Vincenzo Innocente

New Heterogeneous Memory Pool

42

This PR replaces the old "notcub" cache allocator with a memory pool featuring lockfree operations backend agnostic implementation The data interface is based on a simple Buffer that is completely...

[Do Not Merge] New "CUDA" Memory Pool [12.4.X]

18

This is a new version of #37952 mostly for benchmarking purposes. It extends the use of the new memory pool to the calorimeters (including the HostAllocator). A new feature comes...

recent improvements in ocml NOT in release 4.5

2

From my tests I think that the recent improvements in ocml are not included in release 4.5. Am I missing anything? Is the improved ocml foreseen to be included soon...

j1 (both double and float) return NaN for "large" negative numbers

9

I suspect this is due to the use of "x" instead of "ax" in this portion of the code https://github.com/RadeonOpenCompute/ROCm-Device-Libs/blob/a1f82f37cf105f0091e0f9e850a0a75aabde6eb7/ocml/src/j1F.cl#L81

tgammaf becomes inf for tiny values of its argument

1

tgammaf returns inf for |x|

lgamma becomes inf for tiny negative values

2

for values that corresponds to x*x ~ 0 lgamma becomes inf for negative x. ``` __global__ void doit(double x) { auto y =log(x); auto z1= lgamma(x); auto z2= lgamma(-x); printf...

y1(-0.0) returns +inf (instead of -inf or Nan)

4

I noticed that both y1f and y1 return +inf for -0.0 (and NaN for any other negative number) y0(-0.0) return -inf linux libm return -inf for any negative number (NaN...

not all "pi" functions present in ROCm math library are available through hipcc

4

ROCm math library implements all "pi" functions: sinpi, cospi, tanpi and their inverse (including atan2pi) hip seems to provide access only to sinpi and cospi. see https://github.com/RadeonOpenCompute/ROCm-Device-Libs/issues/99

Shall we fix all those inefficient (improper?) use of "pow"?

7

most of our use of pow is inefficient if not improper: https://github.com/search?q=repo%3Acms-sw%2Fcmssw+pow%28+language%3AC%2B%2B&type=code&l=C%2B%2B&p=5 1) pow(x,2) : no the compiler will not substitute with x*x: we need to introduce our own inline...

enable backtrace in gcc12

7

Would it be possible to enable backtrace in gcc12 adding ```--enable-libstdcxx-backtrace=yes``` as an option to configure ?