The randomly generated offset+width may exceeds 32, which causes
a left shift operation with 32-offset-width. As an unsigned number
that is greater than 32 and causes undefined behavior. When the
test is compiled without -mavx it is still OK. However when
the test is compiled with -mavx, the undefined behavior causes
wrong results and test failure.
This patch adjusts width so that offset+width<=32 always.
HIP_VERSION_MAJOR, HIP_VERSION_MINOR, HIP_VERSION_PATCH and HIP_VERSION pre-processor macros are now defined in hipVersion.h instead of being set by hipcc.
Added new memory API's hipMemAllocPitch, hipMemAllocHost, hipMemsetD16, hipMemsetD16Async, hipMemsetD8Async
Modified to support all scenarios hipMemcpyParam2DAsync, hipMemcpyParam2D.
Changed the third arg of the functions __hip_as_write_block and __ockl_as_write_block from ulong to uint64_t so as to fix the compilation error in windows
* Put 3-wide vector types on a ketogenic diet.
* Remove needless include.
* Do not be narrow-minded.
* Do not be narrow-minded.
* Put the C people on a diet too.
- Current clang disallows any invocation of wrong-side functions even
under context with type-inspection only. Work around that by adding a
variant of `std::decl` with `__device__` attribute.
- It's a common mistake by assuming 1 << shamt would be promoted to
64-bit, if shamt is a 64-bit integer. That's not the case. Replace
that left shift to a 64-bit one to ensure it won't fall into undefined
behavior.
- Fix the host-side implementation as well for device function testing.
Since rcp implementations of non-default rounded versions are not correct or supported in OCML, guard them using the same macro OCML_BASIC_ROUNDED_OPERATIONS. Also update the docs and tests.
In ROCm-Device-Libs, they have dropped the non-default-rounded versions of add, sub, mul, div, sqrt and fma. Therefore, ocml has removed the rte, rtp, rtn, and rtz counterparts. This will remove the same math APIs in HIP for _ru, _rd, _rn, and _rz.
Device library has removed the non-default-rounded functions, so hipFloatMath will fail to build. These include the removal of __ocml_sqrt_rte, __ocml_sqrt_rtn, __ocml_sqrt_rtp, and __ocml_sqrt_rtz. As seen here: https://github.com/RadeonOpenCompute/ROCm-Device-Libs/commit/2fc04e10e1354edee331ce700f98a60f8255effb . Disable these function tests for now, until they are re-enabled, or deleted completely.
Jatin Chaudhary
Yaxun (Sam) Liu
Sarbojit2019
Maneesh Gupta
Alex Voicu
vsytch
ansurya
AlexBinXie
Sameer Sahasrabuddhe
amd-lthakur
lthakur
Vladislav Sytchenko
Aaron Enye Shi
Michael LIAO
Michael Kuron
Michael Kuron