Code

Added an OpenCL kernel preprocessor#230

Code

Merged

CNugteren merged 32 commits intomasterCNugteren/CLBlast:masterfrom kernel_preprocessorCNugteren/CLBlast:kernel_preprocessorCopy head branch name to clipboardDec 17, 2017

+1,961-1,003

Commits

Commits on Nov 25, 2017

• Added stub for a preprocessor and a corresponding compilation test

  CNugteren

  committed
  c0c6d00

  Copy full SHA for c0c6d00

• Moved string splitting functions; added string character removal function

  CNugteren

  committed
  f01bcde

  Copy full SHA for f01bcde

• Implemented first simple pre-processor: defines parser and loop unrolling based on assumptions

  CNugteren

  committed
  69aa3b3

  Copy full SHA for 69aa3b3

Commits on Nov 26, 2017

• Improved the for-loop pre-processing

  CNugteren

  committed
  9c643b2

  Copy full SHA for 9c643b2

Commits on Nov 27, 2017

• Added simple implementation of array-to-register promotion

  CNugteren

  committed
  35956f9

  Copy full SHA for 35956f9

Commits on Nov 28, 2017

• Improved the kernel pre-processor in various ways

  CNugteren

  committed
  1404786

  Copy full SHA for 1404786

• Improved the pre-processor tester, added GEMV and GER kernels

  CNugteren

  committed
  4264066

  Copy full SHA for 4264066

Commits on Nov 29, 2017

• Improves the array-to-register promotion in the pre-processor

  CNugteren

  committed
  1d35f65

  Copy full SHA for 1d35f65

• Extended the preprocessor tests to include CopyFast and CopyPad

  CNugteren

  committed
  0dde6af

  Copy full SHA for 0dde6af

• Reformatted unrollable kernel loops and added the new promote_to_registers pragma for several kernels

  CNugteren

  committed
  93ffb87

  Copy full SHA for 93ffb87

Commits on Nov 30, 2017

• Integrated pre-processor in compilation flow, default is still disabled

  CNugteren

  committed
  13eb772

  Copy full SHA for 13eb772

• Improved the pre-processor's handling of defines; added a special nested defines test

  CNugteren

  committed
  bf7aeb8

  Copy full SHA for bf7aeb8

Commits on Dec 3, 2017

• Improved array to register promotion in the pre-processor

  CNugteren

  committed
  9284202

  Copy full SHA for 9284202

• Reformated transpose kernels for the pre-processor; extended the amount of tests

  CNugteren

  committed
  60312e5

  Copy full SHA for 60312e5

• Added basic bracket parsing in defines and loop expressions

  CNugteren

  committed
  0a1a3de

  Copy full SHA for 0a1a3de

• Added GEMM (direct and in-direct) to the pre-processor testing; modified the loops in kernel accordingly

  CNugteren

  committed
  cf4555d

  Copy full SHA for cf4555d

Commits on Dec 5, 2017

• Improved array-to-register promotion, now handling function calls as well

  CNugteren

  committed
  0f9637b

  Copy full SHA for 0f9637b

Commits on Dec 7, 2017

• Added register promotion to the main GEMM kernel

  CNugteren

  committed
  5408964

  Copy full SHA for 5408964

Commits on Dec 9, 2017

• Fixed defines parsing and substituting in pre-processor; fixed some variable names in kernels

  CNugteren

  committed
  d9df62b

  Copy full SHA for d9df62b

• Reformatted GEMM kernel to support array-to-register promotion

  CNugteren

  committed
  23e3a85

  Copy full SHA for 23e3a85

• Modified the direct GEMM kernel to support array-to-register promotion

  CNugteren

  committed
  02c0d64

  Copy full SHA for 02c0d64

• Made the pre-processor run by default for ARM and Qualcomm GPUs

  CNugteren

  committed
  ca5dbcd

  Copy full SHA for ca5dbcd

• Completed kernel modifications for pre-processor of all other kernels

  CNugteren

  committed
  9f02fb5

  Copy full SHA for 9f02fb5

Commits on Dec 10, 2017

• Fixed an Android compilation issue

  CNugteren

  committed
  9112e58

  Copy full SHA for 9112e58

• Fixed an issue in the tuners to prevent error -14 from persisting (CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST)

  CNugteren

  committed
  c2f08fa

  Copy full SHA for c2f08fa

• Fixed a missing include

  CNugteren

  committed
  82467b6

  Copy full SHA for 82467b6

• Updated roadmap: completed pre-processor implementation

  CNugteren

  committed
  11489e6

  Copy full SHA for 11489e6

• Split GEMM kernel in 4 files instead of 3 due to MSVC 2013 string length limit

  CNugteren

  committed
  b4d3a50

  Copy full SHA for b4d3a50

• Fixed for error C1091 in MSVC 2013

  CNugteren

  committed
  4a58efc

  Copy full SHA for 4a58efc

Commits on Dec 16, 2017

• Updated the known issues

  CNugteren

  committed
  35e2b3e

  Copy full SHA for 35e2b3e

Commits on Dec 17, 2017

• Fixed an unnecessary overflow issue on 32-bit systems

  CNugteren

  committed
  7408f6e

  Copy full SHA for 7408f6e

• Removed all ARM Mali tuning results; re-added Mali-T760 and Mali-T628 results based on kernel pre-processor

  CNugteren

  committed
  69f6591

  Copy full SHA for 69f6591