GT4Py v1.0.10

Summary of changes since v1.0.9

Cartesian

• New backend dace:cpu_kfirst.
• New experimental features:
  • absolute indexing in K.
  • expose K index.
• Fixes and performance improvements in DaCe backends.
• Improved error messages.

Development

• Enabled DaCe backends on AMD MI300 CI.

Next

See commit history.

All changes

• CI: enable dace tests on beverin by @edopao in #2263
• fix[next][dace]: check for gpu schedule before adding cudaStreamSynchronize by @edopao in #2260
• feat[next]: Transformation pass from runtime to compile-time domains by @SF-N in #2151
• feat[cartesian]: Add backend dace:cpu_kfirst by @romanc in #2255
• fix[cartesian]: dace backend fixes by @romanc in #2259
• refactor[dace]: Removed Dublicated Code by @philip-paul-mueller in #2272
• fix[next][dace]: move cuda-codegen setting to gt4py config by @edopao in #2271
• fix[next][dace]: ensure unique names for conditional blocks and loop regions by @edopao in #2270
• perf[next][dace] Remove scalar views by @iomaganaris in #2166
• bug[next]: Fix non scan projector in global tmp pass by @tehrengruber in #2274
• tests[cartesian]: attempt to stabilize macos daily (uv resolution highest) by @romanc in #2269
• refactor[next]: Refactor static args mechanism by @tehrengruber in #2258
• fix[next]: Fix multi-node scaling issues of FileCache by @tehrengruber in #2261
• build[next]: update dace version by @edopao in #2275
• feat[next][dace]: extend CopyChainRemover to remove full-write copies by @edopao in #2273
• feature[cartesian]: absolute indexing in K (experimental) by @romanc in #2276
• fix[next]: Fix allocator typing by @havogt in #2279
• feat[next]: synchronous and synchronize compilation by @havogt in #2096
• feat[next][dace]: collect sdfg execution time with tasklet by @edopao in #2280
• feat[dace][next]: Added Inline Fuser by @philip-paul-mueller in #2283
• feat[next][dace]: Hooksystem for gt_auto_optimizer() by @philip-paul-mueller in #2291
• feat[next]: flexible metric collections by @egparedes in #2287
• refactor[cartesian]: shiny error messges from the gt_script frontend by @romanc in #2042
• feat[dace][next]: Set GPU Thread Block Size Per Dimension from gt_auto_optimizer() by @philip-paul-mueller in #2298
• cartesian[feat]: Expand push vertical map down to include ForScope by @FlorianDeconinck in #2290
• feat[dace][next]: Updated Splitting Fusion Transformations by @philip-paul-mueller in #2296
• bug[next]: Fix dead-code elimination on some concat_where expressions by @tehrengruber in #2292
• refactor[next]: Cleanup implicit offsets by @havogt in #2299
• feat[bug]: Fix name collision in collapse tuple pass by @tehrengruber in #2301
• refactor[next][dace]: Move zero-origin constant substitution out of lowering module by @edopao in #2303
• doc[next]: Document concat_where domain arg canonicalization by @tehrengruber in #2223
• feat[next]: Prune empty concat where branches pass by @tehrengruber in #2286
• fix[next][dace]: Avoid name conflict between tasklet connector and data node by @edopao in #2306
• build[cartesian]: update DaCe version by @romanc in #2307
• feat[next][dace]: Enable customization of auto-optimize in dace backend workflow by @edopao in #2295
• bug[next]: change flaky wait_for_compilation test by @havogt in #2297
• feat[dace][next]: Demotion of Fields by @philip-paul-mueller in #2288
• perf[next][dace] Avoid unnecessary vertical map splitting by @iomaganaris in #2278
• fix[next]: Fix offset provider access by @tehrengruber in #2310
• bug[next]: Add support for nan and inf literals in GTIR by @tehrengruber in #2308
• feat[next][dace]: Updated gt_auto_optimizer()'s Hook-System by @philip-paul-mueller in #2294
• Revert "fix[next][dace]: Avoid name conflict between tasklet connector and data node" by @edopao in #2317
• perf[next][dace] Set GPU thread block size properly by @iomaganaris in #2313
• fix[next][dace]: Avoid name conflict between tasklet connector and data node by @edopao in #2318
• feat[next]: upgrade tree_map with custom constructor per original type by @havogt in #2285
• build[cartesian]: remove reference to gtpyc from pyproject.toml by @romanc in #2320
• fix[next][dace]: SDFG instrumentation with empty else-branch caused memory leak by @edopao in #2321
• test[cartesian]: remove duplicate dace parsing tests by @romanc in #2325
• feat[cartesian]: utils.warn_experimental_feature() by @romanc in #2324
• refactor[next][dace]: cleanup update_sdfg_args by @edopao in #2312
• refactor[next]: Reduce programs runtime overhead by @egparedes in #2305
• feat[next][dace]: enable overriding of cxx/cuda compiler arguments by @edopao in #2327
• Fix thread block size setting assertion by @iomaganaris in #2331
• feat[next]: Implement faster data pointer retrieval in ndarray-based fields by @egparedes in #2332
• feature[cartesian]: expose iteration index in K (experimental) by @romanc in #2300
• build[dace][next]: Updated DaCe Dependency by @philip-paul-mueller in #2328
• refactor[next][dace]: Use global array instead of return value to collect the SDFG compute time by @edopao in #2333
• Releasing v1.0.10 by @havogt in #2330

Full Changelog: v1.0.9...v1.0.10