Add PTOAS support for TAXPY, THISTOGRAM, TGET_SCALE_ADDR, TROWARGMAX and TROWARGMIN by HecreReed · Pull Request #461 · hw-native-sys/PTOAS

HecreReed · 2026-04-08T08:56:29Z

Summary

add PTO IR ops and verifiers for pto.taxpy, pto.thistogram, pto.tget_scale_addr, pto.trowargmax, and pto.trowargmin
add EmitC lowering for the 5 new ops
update ptobc opcode mapping and add basic smoke/verifier coverage

Validation

ninja -C build ptoas
manual smoke with build/tools/ptoas/ptoas on the 7 new test/basic/*.pto cases
ctest --output-on-failure -R ptobc_opcode_coverage_check

gemini-code-assist

Code Review

This pull request adds several new operations to the PTO dialect, including THistogramOp, TGetScaleAddrOp, TAxpyOp, TRowArgMaxOp, and TRowArgMinOp, along with their associated TableGen definitions, IR verifiers, memory effects, and EmitC lowering patterns. Feedback focuses on improving the IR verifiers: specifically, the TAxpyOp verifier should account for bf16 to f32 widening on the A5 architecture, and both TRowArgMaxOp and TRowArgMinOp should be extended to support bf16 source elements on A5 for consistency with other operations.

gemini-code-assist · 2026-04-08T08:59:01Z

lib/PTO/IR/PTO.cpp

+    bool widenF16ToF32 = srcElem.isF16() && dstElem.isF32();
+    if (!(sameType || widenF16ToF32))
+      return emitOpError(
+          "expects dst/src element types to match, or dst=f32 and src=f16");


The verifyA5 lambda for TAxpyOp does not account for bf16 to f32 widening. Since bf16 is explicitly allowed for both src and dst in the subsequent checks (lines 2641-2644), it should also be allowed to widen to f32. The error message should also be updated to reflect this.

Suggested change

bool widenF16ToF32 = srcElem.isF16() && dstElem.isF32();

if (!(sameType || widenF16ToF32))

return emitOpError(

"expects dst/src element types to match, or dst=f32 and src=f16");

bool widenF16ToF32 = srcElem.isF16() && dstElem.isF32();

bool widenBF16ToF32 = srcElem.isBF16() && dstElem.isF32();

if (!(sameType || widenF16ToF32 || widenBF16ToF32))

return emitOpError(

"expects dst/src element types to match, or dst=f32 and src=f16/bf16");

gemini-code-assist · 2026-04-08T08:59:01Z

lib/PTO/IR/PTO.cpp

+    auto srcElem = getElemTy(srcTy).dyn_cast<mlir::FloatType>();
+    if (!srcElem || (!srcElem.isF16() && !srcElem.isF32()))
+      return emitOpError("expects src element type to be f16 or f32");


The TRowArgMaxOp verifier currently only allows f16 and f32 for the source element type. Given that TAxpyOp supports bf16 on A5, it is likely that TRowArgMaxOp should also support bf16 on A5. Consider updating the verifier to allow bf16 when the architecture is A5 by splitting the verifyByArch lambda or adding an architecture check.

gemini-code-assist · 2026-04-08T08:59:01Z

lib/PTO/IR/PTO.cpp

+    auto srcElem = getElemTy(srcTy).dyn_cast<mlir::FloatType>();
+    if (!srcElem || (!srcElem.isF16() && !srcElem.isF32()))
+      return emitOpError("expects src element type to be f16 or f32");


Similar to TRowArgMaxOp, the TRowArgMinOp verifier should likely support bf16 source elements on A5 architecture.

HecreReed · 2026-04-08T09:29:33Z

/run a3

reedhecre · 2026-04-08T10:17:28Z

A3 板测失败

触发方式：manual
源码提交：e83f4e4bbe3e
结果汇总：OK 179 / FAIL 3 / SKIP 1
日志：/home/zhongxuan/ptoas-board-monitor/runtime/logs/20260408_175504_manual_pr461.log
手动指令：/run a3
触发人：HecreReed
触发评论：Add PTOAS support for TAXPY, THISTOGRAM, TGET_SCALE_ADDR, TROWARGMAX and TROWARGMIN #461 (comment)
失败阶段：board-validation / exit=1

失败用例

rsqrt (run, exit=2)
rems (run, exit=2)
rem (run, exit=2)

reedhecre · 2026-04-08T10:17:30Z

A3 板测失败详情：PR #461

rsqrt

stage=run info=exit=2

[ERROR] Mismatch: golden_v2.bin vs v2.bin, max diff=1.618960976600647 at idx=417 (golden=1.742221474647522, out=3.361182451248169, dtype=float32)
[ERROR] compare failed
[2026-04-08 18:04:45] ERROR: testcase failed (exit 2): rsqrt

