ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 20:51:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] moe sorting ex kernel to support expert > 128 (#1840)

Browse Source 
* moe sorting ex

* fix bug for race condition

* fix bug and optimze large expert

* fix

* optimize with sub_token_oneshot

* support skip empty tokens for expert sorting

* update moe_sorting

* tidy code
This commit is contained in:
carlushuang
2025-02-11 17:49:17 +08:00
committed by GitHub
parent 2312f4aaf9
commit c0adab4850
12 changed files with 936 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp
View File

@@ -22,7 +22,7 @@
// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5
// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
//
// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated)
// * this could be larger than actual, since actual tokens are on GPU
//
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]

693
include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp
View File

@@ -15,6 +15,10 @@ namespace ck_tile {
#define MOE_SORTING_MOCK_ID(token_id_, topk_id_) \
static_cast<uint32_t>(((token_id_)&0x00ffffff) | (((topk_id_)&0xff) << 24))
#ifndef MOE_SORTING_USE_EX_KERNEL
#define MOE_SORTING_USE_EX_KERNEL 1
#endif
// clang-format off
// [indexing implementation-1]
// using M_a as constexpr block_size to partition all tokens into different slices
@@ -28,7 +32,7 @@ namespace ck_tile {
// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5
// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
//
// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated)
// * this could be larger than actual, since actual tokens are on GPU
//
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]
@@ -55,6 +59,34 @@ namespace ck_tile {
// num_tokens_post_padded_ptr : [28]
// num_sorted_tiles_ptr : [7]
//
// skip_experts_with_zero_tokens(SkipExpertsWithZeroTokens)
// if enabled, the expert with no tokens will be skipped, in stead of padding to at least 1 unit_size(M_a)
//
// (pack below tensor, skip element marked with `-`)
// Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y - - - - Y Y Y Y
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]
// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 -|- exp-5 -|
// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, c, f, i, o]
//
//
// sorted_expert_ids_ptr : [0, 1, 2, 3, 3, 5]
// num_tokens_post_padded_ptr : [24]
//
// * local_expert_mask : indicate local expert mask used on current GPU (used for EP case)
// and modify the output expert-ID, because we will only have enbaled expert on specific GPU.
// we call expert input to this kernel as "global expert id", output as "local expert id"
//
// * local_expert_mask : [1, 0, 1, 1, 0, 1] (mask out expert-id=1, 4)
//
// (pack below tensor, skip element marked with `-`)
// Y Y Y Y - - - - Y Y Y Y Y Y Y Y Y Y Y Y - - - - Y Y Y Y
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]
// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 -|- exp-5 -|
// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, c, f, i, o]
//
// sorted_expert_ids_ptr : [0, 1, 2, 2, 3] (note original it was exper-id= 0, 2, 3, 5, but we produce "local expert id")
// num_tokens_post_padded_ptr : [20]
//
// * different from vLLM
// 1) token_id stored in sorted_token_ids_ptr is actual token_id, not token_id*top_K expanded id
// 2）need sorted_weight_ptr
@@ -67,10 +99,80 @@ namespace ck_tile {
// 4）num_tokens_post_padded_ptr/num_sorted_tiles_ptr (select one)
//
// max_num_tokens_padded: opk_ids.numel() + num_experts * (block_size - 1)
CK_TILE_HOST constexpr auto moe_sorting_get_smem_row_col(int num_tokens_, int num_experts_)
