[SparseTIR] Add SparseBufferLoad/SparseBufferStore #5

MasterJH5574 · 2021-10-23T02:30:07Z

This PR mainly adds SparseBufferLoad/SparseBufferStore, similar to BufferLoad/BufferStore in TIR.

Besides, this PR also does the following:

remove the namespace sparse;
add fields dtype and name to SparseBuffer;
remove field ndim from SparseBuffer, since it can be inferred from axes by quering the array length.

MasterJH5574 · 2021-10-23T02:35:05Z

Several trivial things may still be in the to-do list:

add ReprPrinter & TVM script printer support for the new nodes introduced in recent PRs;
add TVM script printer support for SparseBufferLoad/Store;
add TVM script parser support for both the new nodes and SpareBufferLoad/Store.

I'll try to do them soon. May need some from Siyuan and Bohan on the parser/printer stuff, I don't know. Anyway, the parser/printer isn't a huge block on our progress, until we're going to write lots of unittests.

yzh119 · 2021-10-23T05:19:58Z

include/tvm/tir/sparse.h

  /* Buffer corresponding to flattened value */
  Buffer data;
+  /* Buffer Name */


I suggest adding a function ndim() that returns the length of axes.

Makes sense! Will do.

* Add dtype for SparseBuffer * Add name for SparseBuffer. Remove `ndim` * Remove namespace sparse * Add SparseBufferLoad/Store * Add method `ndim()`

* Format and Buffer data structure (#1) * [SparseTIR] Constructors and Python Interface for `Axis` and `SparseBuffer` (#2) * add methods for Object * axis constructors * methods for SparseBuffer * put into registry * python interface * [CherryPick][Intrinsic] lower_bound and upper_bound for binary search in Sparse TIR. (apache#483) (#4) * upd * upd * fix * upd * upd * upd * upd * upd * fix * upd * upd * upd * upd * upd * upd * upd * codegen-rule * upd * upd * test * upd * fix * two arguments Co-authored-by: Zihao Ye <expye@outlook.com> * Fix AxisTree (#3) * fix axis tree * upd * Format and Buffer data structure (#1) * [SparseTIR] Constructors and Python Interface for `Axis` and `SparseBuffer` (#2) * add methods for Object * axis constructors * methods for SparseBuffer * put into registry * python interface * fix axis tree * upd * Format and Buffer data structure (#1) * [SparseTIR] Constructors and Python Interface for `Axis` and `SparseBuffer` (#2) * add methods for Object * axis constructors * methods for SparseBuffer * put into registry * python interface * [CherryPick][Intrinsic] lower_bound and upper_bound for binary search in Sparse TIR. (apache#483) (#4) * upd * upd * fix * upd * upd * upd * upd * upd * fix * upd * upd * upd * upd * upd * upd * upd * codegen-rule * upd * upd * test * upd * fix * two arguments Co-authored-by: Zihao Ye <expye@outlook.com> * Fix AxisTree (#3) * fix axis tree * upd * [SparseTIR] Add SparseBufferLoad/SparseBufferStore (#5) * Add dtype for SparseBuffer * Add name for SparseBuffer. Remove `ndim` * Remove namespace sparse * Add SparseBufferLoad/Store * Add method `ndim()` * Format and Buffer data structure (#1) * [SparseTIR] Constructors and Python Interface for `Axis` and `SparseBuffer` (#2) * add methods for Object * axis constructors * methods for SparseBuffer * put into registry * python interface * [CherryPick][Intrinsic] lower_bound and upper_bound for binary search in Sparse TIR. (apache#483) (#4) * upd * upd * fix * upd * upd * upd * upd * upd * fix * upd * upd * upd * upd * upd * upd * upd * codegen-rule * upd * upd * test * upd * fix * two arguments Co-authored-by: Zihao Ye <expye@outlook.com> * Fix AxisTree (#3) * fix axis tree * upd * [SparseTIR] Add SparseBufferLoad/SparseBufferStore (#5) * Add dtype for SparseBuffer * Add name for SparseBuffer. Remove `ndim` * Remove namespace sparse * Add SparseBufferLoad/Store * Add method `ndim()` * [SparseTIR] Introduce SpIterVar (#6) * [SparseTIR] Introduce SpIterVar * Add conversion to PrimExpr * [BugFix] Fix binary search & SpIterVar (#7) * [BugFix] Add field `is_reduction` for SpIterVar (#9) * [BugFix] Add field `is_reduction` for SpIterVar * Formatting * upd * upd Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com>