ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 13:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fp16/fp8 mixed-precision Gemm with multiply+add fusion (#865)

Browse Source 
* add compute_type

* add multiply_add ckProfiler

* add f8_fp16 support

* clean

* clean

* fixed lds size calc

* format

---------

Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
zjing14
2023-08-28 16:27:32 -05:00
committed by GitHub
parent c8a8385fdd
commit 31ea132aa2
22 changed files with 2082 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View File

@@ -195,6 +195,51 @@ struct AddMultiply
}
};
// C = A * B
// E = C x D0 + D1
struct MultiplyAdd
{
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
template <>
__host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
const half_t& c,
const half_t& d0,
const half_t& d1) const
{
const half_t y = (c * d0) + d1;
e = y;
}
template <>
__host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
const float& c,
const half_t& d0,
const half_t& d1) const
{
const half_t y = type_convert<half_t>(c) * d0 + d1;
e = y;
}
template <>
__host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
const float& c,
const half_t& d0,
const half_t& d1) const
{
const float y = c * d0 + d1;
e = y;
}
template <>
__host__ __device__ void operator()<half_t, float, float, float>(half_t& e,
const float& c,
const float& d0,
const float& d1) const
{
const float y = c * d0 + d1;
e = y;
}
};
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu
{