Code

History

Commits on move/encoder_kernel to dev210+
  1. added missing checksErik Schultheis committed 2 years ago7005e09
  2. fix the constness and the warning C4305: 'initializing': truncation from 'double' to 'float'lancer committed 2 years ago3cb993b
  3. separating the loops leads to yet more improvementsErik Schultheis committed 2 years agob2e6b68
  4. Merge pull request #179 from ngc92/even-better-attentionAndrej committed 2 years agob556ad9
  5. Merge pull request #169 from lancerts/fix-crossentropy-bkAndrej committed 2 years ago1d3b673
  6. further cleanup and ability to handle arbitrary sequence lengthsErik Schultheis committed 2 years ago50e105a
  7. don't do stupid redundant work; the inner loop is just a block-level sum!!!!Erik Schultheis committed 2 years ago4294820
  8. further simplification of the loopErik Schultheis committed 2 years ago02aa8e4
  9. towards an even better backward attention kernelErik Schultheis committed 2 years ago16123c3
  10. fix the 0.0f initialization for doublelancer committed 2 years ago0c92ea9
  11. fix the 0.0f initialization for doublelancer committed 2 years agoe991fe1
  12. size_t instead of int. everywhere. this will come to bite us at one point...Andrej Karpathy committed 2 years agoc393d82
  13. add kernel3, which is like kernel 2 but no float4Andrej Karpathy committed 2 years agodcff561
  14. make default dloss be 1/BTAndrej Karpathy committed 2 years ago8a2cab3
  15. small fixes and more comments very cool kernelsAndrej Karpathy committed 2 years agoea3e2ab
  16. splitting the loop to get rid of the indicatorErik Schultheis committed 2 years ago09d35a5
  17. another 15% by getting rid of ifsErik Schultheis committed 2 years ago6d98af1
  18. significantly faster kernel by reusing data in registersErik Schultheis committed 2 years ago314661f
  19. add comments to classifier_fusedAndrej Karpathy committed 2 years agode1e871
  20. Merge pull request #150 from ademeure/optimised_fused_classifierAndrej committed 2 years ago2d2f1df
  21. fix(attention_forward.cu): Fix the number of elements to check when verifying attention implementations.FeSens committed 2 years ago0c09a37
  22. fix the bug that yields cpu, gpu results mismatchlancer committed 2 years agoa1abcbc
  23. bugfix for bounds checkingademeure committed 2 years ago557fecd
  24. Merge remote-tracking branch 'karpathy/master' into optimised_fused_classifierademeure committed 2 years ago2fd8ada
  25. mini fixademeure committed 2 years agoea77a26
  26. fixes + bounds checkingademeure committed 2 years ago3ba2cdc
  27. fix the bug that yields cpu, gpu results mismatchlancer committed 2 years agoac055f7
  28. enable benchmarkingErik Schultheis committed 2 years ago58a7b57
  29. coalesced reading in attention backward passErik Schultheis committed 2 years agoe278112
  30. Merge branch 'master' into optimised_fused_classifierademeure committed 2 years ago4ee98f5
  31. Constness fix layernorm_backward.culancer committed 2 years agoc327923
  32. Constness fix layernorm_backward.culancer committed 2 years ago81aa596
  33. warp-wide reduction for layernorm backwardErik Schultheis committed 2 years agoc1bfe4a
  34. kernel2 for softmax backwardAndrej Karpathy committed 2 years agofa4ebfa
  35. small adamw fixes, needs more workAndrej Karpathy committed 2 years agod74d7ae