FrozenGene edited a comment on issue #3859: [WIP] [AutoTVM][TOPI] NHWC conv2d
templates for ARM
URL: https://github.com/apache/incubator-tvm/pull/3859#issuecomment-551422580
@jackwish I agree @tqchen's comment. I think we could bring this in, which
doesn't hurt anything.
FrozenGene commented on issue #3859: [WIP] [AutoTVM][TOPI] NHWC conv2d
templates for ARM
URL: https://github.com/apache/incubator-tvm/pull/3859#issuecomment-551422580
I agree @tqchen's comment. I think we could bring this in, which doesn't
hurt anything.
FrozenGene commented on issue #4249: [TOPI][AlterOpLayout][ARM] Enabling NHWC
to NCHW layout transformation.
URL: https://github.com/apache/incubator-tvm/pull/4249#issuecomment-551422285
> Ping for review - @jackwish @merrymercy @FrozenGene @yzhliu
>
> AutoTVM would not work for
anijain2305 commented on issue #4249: [TOPI][AlterOpLayout][ARM] WIP - Enabling
NHWC to NCHW layout transformation.
URL: https://github.com/apache/incubator-tvm/pull/4249#issuecomment-551421458
Ping for review - @jackwish @merrymercy @FrozenGene @yzhliu
AutoTVM would not work for
kimishpatel commented on issue #4243: Fix broken loop partitioning due to
recent changes.
URL: https://github.com/apache/incubator-tvm/pull/4243#issuecomment-551408119
@ZihengJiang
Output without fix:
```
F1107 22:50:11.396204 3352620 unroll_loop.cc:85] Check failed: value >= 0
kimishpatel commented on issue #4243: Fix broken loop partitioning due to
recent changes.
URL: https://github.com/apache/incubator-tvm/pull/4243#issuecomment-551405311
@ZihengJiang, this PR keeps the fixes for tails loop from #3734. The changes
here
vinx13 merged pull request #4276: [TOPI][CUDA] Fix Winograd Kernel Size Support
URL: https://github.com/apache/incubator-tvm/pull/4276
This is an automated message from the Apache Git Service.
To respond to the message,
This is an automated email from the ASF dual-hosted git repository.
wuwei pushed a change to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-tvm.git.
from 5bcd331 [Relay][Frontend][ONNX] Add support for broadcasting to Where
and MatMul (#4267)
add 76b7967
DamonDeng commented on issue #4279: Creating tvm.target twice will create empty
params
URL: https://github.com/apache/incubator-tvm/issues/4279#issuecomment-551385452
One troubleshooting thread created in discuss.tvm.ai
This
FrozenGene opened a new pull request #4281: Support C++ RPC
URL: https://github.com/apache/incubator-tvm/pull/4281
C++ RPC is very useful for embedded devices, which can not have Python
environment easily. And thanks @siju-samuel ' great work, based on
tobegit3hub commented on issue #2062: Check failed: override Global PackedFunc
__tvm_set_device is already registered
URL: https://github.com/apache/incubator-tvm/issues/2062#issuecomment-551382711
Thanks @sunstarchan . We have the similar issue when linking the
`libtvm_runtime` in our
zhiics opened a new pull request #4280: [tvm][runtime] A minimum runtime for
external library
URL: https://github.com/apache/incubator-tvm/pull/4280
A minimum runtime for external codegen as mentioned in #4258
This is an
tqchen commented on a change in pull request #4268: [Codegen][cuda-fp16]
fallback to fp32 simulation when cuda arch < sm53
URL: https://github.com/apache/incubator-tvm/pull/4268#discussion_r343985590
##
File path: src/codegen/codegen_cuda.cc
##
@@ -50,6 +50,9 @@ void
tqchen commented on issue #4279: Creating tvm.target twice will create empty
params
URL: https://github.com/apache/incubator-tvm/issues/4279#issuecomment-551379928
Thanks for reporting this, please open a new troubleshooting thread on
https://discuss.tvm.ai/
tqchen closed issue #4279: Creating tvm.target twice will create empty params
URL: https://github.com/apache/incubator-tvm/issues/4279
This is an automated message from the Apache Git Service.
To respond to the message,
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343982583
##
File path: tutorials/optimize/opt_matmul_auto_tensorcore.py
##
@@ -0,0 +1,317 @@
+# Licensed to the Apache
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343982375
##
File path: tutorials/optimize/opt_matmul_auto_tensorcore.py
##
@@ -0,0 +1,317 @@
+# Licensed to the Apache
jcf94 commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343981732
##
File path: src/api/api_pass.cc
##
@@ -94,6 +94,13 @@ TVM_REGISTER_API("ir_pass.StorageFlatten")
}
});
DamonDeng opened a new issue #4279: Creating tvm.target twice will create empty
params
URL: https://github.com/apache/incubator-tvm/issues/4279
Calling `target = tvm.target.create('llvm')` twice will create empty params.
