MindAKG 2.1.0 Release Notes

Bug Fixes

• Fixed the inconsistencies issue during build to support PolyTops on Ascend. !908:fix ci and build issues (Ascend)
• Fixed the adaptability issue to CANN 6.4. !911:akg适配CANN-6.4 (Ascend)
• Fixed the adaptability issue to MatMul on MSLite Ascend. !914:add mindspore lite sigma matmul repo on ascend (Ascend)

Contributors

Thanks goes to these wonderful people:

liuchao, huoxinyou, yangsijia, Harenome Ranaivoarivony-Razanajato

MindAKG 2.0.0 Release Notes

Bug Fixes

• Fixed the dependency issue on PolyTOPS during building. !873:fix polytops dependence (All)
• Fixed the data type issue of int64 in GPU code generation. !868:fixbugs: wrong int64 cuda codegen & tuning axis names from memroy constraints (GPU)

Contributors

Thanks goes to these wonderful people:

liuchao, yiyanzhi, hujiahui8, xuhui, yangshuo, nelson.lossing

MindAKG 1.9.0 Release Notes

Major Features and Improvements

• Support the auto-generation of cpu-pool2d(avg/max/global). !763:AKG supports cpu-pool2d(avg/max/gloobal) (CPU)
• Integerate Symbolic Tiling algorithm for AKG. !764:[Symbolic Tiling] advanced classes to be used at symbolic tiling algorithm (Ascend)
• AutoTiling optimization on GPU Backend. !784:auto-tiling opt1: increase proposed block sizes for extremely large shape !785:auto-tiling opt2: update condition for divisible tile size !783:auto-tiling opt3: fix fuse axis size limit (GPU)
• optimization for Transpose Operator. !811:adapt transpose operator for cpu (CPU)

Bug fixes

• Fix the bug that analysis error of read-write relationship in multi-filter.!778:fix_bug: analysis error of read-write relationship in multi-filter (All)
• Fix the bug when batch_matmul recognizes the axis. !807:fix bug: batch_matmul operator recognizes the axis error (CPU)
• Fix the bug when gcc version is less then 7.3. !817:fix bug: raise error when gcc version is not meet (All)

Contributors

Thanks goes to these wonderful people:

yangsijia, polyhedral, zhangrenwei, yiyanzhi, hujiahui8, zhengzuohe, zhangzhaochuang, xuhui, liuchao, xiaruijie, DeshiChen, nelson.lossing, zhenzhang, chenlei_autodiff, wYann, hanhuifeng, gaoxiong, zichun_ye, chengbin, wangrao124, huangmengxi, harenome, huoxinyou, yanglinfeng, Etienne

Major Features and Improvements

• Integrate the CPU compiling process for CustomOp. !612:Update custom op compiling process (CPU)
• Optimize high-dimensional CSR OP. !664:Optimize high-dimensional CSR OP (GPU)
• Update the restart process of Ascend backend. !701:Optimize the restart process of ascend backend (Ascend)
• Support the generation of CSRMM operator. !709:AKG支持SPMM算子 (GPU)
• Integrate usage of intrinsic directives for the new Scheduler. !715:[CustomOp] integrate usage of intrinsic directives for the scheduler (All backends)

Bug fixes

• Fix the bug of reshape elimination for the fused operators. !707:fix reshape bug (All)
• Fix the bug of to_three_address pass. !728:FIX to_three_address pass (Ascend)
• Fix the bug of Operator types matching rules. !730:fix wrong template type. (All)

Contributors

Thanks goes to these wonderful people:

yangsijia, polyhedral, zhangrenwei, yiyanzhi, hujiahui8, zhengzuohe, zhangzhaochuang, xuhui, liuchao, xiaruijie, DeshiChen, nelson.lossing, zhenzhang,

chenlei_autodiff, wYann, hanhuifeng, gaoxiong, zichun_ye, chengbin, wangrao124, huangmengxi, harenome, huoxinyou, yanglinfeng, Etienne

Release 1.7.0

Major Features and Improvements

• Support a new loop intrinsic for MindSpore Hybrid DSL. !560:Support new loop intrin: grid (ALL)
• Update the shared and local promotion strategy on GPU backend. !556:Update the shared and local promotion strategy (GPU)
• Refactoring the reschedule pass after pluto scheduling pass on Ascend backend.!512:Refactoring the reschedule pass (Ascend)

Bug fixes

• Fix the bug of gemm operator due to repeated tiling. Especially for the conv operator, the h and w axes need to be mapped to BlockIdx.y. !582:fix the bug of gemm operator tiling error (GPU)
• Fix the bug of paralleize axis y in reduce of Y_DIRECTION with small x. !559:Fix a bug of fused reduce_y (CPU)

Contributors

Thanks goes to these wonderful people:

yangsijia, polyhedral, zhangrenwei, yiyanzhi, hujiahui8, zhengzuohe, zhangzhaochuang, xuhui, liuchao, xiaruijie, DeshiChen, nelson.lossing, zhenzhang,

chenlei_autodiff, lingyunli63, wYann, hanhuifeng, gaoxiong, zichun_ye, chengbin, wangrao124, huangmengxi, harenome, lear, huoxinyou, yanglinfeng, Etienne, Trump

Contributions of any kind are welcome!

