-b1dc-9977b263c430%40googlegroups.com.

Sudi Sabet

unread,

Mar 18, 2020, 1:22:03 PM3/18/20

to Stella Laurenzo, Mahesh Ravishankar, iree-discuss

Thank you Stella and all others for the feedback!

Can you give me a commit in which I can rely on iree/samples/vulkan/vulkan_inference_gui.cc and colab examples working?

Thanks,

Sudi

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAH8pnHbRQrcKKutMEq67hhMqMTD7SFdmOSwMDT5V9Yxg%2BiMAUA%40mail.gmail.com.

Stella Laurenzo

unread,

Mar 18, 2020, 1:29:53 PM3/18/20

to Sudi Sabet, Stella Laurenzo, Mahesh Ravishankar, iree-discuss

This commit is where the change that invalidate conv and disabled vulkan_conv_test.py took place.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAJuFDWTJYHtiUOm6Xj8A4b0N3eOjjkoprDioXp9FDv9MgfT%2BhA%40mail.gmail.com.

Geoffrey Martin-Noble

unread,

Mar 18, 2020, 4:21:32 PM3/18/20

to Stella Laurenzo, Sudi Sabet, Stella Laurenzo, Mahesh Ravishankar, iree-discuss

Sorry for the confusion Sudi. Please let us know if you get things working. My initial push of the stable branch was actually before that commit (I subsequently bumped it), so I don't think that explains the test failures you were seeing, but maybe they're a separate issue.

Regarding stable branch bumping, I agree that we can't guarantee bumping it particularly frequently, but I'm not sure that just keeping it stale is really a solution in and of itself. In this case lack of automated testing wasn't the issue: the tests were disabled. I think this was a reasonable tradeoff for overall project velocity, but we need to have a better way of tracking such a thing. I don't want to create a super heavy-weight process, but maybe this can be something similar to "launch blocking" bugs for releases. Maybe just notes in our build cop doc? I was around for the disabling conv discussion, but I didn't think about implications for colab notebooks and I didn't consider it when bumping the stable branch. We need something other than state in people's heads for tracking that sort of thing. And if we just bump stable infrequently, I don't think that really helps for things not covered by automated testing, since it's not clear they would be caught in other ways.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/CAGS7HzbYuOKnfjPgtnA_FETwM3RsVVjgxJDc%2B02iixQO103m6A%40mail.gmail.com.

Sudi Sabet

unread,

Mar 19, 2020, 1:30:34 PM3/19/20

to iree-discuss

Hi Geoffrey,

Thanks for the explanation! I still don't have a working colab. Here is what I want to do. I have generated an MLIR from a tensorflow mnist training function. I wanted to use colab source base to see if I can run it through IREE.

I have attached the mlir code. Is there another example I can use? Or again, which commit should I use if the stable branch is not working?

Thanks,

Sudi

...

mark_for_compilation.mlir

Scott Todd

unread,

Mar 19, 2020, 7:18:59 PM3/19/20

to Sudi Sabet, iree-discuss

I just bumped the 'stable' branch after several build fixes went in today. The edge detection Colab notebook using Vulkan now works as expected for me on Windows again.

As far as I'm aware, we don't expect model training to work yet through IREE (I'm skeptical of stuff like "tf.SparseSoftmaxCrossEntropyWithLogits" and "tf.BiasAddGrad" in that IR). The "mnist_tensorflow" notebook shows training ahead of time in TensorFlow, exporting to SavedModel, then loading/compiling/executing the exported model in IREE. You could also give the CPU backend (VMLA / interpreter) a try if you aren't successful in getting Vulkan going within Colab.

--

You received this message because you are subscribed to the Google Groups "iree-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to iree-discuss...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/iree-discuss/566b512f-5473-4469-80b6-d69937b836a6%40googlegroups.com.

Sudi Sabet

unread,

Mar 23, 2020, 12:44:40 PM3/23/20

to iree-discuss

Thanks Scott! That is helpful to know. We will suspend our efforts for now.

Sudi

Reply all

Reply to author

Forward