The following is the code snippet I used, which was copied
yzhliu commented on issue #4268: [Codegen][cuda-fp16] fallback to fp32
simulation when cuda arch < sm53
URL: https://github.com/apache/incubator-tvm/pull/4268#issuecomment-551367782
@vinx13 @tqchen please review again
This
FrozenGene commented on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial
Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-551362798
> @jackwish Yes, sharing code will be very helpful. For now, do you prefer
changing schedule OR Legalize pass?
FrozenGene commented on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial
Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-551359926
> @jackwish I'd be very interested in those results. I got some good results
for NHWC on ARMv7 by porting the
ajtulloch commented on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial
Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-551356298
@jackwish I'd be very interested in those results. I got some good results
for NHWC on ARMv7 by porting the QNNPACK
PeikeLi commented on issue #4278: Fix the problem that android_rpc compilation
failed
URL: https://github.com/apache/incubator-tvm/pull/4278#issuecomment-551354634
@tqchen
Please help to review this fix.
This is an
PeikeLi opened a new pull request #4278: Fix the problem that android_rpc
compilation failed
URL: https://github.com/apache/incubator-tvm/pull/4278
This small PR address the issue of header path in
/apps/android_rpc/app/src/main/jni/tvm_runtime.h .
Description:
Error msg when
anijain2305 commented on issue #4277: [ARM][Topi] Improving Int8 Perf in
Spatial Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-551351937
@jackwish Yes, sharing code will be very helpful. For now, do you prefer
changing schedule OR Legalize pass?
adityaatluri commented on a change in pull request #4234: Auto TensorCore
CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343961695
##
File path: src/pass/tensor_core.cc
##
@@ -0,0 +1,1247 @@
+/*
+ * Licensed to the Apache Software Foundation
jackwish commented on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial
Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-551349270
Ha! This is so interesting - it is very similar to our internal try many
months ago. I think legalization is a good
anijain2305 commented on a change in pull request #4277: [ARM][Topi] Improving
Int8 Perf in Spatial Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#discussion_r343954134
##
File path: topi/python/topi/arm_cpu/conv2d_spatial_pack.py
##
@@ -93,24
anijain2305 commented on a change in pull request #4277: [ARM][Topi] Improving
Int8 Perf in Spatial Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#discussion_r343953530
##
File path: topi/python/topi/arm_cpu/conv2d_spatial_pack.py
##
@@ -93,24
comaniac commented on issue #4276: [TOPI][CUDA] Fix Winograd Kernel Size Support
URL: https://github.com/apache/incubator-tvm/pull/4276#issuecomment-551337905
Thanks for the suggestion. The unit test has been added and I've confirmed
the unit test covers the desire part.
One miner
ajtulloch commented on a change in pull request #4277: [ARM][Topi] Improving
Int8 Perf in Spatial Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#discussion_r343949837
##
File path: topi/python/topi/arm_cpu/conv2d_spatial_pack.py
##
@@ -93,24
anijain2305 opened a new pull request #4277: [ARM][Topi] Supporting Int8 in
Spatial schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277
I am working on improving the performance of Int8 conv on Raspberry Pi 3.
For Conv2D, there is an upcast from int8 to int32 before
weberlo commented on issue #4274: [µTVM] Enable AutoTVM for ARM STM32F746XX
Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#issuecomment-551330769
@tqchen It looks like the CI doesn't allow assembly---namely,
weberlo commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343940866
##
File path: python/tvm/micro/rpc_server.py
##
@@ -0,0 +1,90 @@
+# Licensed to the Apache
ZihengJiang commented on issue #4243: Fix broken loop partitioning due to
recent changes.
