Float math intrinsics: FEXP, FSQRT, FLOG, FABS, FNEG, FMAX, FMIN

[tetsuo-cpp]

Summary

Add floating-point math intrinsic words that lower to MLIR math dialect operations (and ultimately to NVVM/libdevice intrinsics on GPU).

Words to implement

Unary operations

Word	Stack effect	MLIR op	Description
FEXP	( f -- f )	math.exp	Exponential (e^x)
FSQRT	( f -- f )	math.sqrt	Square root
FLOG	( f -- f )	math.log	Natural logarithm
FABS	( f -- f )	math.absf	Absolute value
FNEG	( f -- f )	arith.negf	Negation

Binary operations

Word	Stack effect	MLIR op	Description
FMAX	( f f -- f )	arith.maximumf	Maximum of two floats
FMIN	( f f -- f )	arith.minimumf	Minimum of two floats

Motivation

• FEXP + FMAX: Required for softmax (exp(x - max)) — blocks flash attention, transformer kernels, and any probability-based computation.
• FSQRT: Required for score scaling (1/sqrt(d_k)) in attention, and for normalization (LayerNorm, RMSNorm).
• FLOG: Log-space softmax, cross-entropy loss.
• FABS: Numerical stability checks, absolute error computation.
• FNEG: Cleaner than 0.0 SWAP F-; needed for initializing accumulators to -inf patterns.
• FMAX/FMIN: Online reductions (running max/min across tiles), clamping values.

Implementation notes

• All follow the same pattern as existing float ops (F+, F-, etc.): bitcast i64↔f64 around the math op.
• Unary ops: pop one value, bitcast to f64, apply math op, bitcast back, push result.
• Binary ops (FMAX/FMIN): pop two values, bitcast both to f64, apply op, bitcast result back, push.
• MLIR's math dialect lowers to LLVM intrinsics, which NVVM maps to libdevice calls (e.g., __nv_exp, __nv_sqrt).
• FNEG uses arith.negf rather than math dialect.
• FMAX/FMIN use arith.maximumf/arith.minimumf (IEEE 754 semantics: propagate NaN).

Files to modify

1. include/warpforth/Dialect/Forth/ForthOps.td — Define new ops
2. lib/Translation/ForthToMLIR/ForthToMLIR.cpp — Parse words
3. lib/Conversion/ForthToMemRef/ForthToMemRef.cpp — Add conversion patterns
4. test/Translation/Forth/ — Parser tests
5. test/Conversion/ForthToMemRef/ — Conversion tests
6. test/Pipeline/ — End-to-end pipeline tests

Priority

High — blocks flash attention and most non-trivial GPU compute kernels.

Related

• Warp-level primitives: shuffle and reductions #10 — Warp-level primitives (needed together for performant reductions)
• Tensor core / MMA intrinsics #11 — Tensor core intrinsics (complementary GPU capability)