[Relay][Quantize] Use fixed point mulplications #4160

vinx13 · 2019-10-19T16:41:31Z

Previously we fall back to float computation when the scale is not power of 2. This PR uses fixed point multiplications instead.
See discussion in https://discuss.tvm.ai/t/quantization-float-multiplication-after-conv/4135/3

@ZihengJiang @anijain2305

anijain2305 · 2019-10-19T17:03:39Z

src/relay/pass/quantize/realize.cc

@@ -97,7 +98,9 @@ inline Expr ForwardOp(const Call& ref_call, const Array<Expr>& args) {


 /* calculate `data * s1 / s2`, use shift if possible */
-inline Expr MulAndDiv(Expr data, float s1, float s2, DataType dtype) {
+inline Expr MulAndDiv(Expr data, float s1, float s2, DataType dtype,


This seems like a requantize operation? Is it?

Maybe, we should directly call QNN Requantize operator then. And hide the power of 2 handling inside the requantize/Fixed point multiplication. Special power of 2 handling will help QNN as well. I am fine this PR going in first, and later refactoring. Just putting the point across. Requantize will be a better name for the method.

anijain2305

LGTM.

Only couple of comments

In future, we should think about how can we move to QNN requantize op and clean the realize pass here.
Did you get a chance to run this and check e2e accuracy?

vinx13 · 2019-10-21T19:18:02Z

~~- I tested Resnet18 v2 (with non-power 2 scales from KLD), the accuracy decreased from 69.1% to 68.2%. -~~
There might be something wrong with my evaluation, I'm reevaluating it

anijain2305 · 2019-10-21T20:30:50Z

Ok. I would expect the accuracy drop to some extent (not sure if 1% drop is correct or not).

How do we proceed from here? Do we try to improve KLD (or whatever the calibrator is) to see if these numbers can be brought back up.

tqchen · 2019-10-24T16:54:46Z

@vinx13 please followup and let us know when it can be merged

vinx13 · 2019-10-24T17:00:30Z

Here are the evaluation result on imagenet val, with rounding = "UPWARD" and KLD scales

Resnet18 v2
acc-top1=0.707080 acc-top5=0.898180
Resnet 50 v2
acc-top1=0.764980 acc-top5=0.931160
Resnet 18 v1
acc-top1=0.707360 acc-top5=0.897860
Resnet 50 v1
acc-top1=0.735740 acc-top5=0.915940

Accuracy drop is 0.1%~0.3%.
Previous results are in #3538

@anijain2305 @ZihengJiang Please take a look
Note that the nightly build will not be affected since KLD scales are not used.

tqchen · 2019-10-24T17:01:50Z

Looks reasonable, @vinx13 can you also comment on the speedup? it would be great if you can also send a tutorial about using the quantization infra in a separate PR :)

anijain2305 · 2019-10-24T17:18:30Z

Thanks @vinx13 The performance numbers should improve now, given we are avoiding casting to FP32 and than back to Int :) Moreover, from a quantized network perspective, keeping everything in int seems like the right thing to do (unless there is huge accuracy impact).

anijain2305 · 2019-10-28T21:44:46Z

Ping. I think this can be merged now.

