[Zhonglin1231]

Fixes #3843.

int_to_int widened Signed -> Unsigned with arith.extui, so a runtime Int8(-1).to(Uint32) gave 255, while the constant path, numpy and C give 4294967295. This widens with arith.extsi, as the same-signedness branch right below already does, and keeps extui for 1-bit sources for the same reason that branch does.

        # Signed -> Unsigned: widen with sign extension, as a C cast does,
        # (uint32_t)(int8_t)-1 == 0xFFFFFFFF
        if dst_width > src_width:
            if src_width > 1:
                return arith.extsi(dst_mlir_type, a, loc=loc, ip=ip)
            return arith.extui(dst_mlir_type, a, loc=loc, ip=ip)

Narrowing and the other branches are unchanged.

Test

test/python/CuTeDSL/test_signed_to_unsigned.py converts [-1, -128, 0, 5, 127] in a kernel for all six signed to wider unsigned pairs and checks the bits against torch's signed widening. It needs a CUDA GPU and skips without one, like test_keep_ir.py.

On an A10G with nvidia-cutlass-dsl 4.8.1, all six pairs fail without the fix (Int8 -1 to Uint32 gives 255) and pass with it.

[fengxie]

Thanks for fixing.