mirror of
https://github.com/vosen/ZLUDA.git
synced 2025-07-25 13:16:23 +03:00
5cb0a9b8e8
Implements bar.red.and.pred and bar.red.or.pred, using the undocument __ockl_wgred functions. Doesn't yet add support for numbered barriers and threadcount, as these are not needed for llm.c.
61 lines
1.5 KiB
Plaintext
61 lines
1.5 KiB
Plaintext
.version 6.5
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
.visible .entry bar_red_and_pred(
|
|
.param .u64 input,
|
|
.param .u64 output
|
|
)
|
|
{
|
|
.reg .u64 out_addr;
|
|
.reg .u64 out_index;
|
|
.reg .u32 thread_id;
|
|
.reg .u32 thread_mod_2;
|
|
.reg .pred pred;
|
|
.reg .pred cond;
|
|
.reg .u32 result;
|
|
|
|
ld.param.u64 out_addr, [output];
|
|
|
|
mov.u32 thread_id, %tid.x;
|
|
rem.u32 thread_mod_2, thread_id, 2;
|
|
setp.eq.u32 cond, thread_mod_2, 0;
|
|
|
|
mov.u32 result, 0;
|
|
|
|
// Basic functionality
|
|
|
|
// result += AND(tid.x % 2 == 0) forall threads
|
|
bar.red.and.pred pred, 1, cond;
|
|
@pred add.u32 result, result, 1;
|
|
// result += OR(tid.x % 2 == 0) forall threads
|
|
bar.red.or.pred pred, 1, cond;
|
|
@pred add.u32 result, result, 1;
|
|
|
|
// result += AND(true) forall threads
|
|
setp.eq.u32 cond, 1, 1;
|
|
bar.red.and.pred pred, 1, cond;
|
|
@pred add.u32 result, result, 1;
|
|
// result += OR(false) forall threads
|
|
setp.eq.u32 cond, 1, 0;
|
|
bar.red.or.pred pred, 1, cond;
|
|
@pred add.u32 result, result, 1;
|
|
|
|
// Negated condition
|
|
// result += AND(!true) forall threads
|
|
setp.eq.u32 cond, 1, 1;
|
|
bar.red.and.pred pred, 1, !cond;
|
|
@pred add.u32 result, result, 1;
|
|
|
|
// Return result
|
|
|
|
cvt.u64.u32 out_index, thread_id;
|
|
mul.lo.u64 out_index, out_index, 4;
|
|
add.u64 out_addr, out_addr, out_index;
|
|
st.u32 [out_addr], result;
|
|
|
|
// result should be 2
|
|
|
|
ret;
|
|
}
|