| Issue |
244593
|
| Summary |
[fuzzer-found][NVPTX] LowerINT_TO_FP converts int -> bf16 through f32 below sm_90: double rounding, (__bf16)16842753 is 0x4B80 instead of 0x4B81
|
| Labels |
miscompilation,
backend:NVPTX,
floating-point,
generated by fuzzer
|
| Assignees |
|
| Reporter |
YuanchengJiang
|
The following code:
```cpp
/* [NVPTX] (unsigned/int) -> __bf16 below sm_90 is double-rounded through float:
* NVPTXTargetLowering::LowerINT_TO_FP emits int -> f32 (cvt.rn.f32.u32 /
* cvt.rn.f32.s32, rounds when |x| > 2^24) followed by FP_ROUND f32 -> bf16
* (cvt.rn.bf16.f32, a second rounding). The result is not the correctly
* rounded bf16 value. sm_90 has a direct cvt.rn.bf16.u32 and is correct.
*
* x = 16842753 = 2^24 + 2^16 + 1 lies just above the midpoint 2^24 + 2^16 between
* the bf16 neighbours 2^24 (0x4B80) and 2^24 + 2^17 (0x4B81), so the exact
* result is 0x4B81. Through float, x first rounds (tie, to even) to exactly
* 2^24 + 2^16 (0x4B808000), the midpoint, which then ties to even: 0x4B80.
*
* Godbolt: compiler "x86-64 clang (trunk)" with flags:
* --target=nvptx64-nvidia-cuda -march=sm_80 -O2
* (sm_50..sm_89 are affected; sm_50 rounds f32 -> bf16 with integer ops).
* Godbolt cannot execute PTX: view the PTX of cvt_u32 / cvt_s32:
* buggy: cvt.rn.f32.u32 %r2, %r1 ; cvt.rn.bf16.f32 %rs1, %r2
* correct: one rounding, e.g. cvt.rn.f64.u32 (exact), a round-to-odd
* f64 -> f32 step, then cvt.rn.bf16.f32 (or sm_90's cvt.rn.bf16.u32).
* Run: on any sm_80 GPU (e.g. as a CUDA device function); the arithmetic is the same.
*
* EXPECTED (correct compiler): u32 = 19329 (0x4B81), s32 = 52097 (0xCB81)
* BUGGY (vanilla trunk): u32 = 19328 (0x4B80), s32 = 52096 (0xCB80)
* MISMATCH (exit 1)
*/
#include <stdio.h>
__attribute__((noinline)) __bf16 cvt_u32(unsigned x) { return (__bf16)x; }
__attribute__((noinline)) __bf16 cvt_s32(int x) { return (__bf16)x; }
volatile unsigned gu = 16842753u; /* 2^24 + 2^16 + 1 */
volatile int gs = -16842753;
int main(void) {
__bf16 a = cvt_u32(gu), b = cvt_s32(gs);
unsigned short ua, ub;
__builtin_memcpy(&ua, &a, 2);
__builtin_memcpy(&ub, &b, 2);
printf("u32 = %d (expected 19329)\n", ua);
printf("s32 = %d (expected 52097)\n", ub);
if (ua != 19329 || ub != 52097) {
printf("MISMATCH\n");
return 1;
}
return 0;
}
```
Resulted in this output:
```
PTX (--target=nvptx64-nvidia-cuda -march=sm_80 -O2):
cvt_u32(...)
ld.param.b32 %r1, [cvt_u32_param_0];
cvt.rn.f32.u32 %r2, %r1; // 1st rounding: 16842753 -> 0x4B808000
cvt.rn.bf16.f32 %rs1, %r2; // 2nd rounding: exact midpoint, ties to even -> 0x4B80
st.param.b16 [func_retval0], %rs1;
(cvt_s32: cvt.rn.f32.s32 + cvt.rn.bf16.f32; sm_50 uses cvt.rn.f32.u32 + an integer RNE sequence;
sm_90 emits a single cvt.rn.bf16.u32 and is correct.)
input | expected (correctly rounded) | actual
16842753 | 19329 (0x4B81) | 19328 (0x4B80) <-- MISMATCH
-16842753 | 52097 (0xCB81) | 52096 (0xCB80) <-- MISMATCH
(Not executed: no NVIDIA GPU here. PTX cvt.rn is IEEE RNE, so the values follow from the two
conversions; checked with exact arithmetic: RNE8(x) = 0x4B81, RNE8(RNE24(x)) = 0x4B80.)
```
Trunk: https://godbolt.org/z/jo78svTPj (x86-64 clang trunk, --target=nvptx64-nvidia-cuda -march=sm_80 -O2; PTX only)
Root cause: Below sm_90 there is no direct integer -> bf16 cvt, and NVPTXTargetLowering::LowerINT_TO_FP (llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp) lowers it as [su]int_to_fp to f32 followed by FP_ROUND f32 -> bf16. int -> f32 rounds for |x| > 2^24, and the second rounding to 8 significand bits then double-rounds: 2^24 + 2^16 + 1 first ties to exactly the bf16 midpoint and then ties to even downward. sm_50..sm_89 are affected for i32 and i64 sources; this is a separate copy from the generic/X86 path in #234116.
Suggested fix: For sources of <= 32 bits convert to f64 first (exact), then let FP_ROUND f64 -> bf16 expand with a round-to-odd f64 -> f32 step before cvt.rn.bf16.f32; i64 sources need the same round-to-odd treatment or a libcall. Validated locally for i32 (a patched tree carries this fix and the generated PTX computes 0x4B81 / 0xCB81, checked by modeling the PTX since no GPU is available).
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs