Skip to content

Fix UE8M0 packing: mask mantissa bits when extracting fp32 exponents #35

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
yhyang201 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix UE8M0 packing: mask mantissa bits when extracting fp32 exponents #35

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 12 additions & 9 deletions deep_gemm/include/deep_gemm/impls/smxx_layout.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,13 @@

namespace deep_gemm {

/// Pack exponent bytes from 4 fp32 bit-patterns into one uint32 (UE8M0 format).
__device__ __forceinline__ uint32_t pack_4_fp32_exponents(
uint32_t v0, uint32_t v1, uint32_t v2, uint32_t v3) {
return ((v0 >> 23u) & 0xFFu) | (((v1 >> 23u) & 0xFFu) << 8u)
| (((v2 >> 23u) & 0xFFu) << 16u) | (((v3 >> 23u) & 0xFFu) << 24u);
}

template <uint32_t kNumThreads, uint32_t BLOCK_MN, uint32_t SF_K,
uint32_t PADDED_SF_K = SF_K + (1 - (SF_K % 2))>
CUTLASS_GLOBAL void transpose_fp32(const float* sf, float* out, const uint32_t mn) {
Expand Down Expand Up @@ -98,11 +105,7 @@ CUTLASS_GLOBAL void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out,
}

// Pack and store
uint32_t packed = 0;
packed |= (values[0] >> 23u);
packed |= (values[1] >> 15u);
packed |= (values[2] >> 7u);
packed |= (values[3] << 1u);
uint32_t packed = pack_4_fp32_exponents(values[0], values[1], values[2], values[3]);
if (const auto global_mn_idx = blockIdx.x * BLOCK_MN + mn_idx; global_mn_idx < mn)
out[sf_k_pack_idx * tma_aligned_mn + global_mn_idx] = packed;
}
Expand Down Expand Up @@ -178,10 +181,10 @@ CUTLASS_GLOBAL void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,

// Pack and store
uint4 packed;
packed.x = (values[0].x >> 23u) | (values[1].x >> 15u) | (values[2].x >> 7u) | (values[3].x << 1u);
packed.y = (values[0].y >> 23u) | (values[1].y >> 15u) | (values[2].y >> 7u) | (values[3].y << 1u);
packed.z = (values[0].z >> 23u) | (values[1].z >> 15u) | (values[2].z >> 7u) | (values[3].z << 1u);
packed.w = (values[0].w >> 23u) | (values[1].w >> 15u) | (values[2].w >> 7u) | (values[3].w << 1u);
packed.x = pack_4_fp32_exponents(values[0].x, values[1].x, values[2].x, values[3].x);
packed.y = pack_4_fp32_exponents(values[0].y, values[1].y, values[2].y, values[3].y);
packed.z = pack_4_fp32_exponents(values[0].z, values[1].z, values[2].z, values[3].z);
packed.w = pack_4_fp32_exponents(values[0].w, values[1].w, values[2].w, values[3].w);
reinterpret_cast<uint4*>(out + packed_sf_k_idx * mn)[mn_idx] = packed;
}
}
Expand Down