Merging changes for ezpc from internal repo. #1

nishkum · 2019-06-27T10:40:07Z

More specifically, some of these changes include:

Added extern function support to ezpc.
Added SecureNN as backend to ezpc.
More bug fixes.
All test_suite examples' precompiled-output are updated with new compiler changes.
Added an updated and more complex sample compiled test file in ABY_example.

Tests done

"make runtest" works fine for all test_suite examples.
Sample code in ABY_example is also tested to be working with latest ABY.

…nternal repo. More specifically, some of these changes include: - Added extern function support to ezpc. - Added SecureNN as backend to ezpc. - More bug fixes. - All test_suite examples' precompiled-output are updated with new compiler changes. - Added an updated and more complex sample compiled test file in ABY_example. Tests done - "make runtest" works fine for all test_suite examples. - Sample code in ABY_example is also tested to be working with latest ABY.

BUG DESCRIPTION: Arithmetic right shift used to lead to segfaults when used with ABY. FIX: - Updated function arithmetic_right_shift in file ABY_example/common/ezpc.h - Adding correct gate ids before feeding the new share value to create a share.

EzPC/SecureNN_example/ezpc.h

EzPC/codegen.ml

EzPC/codegensecurenn.ml

EzPC/infer.ml

EzPC/tcenv.ml

Gcf psi

Add information about supported nodes in ONNX.

* conv3d hotfix * conv transpose * ConvTranspose3D layer * graph gen refactor #1 * general refactor * general refactor oops * refactor functional layers * WIP: mutliple input support * updates in OnnxBridge for 3d conv and convtranspose --------- Co-authored-by: Kanav Gupta <[email protected]>

* rename Tensor to Tensor1D * save progress * General tensor * 5d model support * conv2d multiple filters * OnnxBridge modification for mlp and input code in sytorch for mlp * input function generalised * added function to get value at index in Tensor * 3d conv cleartext * broadcasting added and input in Tensor directly * conv3d hotfix * conv transpose * ConvTranspose3D layer * graph gen refactor #1 * general refactor * general refactor oops * refactor functional layers * WIP: mutliple input support * updates in OnnxBridge for 3d conv and convtranspose * bugfix and multiple padding support in conv3d * function call updated * better functional api * function call updated * conv3d and convtranspose3d refactor * function call updated * add concat format modified in OnnxBridge * printing api change * Update sytorchBackendRep.py * Update sytorch_func_calls.py * batchnorm is general now * Update compile_llama.sh * Update sytorchBackendRep.py * deleted take_input function, as not in use rn * add input_nchw --------- Co-authored-by: Saksham Gupta <[email protected]>

* add gpu code * carvanha codegen * average softmax and more stats * massive update * forward + softmax works * fixed seed and cpu fwd activation matches gpu * fix: push 3layer code * local truncation emulation in cleartext and pvgg16 * reluextend in cpu code * new truncation design, somewhat tested * wip commit * wip 2 * reluextend hotfix, might still be wrong? * relu extend probably works now * minor change * llama improved and carvahna * label subtraction in gpu * two round sign extension * save before full run * these bugs make me wonder my sanity * update main * fix maxpool timing bug * deterministic dealer * hot fix * save progress * some changes * delete perf files * benchmarking commit * rt microbenchmark * save code * minor fix

nishkum requested a review from aseemr June 27, 2019 10:40

BUG FIX

d9c1d4e

BUG DESCRIPTION: Arithmetic right shift used to lead to segfaults when used with ABY. FIX: - Updated function arithmetic_right_shift in file ABY_example/common/ezpc.h - Adding correct gate ids before feeding the new share value to create a share.