Add failing test for ftz

This commit is contained in:
Andrzej Janik 2025-01-28 01:46:34 +00:00
parent df5a96d935
commit a6e6454d8b
2 changed files with 29 additions and 0 deletions

View file

@ -0,0 +1,24 @@
.version 6.5
.target sm_30
.address_size 64
.visible .entry add_ftz(
.param .u64 input,
.param .u64 output
)
{
.reg .u64 in_addr;
.reg .u64 out_addr;
.reg .f32 temp<4>;
ld.param.u64 in_addr, [input];
ld.param.u64 out_addr, [output];
ld.f32 temp0, [in_addr];
ld.f32 temp1, [in_addr+4];
add.ftz.f32 temp2, temp0, temp1;
add.f32 temp3, temp0, temp1;
st.f32 [out_addr], temp2;
st.f32 [out_addr+4], temp3;
ret;
}

View file

@ -180,6 +180,11 @@ test_ptx!(activemask, [0u32], [1u32]);
test_ptx!(membar, [152731u32], [152731u32]);
test_ptx!(shared_unify_extern, [7681u64, 7682u64], [15363u64]);
test_ptx!(shared_unify_local, [16752u64, 714u64], [17466u64]);
test_ptx!(
add_ftz,
[f32::from_bits(0x800000), f32::from_bits(0x007FFFFF)],
[0x800000u32, 0xFFFFFF]
);
test_ptx!(assertfail);
test_ptx!(func_ptr);