[libclc] Enable -ffp-contract=fast-honor-pragmas except for exp/trig/hyperbolic funcs #153137
Conversation
…unctions According to OpenCL spec, native_* functions have implementation-defined accuracy and typically have better performance. We can enable floating- point contraction optimizations for them.
|
I think fp contract should be globally enabled in the build, and selectively disabled in the handful of places that it is problematic (namely specific blocks in expF, sinbF, and trig reductions) |
libclc/CMakeLists.txt
Outdated
| @@ -304,7 +304,7 @@ set_source_files_properties( | |||
| ${CMAKE_CURRENT_SOURCE_DIR}/opencl/lib/generic/math/native_sin.cl | |||
| ${CMAKE_CURRENT_SOURCE_DIR}/opencl/lib/generic/math/native_sqrt.cl | |||
| ${CMAKE_CURRENT_SOURCE_DIR}/opencl/lib/generic/math/native_tan.cl | |||
| PROPERTIES COMPILE_OPTIONS -fapprox-func | |||
| PROPERTIES COMPILE_OPTIONS "-fapprox-func;-ffp-contract=fast" | |||
There was a problem hiding this comment.
Also maybe should use -ffp-contract=fast-honor-pragmas, not sure if the stupid interpretation ever got fixed for fast
…o exponential/trigonometric/hyperbolic funcs
wenju-he
changed the title
| @@ -6,6 +6,8 @@ | |||
| // | |||
| //===----------------------------------------------------------------------===// | |||
|
|
|||
| #pragma clang fp contract(off) | |||
There was a problem hiding this comment.
This can be much more targeted. The problematic areas can be specific block scopes inside of individual functions. I'd suggest running the conformance test with it enabled globally, and then finding the specific places that require this
There was a problem hiding this comment.
e.g. in exp f32
- float e = BUILTIN_RINT_F32(ph);
- float a = ph - e + pl;
+ float a, e;
+ {
+ #pragma OPENCL FP_CONTRACT OFF
+ e = BUILTIN_RINT_F32(ph);
+ a = ph - e + pl;
+ }
+
There was a problem hiding this comment.
This can be much more targeted. The problematic areas can be specific block scopes inside of individual functions. I'd suggest running the conformance test with it enabled globally, and then finding the specific places that require this
thanks, I'll run opencl cts on intel gpu to find the places.
Enable -ffp-contract=fast-honor-pragmas globally improves performance.
Disable in functions that may have problem with the flag.