triton-windows 3.2.0.post11__cp312-cp312-win_amd64.whl → 3.3.0a0.post11__cp312-cp312-win_amd64.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Potentially problematic release.

This version of triton-windows might be problematic. Click here for more details.

Files changed (68) hide show
  1. triton/_C/libtriton.pyd +0 -0
  2. triton/__init__.py +3 -3
  3. triton/_internal_testing.py +59 -4
  4. triton/_utils.py +35 -0
  5. triton/backends/amd/compiler.py +121 -74
  6. triton/backends/amd/driver.py +77 -43
  7. triton/backends/amd/include/hip/amd_detail/amd_device_functions.h +28 -49
  8. triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h +35 -9
  9. triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h +761 -284
  10. triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h +9 -3
  11. triton/backends/amd/include/hip/amd_detail/amd_hip_fp8.h +1391 -0
  12. triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h +3 -3
  13. triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h +44 -0
  14. triton/backends/amd/include/hip/amd_detail/amd_warp_sync_functions.h +288 -0
  15. triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp +110 -14
  16. triton/backends/amd/include/hip/amd_detail/hip_prof_str.h +504 -103
  17. triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h +2 -1
  18. triton/backends/amd/include/hip/amd_detail/host_defines.h +4 -0
  19. triton/backends/amd/include/hip/hip_ext.h +4 -2
  20. triton/backends/amd/include/hip/hip_fp8.h +33 -0
  21. triton/backends/amd/include/hip/hip_runtime_api.h +375 -33
  22. triton/backends/amd/include/hip/hip_version.h +3 -3
  23. triton/backends/amd/include/hip/hiprtc.h +25 -25
  24. triton/backends/amd/include/hsa/amd_hsa_elf.h +40 -14
  25. triton/backends/amd/include/hsa/hsa.h +11 -2
  26. triton/backends/amd/include/hsa/hsa_api_trace.h +30 -17
  27. triton/backends/amd/include/hsa/hsa_api_trace_version.h +68 -0
  28. triton/backends/amd/include/hsa/hsa_ext_amd.h +83 -27
  29. triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h +46 -46
  30. triton/backends/amd/include/hsa/hsa_ven_amd_pc_sampling.h +416 -0
  31. triton/backends/amd/include/roctracer/hip_ostream_ops.h +84 -4
  32. triton/backends/amd/include/roctracer/hsa_ostream_ops.h +260 -0
  33. triton/backends/amd/include/roctracer/hsa_prof_str.h +51 -19
  34. triton/backends/amd/lib/asanrtl.bc +0 -0
  35. triton/backends/compiler.py +25 -225
  36. triton/backends/driver.py +7 -2
  37. triton/backends/nvidia/bin/ptxas.exe +0 -0
  38. triton/backends/nvidia/compiler.py +135 -90
  39. triton/backends/nvidia/driver.c +0 -1
  40. triton/backends/nvidia/driver.py +135 -49
  41. triton/backends/nvidia/include/cuda.h +2162 -241
  42. triton/backends/nvidia/lib/x64/cuda.lib +0 -0
  43. triton/compiler/__init__.py +2 -2
  44. triton/compiler/code_generator.py +334 -231
  45. triton/compiler/compiler.py +77 -66
  46. triton/language/__init__.py +22 -5
  47. triton/language/core.py +448 -74
  48. triton/language/extra/cuda/_experimental_tma.py +3 -5
  49. triton/language/math.py +1 -1
  50. triton/language/random.py +2 -1
  51. triton/language/semantic.py +206 -52
  52. triton/language/standard.py +35 -18
  53. triton/runtime/_allocation.py +32 -0
  54. triton/runtime/autotuner.py +27 -32
  55. triton/runtime/build.py +1 -48
  56. triton/runtime/cache.py +6 -6
  57. triton/runtime/errors.py +10 -0
  58. triton/runtime/interpreter.py +179 -45
  59. triton/runtime/jit.py +149 -190
  60. triton/testing.py +39 -11
  61. triton/tools/compile.py +27 -20
  62. triton/tools/{compile.c → extra/cuda/compile.c} +1 -0
  63. triton/tools/mxfp.py +301 -0
  64. {triton_windows-3.2.0.post11.dist-info → triton_windows-3.3.0a0.post11.dist-info}/METADATA +5 -2
  65. {triton_windows-3.2.0.post11.dist-info → triton_windows-3.3.0a0.post11.dist-info}/RECORD +68 -59
  66. {triton_windows-3.2.0.post11.dist-info → triton_windows-3.3.0a0.post11.dist-info}/top_level.txt +2 -0
  67. /triton/tools/{compile.h → extra/cuda/compile.h} +0 -0
  68. {triton_windows-3.2.0.post11.dist-info → triton_windows-3.3.0a0.post11.dist-info}/WHEEL +0 -0
triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h CHANGED
@@ -216,12 +216,18 @@ class thread_block : public thread_group {
216
216
  if (!tile_size || (tile_size > __AMDGCN_WAVEFRONT_SIZE) || !pow2) {
217
217
  __hip_assert(false && "invalid tile size");
218
218
  }
219
+
220
+ auto block_size = size();
221
+ auto rank = thread_rank();
222
+ auto partitions = (block_size + tile_size - 1) / tile_size;
223
+ auto tail = (partitions * tile_size) - block_size;
224
+ auto partition_size = tile_size - tail * (rank >= (partitions - 1) * tile_size);
225
+ thread_group tiledGroup = thread_group(internal::cg_tiled_group, partition_size);
219
226
 
220
- thread_group tiledGroup = thread_group(internal::cg_tiled_group, tile_size);
221
227
  tiledGroup.coalesced_info.tiled_info.size = tile_size;
222
228
  tiledGroup.coalesced_info.tiled_info.is_tiled = true;
223
- tiledGroup.coalesced_info.tiled_info.meta_group_rank = thread_rank() / tile_size;
224
- tiledGroup.coalesced_info.tiled_info.meta_group_size = (size() + tile_size - 1) / tile_size;
229
+ tiledGroup.coalesced_info.tiled_info.meta_group_rank = rank / tile_size;
230
+ tiledGroup.coalesced_info.tiled_info.meta_group_size = partitions;
225
231
  return tiledGroup;
226
232
  }
227
233