vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

SYCL : Improved implementation of dp4a for the Nvidia backend

Browse source
This commit is contained in:
Alberto Cabrera 2024-07-03 16:32:59 +01:00
parent 916248af1f
commit 44cee5dc89
1 changed files with 16 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

18
ggml/src/ggml-sycl/dpct/helper.hpp
View file

@ -1762,8 +1762,21 @@ namespace dpct
uint32_t, int32_t>;
template <typename T1, typename T2, typename T3>
inline auto dp4a(T1 a, T2 b, T3 c)
{
inline auto dp4a(T1 a, T2 b, T3 c) {
#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__) && \
defined(__SYCL_CUDA_ARCH__) && __SYCL_CUDA_ARCH__ >= 610
dot_product_acc_t<T1, T2> res;
if constexpr (std::is_same_v<dot_product_acc_t<T1, T2>, uint32_t>) {
asm volatile("dp4a.u32.u32 %0, %1, %2, %3;"
: "=r"(res)
: "r"(a), "r"(b), "r"(c));
} else {
asm volatile("dp4a.s32.s32 %0, %1, %2, %3;"
: "=r"(res)
: "r"(a), "r"(b), "r"(c));
}
return res;
#else
dot_product_acc_t<T1, T2> res = c;
auto va = extract_and_sign_or_zero_extend4(a);
auto vb = extract_and_sign_or_zero_extend4(b);
@ -1772,6 +1785,7 @@ namespace dpct
res += va[2] * vb[2];
res += va[3] * vb[3];
return res;
#endif
}
struct sub_sat