rocm-5.3.3

ROCm release v5.3.3

rocm-5.3.2

ROCm release v5.3.2

rocm-5.3.1

ROCm release v5.3.1

rocm-5.3.0

Improvements include...

Accuracy update (#1374)
Final performance improvements for release (#1369)
53merge v2 (#1357)
Allow license_stamper.py to be ran from any directory (#1332)
Explicitly set rocblas_pointer_mode in examples (#1331)
Imply type of literal returned based on input protobuff for zero elem… (#1326)
Dynamic ref convolution op (#1224)
Update README.md (#1327)
Improve help and error reporting in driver (#1258)
Add support for tuning db access in mlir kernel (#1307)
Add accuracy checker tool (#1315)
Avoid registering host buffer ptr multiple times during hip copies (#1245)
Add node name to debug output of PARSE_IF (#1318)
Fix literal type in the instance_norm parsing (#1317)
Add onnx mod operator (#1302)
Add fpga target (#1304)
Add performance testing yamls (#1313)
Improve error reporting in the API (#1274)
Change ownership to company email (#1310)
Dynamic check_shapes (#1295)
Fix TF parsing for creating literals and Fix name lookups for input params (#1298)
Dynamic dimension input onnx parser (#1249)
Fix op includes (#1308)
Fix test case for min & max operators (#1305)
Reduce header inclusion in op headers (#1271)
Add tests for C API (#1266)
create the dev package (#1293)
change to a cached github repo for blaze prereq (#1291)
Use current device when constructng context (#1294)
Add restrict to jit kernel params (#1300)
Improve kernel code generation (#1285)
Update perf report to show the number of operators and per operator avg time in summary (#1287)
Add env var to enable debug symbols for gpu kernels (#1284)
Add is_supported and get_target_assignments (#1269)
Dyn shape update (#1199)
Add a step to unsqeeze axis (#1242)
Verify load and save (#1265)
Add jit softmax (#1243)
Horizontally fuse contiguous operators (#1232)
Add mlir fusion (#1251)
Add method to insert multiple instructions (#1178)
Invalid parameter for yolov4 example (#1275)
NMS refactor, enable nonstandard shape (#1257)
Update driver models to use json strings (#1244)
Custom Op example using MIOpen calls (#1208)
Custom Op example using rocBLAS calls (#1211)
Custom Op example using HIP kernel (#1200)
Get parent module in the pass manager (#1181)
bug fix: register the miopen_fusion op. (#1267)
Use jit for contiguous operator (#1217)
Adding in check_stamped.py to tools/ (#1255)
Add compute_method for the experimental custom op (#1194)
remove eliminate_workspace pass (#1254)
Fix code block issue with .ipynb files. (#1263)
Update license files (#1248)
Fixing misspelled macro to enable MIOpen hidden find mode API (#1250)
Update lowering of Dot operator (#1247)
Update tf_parser to have add_common_op() for parse_relu6 (#1241)
Create allocate op and replace_allocate pass (#1183)
Instruction distance check fix (#1237)
Use env var for creds
Add vectorized reduce (#1202)
Prioritizing int8 over int8x4 when it is applicable (#1218)
Group code objects by kernel name in perf report summary (#1234)
Fix compilation on Debian bookworm/sid (#1229)
Fix dangling reference with gemm add fusion (#1233)
Update protobuf version (#1228)
Bump tensorflow from 2.6.4 to 2.7.2 in /examples/nlp/python_bert_squad (#1227)
Improve eliminate contiguous pass (#1223)
renamed to main from master (#1226)
Parallelize evaluations in propagate_constant (#1220)
Upgrade to cppcheck 2.8 and fix new issues found (#1225)
Used wrong path to download the bertsquad-10.onnx model (#1221)
Bump tensorflow from 2.5.3 to 2.6.4 in /examples/nlp/python_bert_squad (#1219)
Improve applicable batched gemms (#1214)
Remove std references in runtime compilation (#1186)
Fuse gemm add with pointwise fusions (#1213)
Fix onnx mean parsing for integral inputs (#1209)
Rename pointwise ops (#1145)
Improve matching with has_value when there are convert operators (#1212)
renamed variables for module from p to m (#1204)
Update install_prereqs.sh for individual use (#1197)
Prefuse layernorm for gpu (#1190)
Updated a path to the bert-squad onnx file after upstream changed path (#1201)
Expose add_literal in C and Python API (#1173)
Refactor vectorization and preloading for pointwise fusions (#1184)
upgrade docker images to ROCm 5.0.2 (#1133)
Add compile tests for gpu math functions (#1182)
Cppcheck fixes (#1195)
Extend lifetimes in C++ API (#1139)
Bumping version to support next ROCm release (#1192)

rocm-5.2.3

No changes from rocm-5.2.1

rocm-5.2.1

Enabled the devel migraphx package
Resolved a bug where migraphx could not run its own binary output files

rocm-5.2.0

Improvements include...

Add GatherND operator (#1089)
Add lane reduction (#1180)
Expose get_queue method for context in API (#1161)
ReverseSequence op (#1177)
Refactor Pooling and implement ONNX LpPool and GlobalLpPool (#1152)
Reduce with runtime compilation (#1150)
Half2 overloads (#1157)
Fix file download for resnet50 example (#1164)
Fix problem with incomplete types with older clang versions (#1174)
Fix out-of-bounds access when generate uses nonpacked tensors (#1160)
parallelize the ref implementation of the gemm operator (#1142)
scatter operator refactoring to include reduction (#1124)
fix a bug in create tensor_view with vec data type (#1155)
Fix comparisons in migraphx::value class (#1146)
Python Binding for the Manual Graph Buidling (#1143)

rocm-5.1.3

Identical to 5.1.1

rocm-5.1.1

Full Changelog: rocm-5.1.0...rocm-5.1.1

rocm-5.1.0

ABI version bumped to 2.1
Added ONNX Operators (HardSigmoid, SoftPlus, SoftSign, GreaterOrEqual, HardSwish, Mean)
Changed MessagePack file extensions to mxr
Updated Examples to include cppcheck
Pointwise operator improvements included Clip, auto-vectorization, and Type enforcement
Add assign_to method for C++ API
Support nonstandard shapes for the Squeeze and Unsqueeze Operator
Improved examples and Documentation
CI and Build improvements