@vinx13 @tqchen

vinx13 · 2019-10-29T04:35:16Z

Yes we can get this in. I'm working on evaluation of speed

* [relay][vm] Separate VM runtime with executable (apache#4100) * [relay][vm] Separate VM runtime with executable * Address comments * move ctx back to vm * make only vm related fields and methods protected * integrate seriliaztion/deserialization to executable * create stream * [Relay][Frontend][TF] Add tensor array ops (apache#3798) * [Relay][Frontend][TF] Add tensor array ops * rename * delete test * Move utility function * Refactor * fix tensor array ops * fix test * fix rebase * Fix serializer bug * Improve tf convert name lookup to use prelude api * Fix lint * Fix test * Fix typo (apache#4144) * [CI] Pin NNPack pthreadtools version (apache#4152) * [QNN][TFLite] Parsing QNN Add op. Adding MobilenetV2. (apache#4142) * Add lift_if_then_else pass (apache#3865) * Add LiftIfThenElse pass * Add more comments * Rename and refactor * Add description for internal data structure * Rename a test * Minor change * Address comments * Improve update_for * [CI] Update cpu docker (apache#4153) * [Refactor] Rename Datatype to ADT (apache#4156) We think it will reduce the confusion with the meaning. https://discuss.tvm.ai/t/discuss-consider-rename-vm-datatype/4339 * [Runtime] Enable option to use OpenMP thread pool (apache#4089) * [REFACTOR][NODE][RUNTIME] Move Node to the new Object protocol. (apache#4161) * [REFACTOR][NODE][RUNTIME] Move Node to the new Object protocol. This PR removes the original node system, and make node as a subclass of Object. This is a major refactor towards a better unified runtime object system. List of changes in the refactor: - We now hide data_ field, use Downcast explicitly to get a sub-class object. - Removed the node system FFI in python. - Removed the node C API, instead use PackedFunc for list and get attrs. - Change relay::Op::set_attr_type_key(attr_key_name) to relay::Op::set_attr_type<AttrType>(). - This change was necessary because of the new Object registration mechanism. - Subsequent changes to the op registrations - The change revealed a few previous problems that is now fixed. - Patched up a few missing node type registration. - Now we will raise an error if we register object that is not registered. - The original node.h and container.h are kept in the same location. - Calling convention: kObjectHandle now equals the old kNodeHandle, kNodeHandle is removed. - IRFunctor now dispatches on ObjectRef. - Update to the new type checking API: is_type, derived_from are replaced by IsInstance. - Removed .hash member function, instead use C++ convention hasher functors. * Address review comments * [CI] Move golang tests to the end (apache#4164) * Add support for quantized multiply to Relay (apache#4141) This patch adds multiply operator for quantized tensors. The details of the quantized multiplication are outlined in the code. This builds on pull request 3927 and includes the changes Animesh mentions in the comments on that request. Change-Id: I555715b53d0266a91d5c03dc3dfe8fc31e7ce4e1 * Fix missspelling (apache#4166) FIX "After connecting he usb" with "After connecting the usb" * [Relay][Pass] Count MAC for BatchMatMul (apache#4157) * count MAC for BatchMatMul * update doc * [Relay][QNN] Add unit test for int8 (apache#4159) * [bugfix][codegen] fix casting bug in llvm codegen * update example * retrigger ci * check llvm version * [relay][vm] Reuse allocated device memory (apache#4170) * add missing gradient check to gradient pass (apache#4169) * merge extract_from_program and extract_from_multiple_progam (apache#4173) * [TOPI] Added support for Mali Bifrost target (apache#4047) * [Relay][Frontend][TF] Fix Size operator (apache#4175) * [Relay][Frontend][TF] Fix Size operator * Uncomment tests * [Pass] Remove dead code (apache#4177) * [rpc] use callback func to do send & recv (apache#4147) * [rpc] use callback func to do send & recv. don't get fd from sock as it is deprecated in java * fix java build * fix min/max macro define in windows * keep the old rpc setup for py * add doc for CallbackChannel * Add support and testing for tf.assert (as no-op) and tf.no_op to TF Relay frontend. (apache#4172) * [DOCS] Add TensorFlow frontend docs (apache#4154) * Start to update TF frontend docs * Add rst * Remove markdown * Update wording * Resolve comments * Revert "[Relay][QNN] Add unit test for int8 (apache#4159)" (apache#4192) This reverts commit 6f9d028. * [cmake][ANTLR] Support setting path to ANTLR jar (apache#4176) * Support setting path to ANTLR jar * Update comment * Split adaptive_pool2d_avg into sum and div (apache#4186) * [Documentation]Fix example code in comment of tvm.build_module.build() (apache#4195) * Fix example code in comment of tvm.build_module.build() * Update build_module.py * [relay] use time_evaluator for measurement (apache#4191) * Add parser support for SUM tflite operator (apache#4182) * [Relay] Fix memory leak in the interpreter (apache#4155) * save lint * address reviewer comment * [TOPI] Tunable Template for Conv2D HWCN on CUDA (apache#4168) * support conv2d