rems

stage=run info=exit=2

/home/zhongxuan/ptoas-board-monitor/runtime/runs/20260408_175504_manual_pr461/npu_validation/Rems/rems/rems_kernel.cpp:93:3: error: no matching function for call to 'TREMS'
  TREMS(v13, v12, v4);
  ^~~~~
/home/zhongxuan/ptoas-board-monitor/runtime/runs/20260408_175504_manual_pr461/payload/pto-isa/include/pto/common/pto_instr.hpp:1446:22: note: candidate function template not viable: requires at least 4 arguments, but 3 were provided
PTO_INST RecordEvent TREMS(TileDataDst &dst, TileDataSrc &src, typename TileDataSrc::DType scalar, TileDataTmp &tmp,
                     ^
1 error generated.
gmake[2]: *** [CMakeFiles/rems_kernel.dir/build.make:76: CMakeFiles/rems_kernel.dir/rems_kernel.cpp.o] Error 1
gmake[2]: *** Waiting for unfinished jobs....
gmake[1]: *** [CMakeFiles/Makefile2:85: CMakeFiles/rems_kernel.dir/all] Error 2
gmake: *** [Makefile:91: all] Error 2
[2026-04-08 18:07:31] ERROR: testcase failed (exit 2): rems

rem

stage=run info=exit=2

/home/zhongxuan/ptoas-board-monitor/runtime/runs/20260408_175504_manual_pr461/npu_validation/Rem/rem/rem_kernel.cpp:99:3: error: no matching function for call to 'TREM'
  TREM(v18, v16, v17);
  ^~~~
/home/zhongxuan/ptoas-board-monitor/runtime/runs/20260408_175504_manual_pr461/payload/pto-isa/include/pto/common/pto_instr.hpp:1679:22: note: candidate function template not viable: requires at least 4 arguments, but 3 were provided
PTO_INST RecordEvent TREM(TileDataDst &dst, TileDataSrc0 &src0, TileDataSrc1 &src1, TileDataTmp &tmp,
                     ^
1 error generated.
gmake[2]: *** [CMakeFiles/rem_kernel.dir/build.make:76: CMakeFiles/rem_kernel.dir/rem_kernel.cpp.o] Error 1
gmake[2]: *** Waiting for unfinished jobs....
gmake[1]: *** [CMakeFiles/Makefile2:85: CMakeFiles/rem_kernel.dir/all] Error 2
gmake: *** [Makefile:91: all] Error 2
[2026-04-08 18:17:02] ERROR: testcase failed (exit 2): rem

gemini-code-assist · 2026-04-09T01:02:30Z

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: db7b6d8d1a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-04-09T01:07:43Z

lib/PTO/IR/PTO.cpp

+  PTO_ADD_READ(getSrcMutable());
+  PTO_ADD_READ(getScalarMutable());
+  PTO_ADD_WRITE(getDstMutable());


Mark TAXPY destination as read in memory effects

pto.taxpy is defined as an in-place accumulate (dst += src * scalar), so it depends on the previous contents of dst. In TAxpyOp::getEffects the destination is marked only as Write, which can let transformations that consume MemoryEffectsOpInterface treat earlier updates to the same tile as dead/overwritable. This can change results when multiple taxpy operations target the same destination.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-04-09T01:07:43Z

lib/PTO/IR/PTO.cpp

+  PTO_ADD_READ(getSrcMutable());
+  PTO_ADD_READ(getIdxMutable());
+  PTO_ADD_WRITE(getDstMutable());


Mark THISTOGRAM destination as read in memory effects

pto.thistogram performs histogram accumulation into dst, so it semantically reads existing bin values before writing updated ones. THistogramOp::getEffects currently models dst as write-only, which can permit incorrect scheduling or dead-store style rewrites for sequences that repeatedly update the same histogram tile.

Useful? React with 👍 / 👎.

Add missing PTOAS support for TAXPY and other tile ops

e9a98b6

gemini-code-assist bot reviewed Apr 8, 2026

View reviewed changes

Update PTO IR manual for new tile ops

db7b6d8

HecreReed marked this pull request as ready for review April 9, 2026 01:02

chatgpt-codex-connector bot reviewed Apr 9, 2026

View reviewed changes

HecreReed added 2 commits April 9, 2026 17:44

Merge main into missing tile op support branch

48fa561

fix(emitc): lower mgather and mscatter to isa intrinsics

349e547

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add PTOAS support for TAXPY, THISTOGRAM, TGET_SCALE_ADDR, TROWARGMAX and TROWARGMIN#461

Add PTOAS support for TAXPY, THISTOGRAM, TGET_SCALE_ADDR, TROWARGMAX and TROWARGMIN#461
HecreReed wants to merge 4 commits intohw-native-sys:mainfrom
HecreReed:codex/add-missing-other-ops

HecreReed commented Apr 8, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Apr 8, 2026

Uh oh!

gemini-code-assist bot Apr 8, 2026

Uh oh!

gemini-code-assist bot Apr 8, 2026

Uh oh!

HecreReed commented Apr 8, 2026

Uh oh!

reedhecre commented Apr 8, 2026

Uh oh!

reedhecre commented Apr 8, 2026

Uh oh!

gemini-code-assist bot commented Apr 9, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Apr 9, 2026

Uh oh!

chatgpt-codex-connector bot Apr 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

HecreReed commented Apr 8, 2026

Summary

Validation

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Apr 8, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Apr 8, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Apr 8, 2026

Choose a reason for hiding this comment

Uh oh!

HecreReed commented Apr 8, 2026

Uh oh!

reedhecre commented Apr 8, 2026

A3 板测失败

失败用例

Uh oh!

reedhecre commented Apr 8, 2026

A3 板测失败详情：PR #461

Uh oh!

gemini-code-assist bot commented Apr 9, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants