Using static for LUTs #354
Conversation
|
This seems to improve timing as well (unless I'm behind on the latest numbers) |
| @@ -7,7 +7,7 @@ | |||
|
|
|||
| template<int kNBitsBuffer> | |||
| static const ap_uint<(1 << (2 * kNBitsBuffer))> nearFullUnit() { | |||
| ap_uint<(1 << (2 * kNBitsBuffer))> lut; | |||
| static ap_uint<(1 << (2 * kNBitsBuffer))> lut; | |||
There was a problem hiding this comment.
@bryates For this to work, I think you also need logic to prevent the reinitialization of the LUT. For example:
template<int kNBitsBuffer>
static const ap_uint<(1 << (2 * kNBitsBuffer))> nearFullUnit() {
static ap_uint<(1 << (2 * kNBitsBuffer))> lut;
static bool initialized = false;
if (!initialized) {
initialized = true;
for(int i = 0; i < (1 << (2 * kNBitsBuffer)); ++i) {
#pragma HLS unroll
ap_uint<kNBitsBuffer> wptr, rptr;
ap_uint<2 * kNBitsBuffer> address(i);
(rptr,wptr) = address;
auto wptr1 = wptr+1;
auto wptr2 = wptr+2;
bool result = wptr1==rptr || wptr2==rptr;
lut[i] = result;
}
}
return lut;
}Of course, this applies to the other functions in this file.
There was a problem hiding this comment.
@bryates You might also remove the static const on the return types of these functions, since they don't really have any effect.
@bryates After addressing the comment above, could you rerun the out-of-context implementation and also do a comparison of the resource utilization, before your changes vs. after? Just to make sure nothing unexpected is happening during C-synthesis. |
These changes reduced the resource utilization a bit. The post-implementation timing is about the same. |
This is an alternative to #351 to address an issue where CMSSW wastes CPU cycles for emptyUnit and nearFull3Unit. All luts are now declared as static, so this should fix the issue.