maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[mlir][sparse] Refine f695554a2a wording a bit and fix the double conversion I broke with aggressive copy&paste

This commit is contained in:
Benjamin Kramer 2022-08-11 11:56:41 +02:00
parent 875694089d
commit 726719e970
1 changed files with 4 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
mlir/lib/ExecutionEngine/Float16bits.cpp
View File

@ -159,10 +159,12 @@ std::ostream &operator<<(std::ostream &os, const bf16 &d) {
#endif
#if defined(__x86_64__)
// On x86 bfloat16 is passed in SSE2 registers. Since both float and _Float16
// On x86 bfloat16 is passed in SSE registers. Since both float and __bf16
// are passed in the same register we can use the wider type and careful casting
// to conform to x86_64 psABI. This only works with the assumption that we're
// dealing with little-endian values passed in wider registers.
// Ideally this would directly use __bf16, but that type isn't supported by all
// compilers.
using BF16ABIType = float;
#else
// Default to uint16_t if we have nothing else.
@ -184,9 +186,5 @@ extern "C" BF16ABIType ATTR_WEAK __truncsfbf2(float f) {
extern "C" BF16ABIType ATTR_WEAK __truncdfbf2(double d) {
// This does a double rounding step, but it's precise enough for our use
// cases.
uint16_t bf = __truncsfbf2(static_cast<float>(d));
// The output can be a float type, bitcast it from uint16_t.
BF16ABIType ret = 0;
std::memcpy(&ret, &bf, sizeof(bf));
return ret;
return __truncsfbf2(static_cast<float>(d));
}