leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

CUTLASS 3.4.0 (#1286)

Browse Source 
* CUTLASS 3.4.0

* Update CHANGELOG.md

---------

Co-authored-by: Pradeep Ramani <prramani@nvidia.com>
This commit is contained in:
Pradeep Ramani
2023-12-29 12:21:31 -08:00
committed by GitHub
parent b7508e3379
commit 8236f30675
211 changed files with 11409 additions and 2763 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
include/cutlass/bfloat16.h
View File

@@ -55,6 +55,7 @@
#include <cstring>
#endif
#include <cuda_bf16.h>
#include "cutlass/cutlass.h"
namespace cutlass {
@@ -83,6 +84,28 @@ struct alignas(2) bfloat16_t {
return h;
}
private:
struct from_32_bit_integer_t {};
static constexpr from_32_bit_integer_t from_32_bit_integer{};
template<class T>
CUTLASS_HOST_DEVICE
explicit bfloat16_t(from_32_bit_integer_t, T x) {
static_assert(cutlass::platform::is_integral<T>::value && sizeof(T) == 4, "Requires 32-bit integer");
float flt = static_cast<float>(x);
uint32_t bits;
#if defined(__CUDA_ARCH__)
bits = reinterpret_cast<uint32_t &>(flt);
#else
std::memcpy(&bits, &flt, sizeof(bits));
#endif
storage = uint16_t(bits >> 16);
}
public:
/// Default constructor
bfloat16_t() = default;
@@ -129,18 +152,10 @@ struct alignas(2) bfloat16_t {
/// Integer conversion - round toward nearest
CUTLASS_HOST_DEVICE
explicit bfloat16_t(int x) {
float flt = static_cast<float>(x);
uint32_t bits;
explicit bfloat16_t(int x) : bfloat16_t(from_32_bit_integer, x) {}
#if defined(__CUDA_ARCH__)
bits = reinterpret_cast<uint32_t &>(flt);
#else
std::memcpy(&bits, &flt, sizeof(bits));
#endif
storage = uint16_t(bits >> 16);
}
CUTLASS_HOST_DEVICE
explicit bfloat16_t(uint32_t x) : bfloat16_t(from_32_bit_integer, x) {}
/// Converts to float
CUTLASS_HOST_DEVICE