Skip to content

Butterfly all reduce for warp level reductions #635

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
alextmagro wants to merge 2 commits into
base: dev
Choose a base branch
from
Open

Butterfly all reduce for warp level reductions #635

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
13dec04
Butterfly all reduce for warp level reductions
alextmagro Jun 18, 2026
e0ca97e
Nest op structs
alextmagro Jun 18, 2026
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
1 change: 1 addition & 0 deletions transformer_engine/common/cast/mxfp8/gated_mxfp8.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,7 @@

#ifdef __HIP_PLATFORM_AMD__
#include "./rocm_vectorized_2d.cuh"
#include "../../util/rocm_device_utils.cuh"
#endif

namespace transformer_engine {
Expand Down
1 change: 1 addition & 0 deletions transformer_engine/common/cast/mxfp8/quantize_mxfp8.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@

#ifdef __HIP_PLATFORM_AMD__
#include "./rocm_vectorized_2d.cuh"
#include "../../util/rocm_device_utils.cuh"
#endif

namespace transformer_engine {
Expand Down
8 changes: 4 additions & 4 deletions transformer_engine/common/cast/mxfp8/rocm_gated_mxfp8.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,7 +176,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
// --- Act rowwise quantization ---
{
__builtin_assume(act_amax >= 0);
const float scale_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(act_amax);
const float scale_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(act_amax, rocm_op::max{});
const e8m0_t biased_exp =
ptx::float_to_e8m0(scale_amax * Quantized_Limits<OType>::max_norm_rcp);
const float scale_inv = ptx::exp2f_rcp(biased_exp);
Expand Down Expand Up @@ -210,7 +210,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
// --- Gate rowwise quantization (BWD only) ---
if constexpr (IS_DGATED) {
__builtin_assume(gate_amax >= 0);
const float scale_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(gate_amax);
const float scale_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(gate_amax, rocm_op::max{});
const e8m0_t biased_exp =
ptx::float_to_e8m0(scale_amax * Quantized_Limits<OType>::max_norm_rcp);
const float scale_inv = ptx::exp2f_rcp(biased_exp);
Expand Down Expand Up @@ -333,7 +333,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
// --- Act rowwise quantization ---
{
__builtin_assume(act_amax >= 0);
const float scale_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(act_amax);
const float scale_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(act_amax, rocm_op::max{});
const e8m0_t biased_exp =
ptx::float_to_e8m0(scale_amax * Quantized_Limits<OType>::max_norm_rcp);
const float scale_inv = ptx::exp2f_rcp(biased_exp);
Expand Down Expand Up @@ -367,7 +367,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
// --- Gate rowwise quantization (BWD only) ---
if constexpr (IS_DGATED) {
__builtin_assume(gate_amax >= 0);
const float scale_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(gate_amax);
const float scale_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(gate_amax, rocm_op::max{});
const e8m0_t biased_exp =
ptx::float_to_e8m0(scale_amax * Quantized_Limits<OType>::max_norm_rcp);
const float scale_inv = ptx::exp2f_rcp(biased_exp);
Expand Down
6 changes: 2 additions & 4 deletions transformer_engine/common/cast/mxfp8/rocm_quantize_mxfp8.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,8 +7,6 @@
// drop-in replacement for rocm quantize_mxfp8 kernels
//#include "hip/hip_runtime.h" //dummy include to prevent hipification adding this header

#include "../../util/rocm_device_utils.cuh"

constexpr size_t MXFP8_CHUNK_DIM_Y = 64;
constexpr size_t MXFP8_CHUNK_DIM_X = 64;
constexpr size_t MXFP8_THREADS_PER_CHUNK = 64;
Expand Down Expand Up @@ -163,7 +161,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
__builtin_assume(thread_amax >= 0);
block_amax = fmaxf(block_amax, thread_amax);

const float subwarp_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(thread_amax);
const float subwarp_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(thread_amax, rocm_op::max{});
const e8m0_t biased_exponent =
ptx::float_to_e8m0(subwarp_amax * Quantized_Limits<OType>::max_norm_rcp);

Expand Down Expand Up @@ -309,7 +307,7 @@ __global__ void __launch_bounds__(THREADS_PER_CHUNK)
__builtin_assume(thread_amax >= 0);
block_amax = fmaxf(block_amax, thread_amax);

const float subwarp_amax = subwarp_reduce_max_broadcast<SUBWARP_WIDTH>(thread_amax);
const float subwarp_amax = rocm_subwarp_allreduce<SUBWARP_WIDTH>(thread_amax, rocm_op::max{});
const e8m0_t biased_exponent =
ptx::float_to_e8m0(subwarp_amax * Quantized_Limits<OType>::max_norm_rcp);

Expand Down
25 changes: 25 additions & 0 deletions transformer_engine/common/util/rocm_device_utils.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,6 +135,31 @@ __device__ __forceinline__ int rocm_upper_bound(const T* arr, int n, T val) {
return lo;
}

// Binary reduction ops for rocm_subwarp_allreduce
struct rocm_op {
struct max {
__device__ __forceinline__ float operator()(float a, float b) const { return fmaxf(a, b); }
};

struct min {
__device__ __forceinline__ float operator()(float a, float b) const { return fminf(a, b); }
};

struct sum {
__device__ __forceinline__ float operator()(float a, float b) const { return a + b; }
};
};

// Butterfly all-reduce within a subwarp. All lanes get the result.
template <int WIDTH, typename T, typename OP>
__device__ __forceinline__ T rocm_subwarp_allreduce(T val, const OP &op) {
#pragma unroll
for (int offset = WIDTH / 2; offset > 0; offset >>= 1) {
val = op(val, __shfl_xor(val, offset, WIDTH));
}
return val;
}

template <int WARPS>
__device__ __forceinline__ float rocm_block_reduce_max(float val, int warp_id) {
__shared__ float staging[WARPS];
Expand Down