Commits

Branch selector

 master

User selector

timshen91

Datepicker

All time

Commit History

Commits on Jun 11, 2019

• [XLA] Add a flag to let convolution layout assignment use TF's layouts (all NCHW).

  Show description for c309df6
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  c309df6

  Copy full SHA for c309df6

• [XLA] Add a flag to explicitly enable a list of passes to run.

  Show description for 543ddf5
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  543ddf5

  Copy full SHA for 543ddf5

Commits on Jun 5, 2019

• [XLA] Scale down fp16 convolution autotuning input so that it's less overflowy.

  Show description for 8ac1075
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  8ac1075

  Copy full SHA for 8ac1075

• [XLA] Consistently pick convolution algorithms with WRONG_RESULT as commented.

  Show description for e6072c0
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  e6072c0

  Copy full SHA for e6072c0

Commits on Jun 4, 2019

• Use determinisitc, default-seeded random inputs for convolution autotuning.

  Show description for e23aafa
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  e23aafa

  Copy full SHA for e23aafa

Commits on Jun 3, 2019

• [XLA] Seed each convolution with the same rng state, so that the conv autotuning input is consistent even when run individually.

  Show description for 3679396
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  3679396

  Copy full SHA for 3679396

Commits on Jun 1, 2019

• [XLA] Add regression tests detected by XLA convolution checker.

  Show description for 68974ba
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  68974ba

  Copy full SHA for 68974ba

Commits on May 31, 2019

• [XLA] Randomize buffer content during convolution autotuning.

  Show description for e734bb0
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  e734bb0

  Copy full SHA for e734bb0

Commits on May 23, 2019

• [XLA] Randomize the F32/F64 convolution checker input to [0, 1). This potentially discovers more cuDNN bugs due to the non-broadcast input.

  Show description for ac279bf
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  ac279bf

  Copy full SHA for ac279bf

• Adjust logging in the algorithm picker. Print more information under VLOG(1) instead of relying on the crasher flag, because they are indepdent intentions.

  Show description for 310a4a4
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  310a4a4

  Copy full SHA for 310a4a4

Commits on May 21, 2019

• Add "compile_only" mode to replay_computation.

  Show description for 329fe4c
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  329fe4c

  Copy full SHA for 329fe4c

Commits on May 17, 2019

• Expose GetCudnnConvParams for testing and debugging. NFC.

  Show description for ec9e993
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  ec9e993

  Copy full SHA for ec9e993

• [XLA] This CL moves most logic in the original RunCudnnConvImpl to GetCudnnConvParams. Previously, parameters were only partially computed in GetCudnnConvParams, and most of the fields were filled …

  Show description for 6b657e3
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  6b657e3

  Copy full SHA for 6b657e3

Commits on May 16, 2019

• Fix --config=monolithic build.

  Show description for 0166739
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  0166739

  Copy full SHA for 0166739

Commits on May 15, 2019

• Add recordio support for replay_computation_gpu.

  Show description for d165200
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  d165200

  Copy full SHA for d165200

• Fix Conv3D's layout assignment. According to cuDNN's documentation, the filter type must be NCHW.

  Show description for fef53af
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  fef53af

  Copy full SHA for fef53af

• Print more information to local logs when `crash_on_checking_failure` flag is on.

  Show description for 6918721
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  6918721

  Copy full SHA for 6918721

Commits on May 11, 2019

• Automated rollback of commit b9b83e245ce2036e6b286d57e391408a70dfd8a6

  Show description for 7ce7ba4
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  7ce7ba4

  Copy full SHA for 7ce7ba4

Commits on May 9, 2019

• Add convolution kind FORWARD_BIAS_ACTIVATION to StreamExecutor. Previously we thought that the combination of `kind = FORWARD` and `ActivationMode != kNone` is a necessary-sufficient condition for …

  Show description for 1ba157f
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  1ba157f

  Copy full SHA for 1ba157f

Commits on May 7, 2019

• Remove unused dependency from XLA to TF conv kernels.

  Show description for 3a459ac
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  3a459ac

  Copy full SHA for 3a459ac

• [XLA] Clarify the dimension ordering.

  Show description for f56ba9f
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  f56ba9f

  Copy full SHA for f56ba9f

Commits on Apr 26, 2019

• Automated rollback of commit a964ec46df9df2e11d51815382e5ad39a52ba50f

  Show description for 72781a0
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  72781a0

  Copy full SHA for 72781a0

Commits on Apr 24, 2019

• Changed RNN's compute type to fp32. This is because

  Show description for cd84019
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  cd84019

  Copy full SHA for cd84019

Commits on Apr 12, 2019

• Bring back ConvolutionProto for logging, instead of dumping NodeDef. There is fair amount of logic between the NodeDef representation and StreamExecutor representation, which is not easy to replica…

  Show description for 9cc5d48
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  9cc5d48

  Copy full SHA for 9cc5d48

Commits on Mar 23, 2019

• Crash the program under debug flags only after logging the convolution protos.

  Show description for 9acfe8a
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  9acfe8a

  Copy full SHA for 9acfe8a

• Add F32 and F64 support for buffer comparator.

  Show description for ec296ed
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  ec296ed

  Copy full SHA for ec296ed

Commits on Mar 22, 2019

• Fix unintended nullptr arithmetic in BFCAllocator. This fixes occasional crashes when dumping OOM message under VLOG(1).

  Show description for cf00155
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  cf00155

  Copy full SHA for cf00155

Commits on Mar 21, 2019

• Push some of the buffer comparator logic to algorithm picker. Buffer comparator should do just comparisons.

  Show description for 822516c
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  822516c

  Copy full SHA for 822516c

Commits on Mar 15, 2019

• Log a convolution only during an Autotuning cache miss.

  Show description for 013e062
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  013e062

  Copy full SHA for 013e062

Commits on Mar 7, 2019

• Roll-forward:

  Show description for 94be8f0
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  94be8f0

  Copy full SHA for 94be8f0

Commits on Mar 5, 2019

• Log convolutions during Tensorflow GPU conv autotuning. Also removed the same functionality from StreamExecutor.

  Show description for 5aefc4e
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  5aefc4e

  Copy full SHA for 5aefc4e

• Roll-forward:

  Show description for 82ff4f8
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  82ff4f8

  Copy full SHA for 82ff4f8

Commits on Mar 2, 2019

• Automated rollback of commit b562be27f705034d46e9920ebbcdd293485a7305

  Show description for fc33a70
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  fc33a70

  Copy full SHA for fc33a70

Commits on Mar 1, 2019

• Move autotuning.proto to tensorflow/core/protobuf. This is in preparation of the next patch, to tensorflow::Logger::LogProto() for TF convolutions.

  Show description for b562be2
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  b562be2

  Copy full SHA for b562be2

Commits on Feb 8, 2019

• Export GPU convolution autotuning results from XLA.

  Show description for 3de71d7
  

  timshen91

  authored and

  tensorflower-gardener

  committed
  3de71d7

  Copy full SHA for 3de71d7

Pagination

Previous

Next