derek.holloway
7502bf64a4
6.3 KiB
StinkyTofu Configuration via GlobalParameters
Control StinkyTofu behavior through Tensile's GlobalParameters system -- either via --global-parameters on the CLI or in the YAML configuration file.
Quick Reference
| Parameter | Values | Default | Purpose |
|---|---|---|---|
StinkyTofuOptLevel |
0, 1, 2, 3 |
0 |
Pipeline optimization level |
StinkyTofuDebugLevel |
0, 1, 2 |
0 |
Diagnostic output verbosity |
StinkyTofuPrintBeforePass |
comma-separated pass names | "" |
Print IR before specific passes |
StinkyTofuPrintAfterPass |
comma-separated pass names | "" |
Print IR after specific passes |
StinkyTofuDebugPass |
comma-separated pass names | "" |
PASS_DEBUG logging |
StinkyTofuVerifyEach |
0, 1 |
0 |
Verify StinkyTofu ASM IR after every pass |
StinkyTofuEnableRemarks |
0, 1 |
0 |
Emit optimization remarks to stderr |
StinkyTofuOptLevel
Maps directly to the pipeline's OptLevel enum (O0--O3). Set to None in YAML to disable the StinkyTofu feature entirely.
StinkyTofuDebugLevel
Debug output is per-PM -- the outer PassManager and each ScopeAdaptor's inner PassManager are configured independently.
| Level | Output | Destination |
|---|---|---|
0 (default) |
Silent | -- |
1 |
Pass names in execution order | stdout |
2 |
Initial IR before passes + IR after each pass | per-PM files (see below) |
DebugLevel 2 file naming
Each PM writes to its own file, named by scope:
| Scope | File prefix |
|---|---|
| Outer PM | kernel-OuterPM |
| Single-region adapter | <groupName> (e.g. loopWithPrefetch) |
| Multi-region adapter | <group1>+<group2> (e.g. loopWithPrefetch+noLoadLoopBody) |
| Whole-kernel adapter | wholeKernel |
Example files for a gfx1250 kernel:
kernel-OuterPM-before_passes.txt
kernel-OuterPM-after_passes.txt
loopWithPrefetch-before_passes.txt
loopWithPrefetch-after_passes.txt
noLoadLoopBody-before_passes.txt
noLoadLoopBody-after_passes.txt
loopWithPrefetch+noLoadLoopBody-before_passes.txt
loopWithPrefetch+noLoadLoopBody-after_passes.txt
StinkyTofuPrintBeforePass / StinkyTofuPrintAfterPass
Comma-separated list of pass names (case-sensitive) to selectively dump IR before or after. Output goes to per-PM files using the same naming convention as DebugLevel 2.
These work independently of StinkyTofuDebugLevel. When both are set, they compose without duplication. Unmatched pass names are silently ignored.
Pass names target the actual pass (e.g. CFGBuilderPass, StinkyDAGSchedulerPass), not the ScopeAdaptor wrapper. The pass runs inside an inner PM, so the IR dump shows the extracted region, not the full kernel.
StinkyTofuDebugPass
Comma-separated list of pass names (case-sensitive). Enables PASS_DEBUG logging -- prints pass-internal debug information (e.g. DAG graphs, scheduling decisions) to stderr for the listed passes.
Unmatched pass names are silently ignored.
Global scope: DebugPass is a global setting that applies to all PMs regardless of nesting. No per-PM configuration needed.
StinkyTofuEnableRemarks
Enables LLVM-style optimization remarks on stderr. Unlike PASS_DEBUG (which targets compiler developers), remarks are designed for kernel developers who want to understand the quality of generated code without reading raw IR.
Currently the only remark-emitting pass is LoopRegionRemarkPass, which reports per-loop:
| Metric | What it means |
|---|---|
| Region count | Number of contiguous instruction stretches separated by non-movable side effects. Fewer regions = more scheduling freedom = better performance. |
| Boundary causes | What splits each region: [wait] (s_waitcnt), [store] (global memory store), [barrier], [branch], or [untokenized_mem] (ds_read/ds_write/tensor_load without MemTokenData -- these could potentially be made movable by adding token support). |
| s_nop count | s_nop instructions inserted by InsertDelayAluPass, indicating scheduling couldn't hide latency. Reports count and total wasted cycles. |
| Branch count | Branches in the loop body. Too many branches hurt the hardware instruction prefetcher. |
Example output:
analysis: LoopRegionRemark: loop 'loopWithPrefetch' summary: 142 insts, 4 regions, 2 s_nop (5 wasted cycles), 3 branches
analysis: LoopRegionRemark: BB 'loopWithPrefetch_body': 4 regions, 3 boundaries:
analysis: LoopRegionRemark: [wait] s_waitcnt_vscnt (inst #12)
analysis: LoopRegionRemark: [barrier] s_barrier (inst #25)
analysis: LoopRegionRemark: [wait] s_waitcnt_lgkmcnt (inst #38)
analysis: LoopRegionRemark: BB 'loopWithPrefetch_body': 2 s_nop instructions (5 wasted cycles)
What to do with the output:
- High region count: Look at the
[untokenized_mem]boundaries -- these are memory operations that lack MemTokenData tokens. Adding token support for these operations would allow the DAG scheduler to move them freely, merging regions. - s_nop waste: Indicates the scheduler couldn't hide instruction latency. Consider reordering independent work to fill the gaps.
- Many branches: Consider loop restructuring to reduce branch pressure on the instruction prefetcher.
Usage
Via CLI
String values require both outer single quotes and inner double quotes ('Key="value"') so that Python's eval() parses them correctly. This applies even for a single pass name.
Tensile.sh config.yaml output/ --global-parameters StinkyTofuOptLevel=3 StinkyTofuDebugLevel=2
Tensile.sh config.yaml output/ --global-parameters StinkyTofuOptLevel=3 'StinkyTofuPrintAfterPass="CFG Builder, StinkyDAGSchedulerPass"'
Tensile.sh config.yaml output/ --global-parameters StinkyTofuOptLevel=3 'StinkyTofuDebugPass="StinkyDAGSchedulerPass"'
Tensile.sh config.yaml output/ --global-parameters StinkyTofuOptLevel=3 StinkyTofuEnableRemarks=1
Via YAML
GlobalParameters:
StinkyTofuOptLevel: 3
StinkyTofuDebugLevel: 2
StinkyTofuPrintAfterPass: "CFG Builder, StinkyDAGSchedulerPass"
StinkyTofuDebugPass: "StinkyDAGSchedulerPass"
StinkyTofuEnableRemarks: 1
Scope
These parameters apply to the Tensile/KernelWriter integration path. For stinkytofu-opt, use the equivalent CLI flags (e.g. --remarks). See stinkytofu-opt README.