HWCN in AutoTVM and Relay * fix lint * fix comments and unit tests * TensorCore Support using Intrinsic (apache#4136) * add tensor core support * avoid memory bank conflict * fix thread sync & better performance * better performance * add schedule test for conv2d * extend into BatchMatMul * support config fragment shape and layout using intrinsic * add TensorCore tutorial * add int support and fix lint * address comment * add 32*16*8 TensorCore test * fix wmma include logic * [NODE][REFACTOR] Refactor reflection system in node. (apache#4189) * [NODE][REFACTOR] Refactor reflection system in node. - Removed the old Node, Node is now just an alias of runtime::Object - Introduce ReflectionVTable, a new columnar dispatcher to support reflection - This allows us to remove vtable from most node objects - The VisitAttrs are registered via TVM_RESGITER_NODE_TYPE, they are no longer virtual. - Consolidated serialization and reflection features into node. * Explicit type qualification when calling destructor. * Fix SPIRV, more comments * hotfix the ci (apache#4199) * [TOPI][x86] Legalize - Support int8xint8 convolution to use VNNI instructions. (apache#4196) * [Relay] crossentropy_with_logits and its gradient (apache#4075) * save * lint * [hotfix] missing include headers (apache#4204) * [Relay][Training] Add checkpoint annotation for checkpointing memory optimization (apache#4146) * add checkpoint annotation for checkpointing memory optimization * add alpha-equivalence checkpoint test and fix gradient type issue * fix build issues * ignore checkpoint annotation when checking missing gradients * refactor, fix checkpoint compute for tuple and add tests * [Relay][Params] Add APIs for storing and retrieving parameters from individual functions. (apache#4194) * Add support for attaching params * Fix types * Fix test * [Relay][Frontend][ONNX] Add support for op Where (apache#4184) * Add support for op Where * Update impl version * [VTA][Chisel] TSIM VTA Source Refactor (apache#4163) * app init push * fix on readme * change name, add bit serial explanantion * rm serialLoadMM, change doc * syntax change for readme * add parallel test functionality * fix readme * add python doc * syntax * init commit * fix empty line * fix typo * [RUNTIME] Separate runtime related contrib into runtime/contrib (apache#4207) * Fix type var docs (apache#4208) * [Relay] Setting Legalize opt_level to 1. (apache#4198) * [TOPI] Fix flaky testcase for check round (apache#4211) * [Relay][Op] Enhance Upsample Operator to support float scales (apache#4206) * :add scale2 for upsample * update unit test for upsampling * support latest upsample op for multiple frontend * fix lint * fix lint * fix lint * fix lint * update scale description and rebase * [Relay][Quantize] Use fixed point mulplications (apache#4160) * Update have_int8 condition to run on compute capability 7.x devices (apache#4214) * Optimizing autotvm task extraction speed (apache#4138) * Optimize task extraction speed * correct pylint errors * Delete unused function * remove unnecessary argument * resolve code review comments * corrent cpp lint errors * remove one more graph_json return value * fix test bugs * [Relay] Add Python type functor and tests (apache#4209) * Add Python type functor and tests * Lint roller * Fix typo in packed_func.h (apache#4219) * Improve the lowering of Qnn Dense (apache#4213) * [QNN] Improving Dense lowering. * - Moving get_shape method to util - Finalizing the test cases and the code structure for optimized dense computation. * - Fixing cpplint. * - Addressing review comments. * - Renaming the variables correctly. * - Renaming the variables correctly. * [ARITH] Fix the rule y < x && x <= y (apache#4220) * [PYTHON] Add __init__ to the generated grammar so that it can be installed properly (apache#4223) * [Relay][Frontend][ONNX] New Operators and Opsets to Support BERT (apache#4197) * Added slice v10 * Added constantofshape operation and small refactor. * Finished one_hot implementation. * Reshape working across all bert layers. * Fixed constantofshape and removed code duplication. * onnx model fully ingested. * Working on improving onnx tests. * Changed onnx testing to use onnxruntime instead of caffe2, also formatted. * Add arbitrary output nodes to onnx frontend. * Added v6 tiling for bert squad 8 support. * Small syntax fixes * Reduced code duplication in split opset versions. * Added batch matmul test * Added unstack split testing. * Adde onehot test, needs a little cleanup probably. * Replaced deprecated constant fill with constantofshape and updated tests accordingly. * Added tests for new opset version of slice and tile. * lint clean up * Lint fixes * Changed onnx dependency * Went back to caffe2 runtime for CI integration. * Rebase and small typo/syntax changes. * Added hard casting of onehot attributes to int. * [Relay][Topi][TensorFlow][ONNX][Lang] Add support for Any op (apache#4205) * Add support for Any op * Support ONNX frontend * Add doc * Add to relay docs * Dummy change to retrigger CI * Update dmlc_tvm_commit_id.txt * Merge from upstream

[Relay][Quantize] Use fixed point mulplications

cd27778

vinx13 force-pushed the feature/quanti_fixed_point branch from d2d2619 to cd27778 Compare October 19, 2019 16:48

anijain2305 reviewed Oct 19, 2019

View reviewed changes

anijain2305 approved these changes Oct 21, 2019

View reviewed changes

tqchen assigned ZihengJiang Oct 24, 2019

tqchen added status: need review status: need update need update based on feedbacks labels Oct 24, 2019

tqchen merged commit e889928 into apache:master Oct 29, 2019

tqchen added status: accepted and removed status: need review status: need update need update based on feedbacks labels Oct 29, 2019

kevinthesun pushed a commit to kevinthesun/tvm that referenced this pull request Oct 30, 2019

[Relay][Quantize] Use fixed point mulplications (apache#4160)

dbc1cc7

tqchen unassigned ZihengJiang Nov 4, 2019

tqchen mentioned this pull request Nov 8, 2019

[RELEASE][DRAFT] TVM v0.6 Release candidate #4259

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Relay][Quantize] Use fixed point mulplications #4160

[Relay][Quantize] Use fixed point mulplications #4160

vinx13 commented Oct 19, 2019

anijain2305 Oct 19, 2019

anijain2305 left a comment

vinx13 commented Oct 21, 2019 •

edited

Loading

anijain2305 commented Oct 21, 2019

tqchen commented Oct 24, 2019

vinx13 commented Oct 24, 2019 •

edited

Loading

tqchen commented Oct 24, 2019

anijain2305 commented Oct 24, 2019

anijain2305 commented Oct 28, 2019

vinx13 commented Oct 29, 2019

[Relay][Quantize] Use fixed point mulplications #4160

[Relay][Quantize] Use fixed point mulplications #4160

Conversation

vinx13 commented Oct 19, 2019

anijain2305 Oct 19, 2019

Choose a reason for hiding this comment

anijain2305 left a comment

Choose a reason for hiding this comment

vinx13 commented Oct 21, 2019 • edited Loading

anijain2305 commented Oct 21, 2019

tqchen commented Oct 24, 2019

vinx13 commented Oct 24, 2019 • edited Loading

tqchen commented Oct 24, 2019

anijain2305 commented Oct 24, 2019

anijain2305 commented Oct 28, 2019

vinx13 commented Oct 29, 2019

vinx13 commented Oct 21, 2019 •

edited

Loading

vinx13 commented Oct 24, 2019 •

edited

Loading