[MetaXGPU] Split the fast-math and precise lowering paths, as CUDA does

[Lfan-ke]

Problem

tirx.enable_fast_math defaults to false, and IntrinInjecter only consults <target>.fastmath.FLowerIntrinsic when it is set (src/tirx/transform/lower_intrin.cc:57-60). The MACA rules register their fast-math variants under the plain maca.FLowerIntrinsic key instead, so the flag never reaches them: seven ops lower to the approximate device function unconditionally.

On a C500, with the default PassContext:

op	upstream	expected
exp	__expf	expf
exp10	__exp10f	exp10f
log	__logf	logf
log2	__log2f	log2f
log10	__log10f	log10f
sin	__sinf	sinf
cos	__cosf	cosf

A program that never asked for fast math gets it anyway, and setting the flag changes nothing. The CUDA rules do not have this problem: src/backend/cuda/codegen/intrin_rule_cuda.cc carries ten cuda.fastmath.FLowerIntrinsic registrations alongside the precise ones.

Changes

Each of these ops now registers twice, the way CUDA does: MACAFastMath on maca.fastmath.FLowerIntrinsic, MACAMath on maca.FLowerIntrinsic. The default becomes precise and tirx.enable_fast_math becomes the thing that selects the approximation.

pow gains a fast-math lowering it did not have before; __powf compiles on a C500.

tan keeps MACAFastMathTan on the fast path, which yields tanf at fp32 — the same function the precise path emits — so the registration only affects it at other widths. tanh stays precise on both paths, because MACA has no __tanhf and mxcc rejects the symbol.

Test Plan

tests/python/codegen/test_target_codegen_maca_fastmath.py on a MetaX C500 (MACA 3.5.3.20), over the eight ops that have both forms:

• under the default PassContext, the generated source calls expf and never __expf
• under PassContext(config={"tirx.enable_fast_math": True}), it calls __expf

Test Result

16 passed in 7.38s

Reverting intrin_rule_maca.cc and rebuilding turns eight of them red — the seven ops above lose their precise default, and pow loses the opt-in path:

8 failed, 8 passed in 7.27s
FAILED test_precise_math_is_the_default[exp]
FAILED test_precise_math_is_the_default[exp10]
FAILED test_precise_math_is_the_default[log]
FAILED test_precise_math_is_the_default[log2]
FAILED test_precise_math_is_the_default[log10]
FAILED test_precise_math_is_the_default[sin]
FAILED test_precise_math_is_the_default[cos]
FAILED test_fast_math_is_opt_in[pow]

[gemini-code-assist]

Code Review

This pull request separates the lowering paths for several mathematical intrinsics (such as exp, log, tan, cos, and sin) into standard and fast-math variants. The reviewer pointed out that using MACAFastMathTan for the fast-math path of tirx.tan breaks bfloat16 support because it lacks support for kDLBfloat, and suggested using MACAMath for both paths instead.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The MACAFastMathTan struct is functionally identical to MACAMath for float types (since it avoids __tanf and uses tanf instead), but it lacks support for other types like kDLBfloat (returning "" instead of "htan"). This breaks bfloat16 support for tan when fast-math is enabled.

We can simplify this registration and fix the bfloat16 bug by using MACAMath for both the default and fast-math paths. Additionally, MACAFastMathTan (lines 83-101) can be completely removed from this file as it is no longer needed.

Suggested change

	TVM_REGISTER_OP("tirx.tan")
	.set_attr<FLowerIntrinsic>("maca.fastmath.FLowerIntrinsic",
	DispatchPureExtern<MACAFastMathTan>)
	.set_attr<FLowerIntrinsic>("maca.FLowerIntrinsic", DispatchPureExtern<MACAMath>);
	TVM_REGISTER_OP("tirx.tan")
	.set_attr<FLowerIntrinsic>("maca.fastmath.FLowerIntrinsic",
	DispatchPureExtern<MACAMath>)
	.set_attr<FLowerIntrinsic>("maca.FLowerIntrinsic", DispatchPureExtern<MACAMath>);

[Lfan-ke]

@codex review please. read code at pr and repositories.