Release 1.6.0

Major Features and Improvements

• [STABLE] CPU backend support: AKG now supports automatic generation of high-performance code for CPUs with different architectures. !413:Add CPU backend support (CPU)
• [STABLE] CSR operators support: Develops csr operators(csrmv/csr mul/csr reduce_sum)，and provides optimizing strategy to handle csr operators with dynamic upper bound. !407:Support csr operators codegen (GPU)
• [STABLE] Scheduler Optimizing: Outermost band node now will be treated as multiple filter nodes if there is no dependency relationship between multiple statements. !460:Adapt to the case where the outermost band is a set or sequence node (GPU)
• [STABLE] Replace gmp with imath on isl.!455:Replace gmp with imath on isl (ALL)
• [STABLE] Apply Autodiff in Custom Op. !464:Apply Autodiff in Custom Op (ALL)

Bug fixes

• fix bugs in cpu tuning process:add "need_warm_up" attrs to avoid too many warm-up on cpu profiling.!495:fix bug: get a wrong op_template & add warm_up mark (CPU)
• fix auto-tiling output mem-flow bug:modify the GEMM's auto-tiling strategy to match the actual target constraint.!504:fix auto-tiling output mem-flow bug (ASCEND)
• fix muilt output bug in emitter: when emit multi-output, should not convert real inputs.!506:fix muilt output bug in emitter (ALL)

Contributors

Thanks goes to these wonderful people:

yangsijia, polyhedral, zhangrenwei, yiyanzhi, hujiahui8, zhengzuohe, zhangzhaochuang, xuhui, liuchao, xiaruijie, DeshiChen, nelson.lossing, zhenzhang, chenlei_autodiff, lingyunli63, wYann, hanhuifeng, gaoxiong, zichun_ye, chengbin, wangrao124, huangmengxi, harenome, lear, huoxinyou, yanglinfeng, Etienne, Trump

Contributions of any kind are welcome!

Release 1.5.0

Major Features and Improvements

• [STABLE] New operators developing: Tensor of Tensor operators(Gather/GatherNd/TensorScatterAdd/UnsortedSegmentSum), which can be used to support GNN networks.!323:support gnn(GPU)
• [STABLE] Add a topi op of UserDefine Op in AKG, which can be compiled from func_source_string or op_imply_path.!319:Add a topi op of userdefine op(GPU)
• [STABLE] Lower interface Refactor: add StageLower for stage lower case.!310:Refactor Lower interface.(GPU)
• [STABLE] The profiling suit for new runtime process.!306:[Ascend]The profiling suit for new runtime proccess(ASCEND)

Bug fixes

• Fixed memory promotion bug: sort the clusters before merging.!338:sort the footprintclusters before merging (ASCEND)
• Fixed irregular-reduce bug: replace shfl.down with shared memory reducetion.!332:fix bugs: shfl.down reductions (GPU)
• Fixed foldDimension bug: build wrong axis relation of relation.!302:FoldDimension bugfix of reshape relation (GPU)

Contributors

Thanks goes to these wonderful people:

yangsijia, xxxxxxw, polyhedral, zhangrenwei, yiyanzhi, xixixian, hujiahui8, zhengzuohe, lishanni, zhangzhaochuang, xuhui, liu
chao, gengzhen, xiaruijie,chenlei_autodiff, lingyunli63, wYann, lvwenyuan, peiwenfang, hanhuifeng, gaoxiong, chengyun
Contributions of any kind are welcome!

Major Features and Improvements

• [STABLE] Support optimizing GEMM && Conv by using polyhedral + Tensorcore, as well as providing an akg::fragment_add/sub/mul/div library for GEMM op fusions. !156:Enable TensorCore for GEMM and CONV. (GPU)
• [STABLE] Optimize layout related operators(Transpose && pad/unpad) by adjusting the autotiling strategy and solving bank conflict for these ops. !152:Optimize the tiling parameters&bank conflict for Transpose OP (GPU)
• [STABLE] Add Kahan algorithm for reducetion operators. !107:kahan add for reduction ops (GPU)
• [STABLE] Support transdata + matmul prefusion pattern in akg. !103:add transdata + matmul prefusion dsl and verification implementation (ASCEND)

Bug fixes

• Fixed stitch_fusion bug when store is shared but load is not shared. !109:fix stitch bug when store is shared but load is not shared (GPU)
• Fixed reshape bug: when add reshape, should update tensor shape. !140:fix reshape bug: when add reshape, should update tensor shape (GPU)
• Fixed autofuse bug: when set autofuse config, should record broadcast. !111:fix autofuse bug, when set autofuse config, should record broadcast (GPU)

Contributors

Thanks goes to these wonderful people:

yangsijia, xxxxxxw, polyhedral, zhangrenwei, yiyanzhi, xixixian, hujiahui8, zhengzuohe, lishanni, zhangzhaochuang, xuhui, liu
chao, gengzhen, xiaruijie,chenlei_autodiff, lingyunli63, wYann, lvwenyuan, peiwenfang, hanhuifeng, gaoxiong, chengyun
Contributions of any kind are welcome!

Release 1.2.0

Major Features and Improvements

• [STABLE] Rebuild the AKG repository for providing a new way to support ascend backend by linking a static library contained all the ascend passes. (Ascend)
• [STABLE] Optimize the reduction add operation in ascend backend. (Ascend)
• [STABLE] Add support for tuning elemwise&&reduction operators. (GPU)

Bug fixes

• Fixed a problem that data prefetch cannot be enabled by attributes in DSL.
• Fixed bugs of autotiling algorithms (tiling too small, cannot adapted matmul+bias, etc.) in Ascend platform.
• Fixed local memory promotion for large thread (2980!)
• Fixed reduce binding dimension issue on gpu platform (ff38!)

Contributors

Thanks goes to these wonderful people:

yangsijia, xxxxxxw, polyhedral, zhangrenwei, yiyanzhi, xixixian, hujiahui8, zhengzuohe, lishanni, zhangzhaochuang, xuhui, liuchao, gengzhen, xiaruijie,
chenlei_autodiff, lingyunli63, wYann, lvwenyuan, peiwenfang, hanhuifeng, gaoxiong, chengyun
Contributions of any kind are welcome!