URL: https://github.com/apache/incubator-tvm/pull/4243#issuecomment-551326453
Hi @kimishpatel , the https://github.com/apache/incubator-tvm/pull/3734 is
for avoiding generating the tail loop with
weberlo commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343934268
##
File path: src/runtime/micro/device/arm/stm32f746xx/utvm_timer.c
##
@@ -0,0 +1,102 @@
weberlo commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343933459
##
File path: python/tvm/contrib/binutil.py
##
@@ -45,25 +58,18 @@ def
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343927217
##
File path: src/runtime/micro/openocd_low_level_device.cc
##
@@ -201,35 +192,21 @@ class
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343926760
##
File path: python/tvm/micro/rpc_server.py
##
@@ -0,0 +1,90 @@
+# Licensed to the Apache
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343926911
##
File path: src/runtime/micro/device/arm/stm32f746xx/utvm_timer.c
##
@@ -0,0 +1,102 @@
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343909330
##
File path: python/tvm/micro/base.py
##
@@ -14,129 +14,156 @@
# KIND, either express or
tqchen commented on issue #3859: [WIP] [AutoTVM][TOPI] NHWC conv2d templates
for ARM
URL: https://github.com/apache/incubator-tvm/pull/3859#issuecomment-551311734
Given how far we have landed the code, it will still be great to bring it
into the collection. @jackwish if you like, please
vinx13 commented on issue #4276: [TOPI][CUDA] Fix Winograd Kernel Size Support
URL: https://github.com/apache/incubator-tvm/pull/4276#issuecomment-551306320
To test with pre_computed case, you can add a relay unit test that runs a
single conv layer model under `opt_level=3`. You may need
comaniac opened a new pull request #4276: [TOPI][CUDA] Fix Winograd Kernel Size
Support
URL: https://github.com/apache/incubator-tvm/pull/4276
The merged PR #4260 fixes the padding issue when building Winograd conv2d
for CUDA, but we found the kernel size is still a problem.
The
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343908611
##
File path: python/tvm/contrib/binutil.py
##
@@ -45,25 +58,18 @@ def
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343909482
##
File path: python/tvm/micro/device/arm/stm32f746xx.py
##
@@ -0,0 +1,91 @@
+# Licensed to
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343909330
##
File path: python/tvm/micro/base.py
##
@@ -14,129 +14,156 @@
# KIND, either express or
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343908899
##
File path: python/tvm/contrib/binutil.py
##
@@ -125,74 +146,73 @@ def
tqchen commented on a change in pull request #4274: [µTVM] Enable AutoTVM for
ARM STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274#discussion_r343908512
##
File path: python/tvm/contrib/binutil.py
##
@@ -22,6 +22,19 @@
from .._ffi.base import
This is an automated email from the ASF dual-hosted git repository.
tqchen pushed a change to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-tvm.git.
from 14a5a35 [AutoTVM] Add batch_matmul to tunable operations (#4242)
add 5bcd331 [Relay][Frontend][ONNX]
tqchen merged pull request #4267: [Relay][Frontend][ONNX] Add support for
broadcasting to Where and MatMul
URL: https://github.com/apache/incubator-tvm/pull/4267
This is an automated message from the Apache Git Service.
To
tqchen commented on issue #3200: [WIP][ANDROID][APPS] Real-Time Android Camera
Demo Sample App
URL: https://github.com/apache/incubator-tvm/pull/3200#issuecomment-551289334
@eqy can you followup? let us bring this in.
This
tqchen edited a comment on issue #3200: [WIP][ANDROID][APPS] Real-Time Android
Camera Demo Sample App
URL: https://github.com/apache/incubator-tvm/pull/3200#issuecomment-551289334
@eqy @srkreddy1238 can you followup? let us bring this in.
were commented on issue #4275: imp module is deprecated
URL: https://github.com/apache/incubator-tvm/pull/4275#issuecomment-551284051
@junrushao1994 Can you take a look?
This is an automated message from the Apache Git
were opened a new pull request #4275: imp module is deprecated
URL: https://github.com/apache/incubator-tvm/pull/4275
Python deprecated `imp` module, so we can no longer use `imp.load_source` to
import a file by a path.
A workaround is to execute the source so that the defined
weberlo opened a new pull request #4274: [µTVM] Enable AutoTVM for ARM
STM32F746XX Boards
URL: https://github.com/apache/incubator-tvm/pull/4274
This PR adds support for autotuning via MicroTVM. To test this
infrastructure on a physical board, I have added support for ARM STM32F746XX
tqchen commented on issue #4270: [Codgen] Thread variable use before define
URL: https://github.com/apache/incubator-tvm/pull/4270#issuecomment-551236823
Perhaps we can discuss your fast-gemm a bit in the forum to see if we can
find alternatives, if it is indeed a limitation in the
yongwww commented on a change in pull request #4083: [tutorial] Relay pass
infra tutorial
URL: https://github.com/apache/incubator-tvm/pull/4083#discussion_r343832715
##
File path: tutorials/dev/relay_pass_infra.py
##
@@ -0,0 +1,253 @@
+# Licensed to the Apache Software
csarofeen edited a comment on issue #4270: [Codgen] Thread variable use before
define
URL: https://github.com/apache/incubator-tvm/pull/4270#issuecomment-551229481
That is a reasonable suggestion, the issue is it does prevent being able to
construct certain loops. I've gone back and forth
yongwww commented on issue #4273: [TOPI] NMS support for Tensorflow
URL: https://github.com/apache/incubator-tvm/pull/4273#issuecomment-551230656
Have closed it, sent it out by mistake, will pr later once it is ready.
Sorry about that.
tqchen commented on issue #4270: [Codgen] Thread variable use before define
URL: https://github.com/apache/incubator-tvm/pull/4270#issuecomment-551226490
In this particular case, shall we do compute_at on the ```op.axis[1]```
instead? Given that B is a local stage and does not populates
yongwww opened a new pull request #4273: Nms
URL: https://github.com/apache/incubator-tvm/pull/4273
Thanks for contributing to TVM! Please refer to guideline
https://docs.tvm.ai/contribute/ for useful information and tips. After the pull
request is submitted, please request code reviews
jwfromm commented on a change in pull request #4271: [Relay][Frontend][ONNX]
operator support: DepthToSpace, SpaceToDepth
URL: https://github.com/apache/incubator-tvm/pull/4271#discussion_r343818683
##
File path: python/tvm/relay/frontend/onnx.py
##
@@ -466,6 +466,76 @@
jwfromm commented on a change in pull request #4271: [Relay][Frontend][ONNX]
operator support: DepthToSpace, SpaceToDepth
URL: https://github.com/apache/incubator-tvm/pull/4271#discussion_r343821615
##
File path: tests/python/frontend/onnx/test_forward.py
##
@@ -142,6
MarisaKirisame commented on a change in pull request #4083: [tutorial] Relay
pass infra tutorial
URL: https://github.com/apache/incubator-tvm/pull/4083#discussion_r343817823
##
File path: tutorials/dev/relay_pass_infra.py
##
@@ -0,0 +1,253 @@
+# Licensed to the Apache
tmoreau89 opened a new issue #4272: [VTA] Tutorial on how to deploy and execute
model on device without RPC
URL: https://github.com/apache/incubator-tvm/issues/4272
This is a follow up to forum post:
vinx13 commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343774224
##
File path: src/api/api_pass.cc
##
@@ -94,6 +94,13 @@ TVM_REGISTER_API("ir_pass.StorageFlatten")
}
});
vinx13 commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343779087
##
File path: tutorials/optimize/opt_matmul_auto_tensorcore.py
##
@@ -0,0 +1,317 @@
+# Licensed to the Apache Software
vinx13 commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343781065
##
File path: tutorials/optimize/opt_matmul_auto_tensorcore.py
##
@@ -0,0 +1,317 @@
+# Licensed to the Apache Software
vinx13 commented on issue #4268: [Codegen][cuda-fp16] fallback to fp32
simulation when cuda arch < sm53
URL: https://github.com/apache/incubator-tvm/pull/4268#issuecomment-551179870
I agree better we can put the code in a separate file instead of string
concatenation. The code itself is
csarofeen opened a new pull request #4270: [Codgen] Thread variable use before
define
URL: https://github.com/apache/incubator-tvm/pull/4270
There are instances where threads can be used before defined. This can
happen, for example, when a consumer stage has a thread dimension bound to an
minminsun commented on issue #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#issuecomment-551094241
Thanks @Laurawly @Hzfengsy @vinx13 for your review comments and suggestions.
We have improved the code accordingly, and have added a formal tutorial. Do
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343661298
##
File path: python/tvm/build_module.py
##
@@ -387,6 +387,7 @@ def lower(sch,
binds, arg_list = get_binds(args,
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343661212
##
File path: tutorials/autotvm/tune_tensor_core_batch_matmul.py
##
@@ -0,0 +1,230 @@
+import logging
+import sys
+
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343659974
##
File path: include/tvm/ir_pass.h
##
@@ -206,6 +206,20 @@ Stmt StorageFlatten(Stmt stmt,
Map
minminsun commented on a change in pull request #4234: Auto TensorCore CodeGen
URL: https://github.com/apache/incubator-tvm/pull/4234#discussion_r343660054
##
File path: src/pass/tensor_core.cc
##
@@ -0,0 +1,1209 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF)
jackwish commented on issue #3859: [WIP] [AutoTVM][TOPI] NHWC conv2d templates
for ARM
URL: https://github.com/apache/incubator-tvm/pull/3859#issuecomment-550989430
> Thanks @jackwish for the update. Makes sense. So, are we planning to close
this PR?
I agree to close this - data
78 matches
Mail list logo