{
/* num_experts + 1
* +--------------------------------------+
* | |
* | |
* | | * -> sub-tokens
* | |
* | |
* +--------------------------------------+
* | | 2 -> cumsum buffer
* +--------------------------------------+
*
*/
int smem_cols = num_experts_ + 1; // usually experts is power of 2. padding here
int smem_rows = [&](){
index_t target_occupancy_ = 2;
constexpr index_t total_ = 65536 / sizeof(int);
constexpr index_t sub_unroll = 8;
constexpr index_t cumsum_bufs = 2; // 1 for cumsum, 1 for cnt
// at lease 2 lines, one for sub_token unroll, one for cumsum
// should be enough
if ((total_ / target_occupancy_) < ((cumsum_bufs+sub_unroll) * smem_cols)) {
if ((total_ / 1) < ((cumsum_bufs+sub_unroll) * smem_cols))
throw std::runtime_error("too many num_experts, can't allocate smem");
target_occupancy_ = 1;
}
int r = total_ / target_occupancy_ / smem_cols;
// round to sub_unroll multipl
int r_for_sub_token = r - cumsum_bufs;
r_for_sub_token = min(r_for_sub_token, num_tokens_);
r_for_sub_token = (r_for_sub_token + sub_unroll - 1) / sub_unroll * sub_unroll;
r_for_sub_token = max(r_for_sub_token, 1);
if(r_for_sub_token > 1)
{
int r_unroll_ = r_for_sub_token / sub_unroll;
// round to 1x/2x/4x/8x number of sub_unroll
int clz_ = __builtin_clz(r_unroll_); // 0b1:31 0b2:30, 0b3:30, 0b4:29
int mask_ = (1 << (31 - clz_)) - 1;
mask_ = mask_ > 0b111 ? 0b111 : mask_; //clamp to 8x at most
mask_ = ~mask_;
//printf("r_unroll_:%d, clz:%d, mask:%x\n", r_unroll_, clz_, mask_); fflush(stdout);
r_for_sub_token = (r_unroll_ & mask_) * sub_unroll;
}
// final check
if( (r_for_sub_token + cumsum_bufs * smem_cols * target_occupancy_ ) >= total_ ) {
throw std::runtime_error("can't run this kernel, request LDS over size");
}
return r_for_sub_token + cumsum_bufs;
}();
// printf("r:%d, c:%d\n", smem_rows, smem_cols);
return ck_tile::make_tuple(smem_rows, smem_cols);
}
struct MoeSortingHostArgs
{
const void* p_topk_ids; // [token, topk]
const void* p_weights; // [token, topk]
const void* p_local_expert_mask;
void* p_sorted_token_ids;
void* p_sorted_weights;
void* p_sorted_expert_ids;
@@ -101,6 +203,7 @@ struct MoeSortingKernel
{
const void* p_topk_ids;
const void* p_weights;
const void* p_local_expert_mask;
void* p_sorted_token_ids;
void* p_sorted_weights;
void* p_sorted_expert_ids;
@@ -111,8 +214,11 @@ struct MoeSortingKernel
index_t moe_buf_bytes;
index_t tokens_per_thread;
index_t smem_rows;
mdiv unit_size_mdiv;
mdiv topk_mdiv;
mdiv expert_mdiv;
// mdiv sub_tokens_mdiv;
};
CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
@@ -123,15 +229,25 @@ struct MoeSortingKernel
CK_TILE_HOST static constexpr auto BlockSize(const Hargs& h)
{
#if MOE_SORTING_USE_EX_KERNEL
(void)h;
return dim3(256);
#else
return dim3(ck_tile::integer_least_multiple(h.num_experts, ck_tile::get_warp_size()));
#endif
}
// in byte
CK_TILE_HOST static constexpr auto GetSmemSize(const Hargs& h)
{
#if MOE_SORTING_USE_EX_KERNEL
auto [smem_rows, smem_cols] = moe_sorting_get_smem_row_col(h.tokens, h.num_experts);
return smem_rows * smem_cols * sizeof(int);
#else
const auto blocks = BlockSize(h);
// usually num_experts is power of 2, we pad 1 dword here for the row-size
return ((blocks.x + 1) * (h.num_experts + 1) + (h.num_experts + 1)) * sizeof(index_t);
#endif
}
CK_TILE_HOST static constexpr auto MakeKargs(const Hargs& h)
@@ -139,6 +255,7 @@ struct MoeSortingKernel
Kargs k;
k.p_topk_ids = h.p_topk_ids;
k.p_weights = h.p_weights;
k.p_local_expert_mask = h.p_local_expert_mask;
k.p_sorted_token_ids = h.p_sorted_token_ids;
k.p_sorted_weights = h.p_sorted_weights;
k.p_sorted_expert_ids = h.p_sorted_expert_ids;
@@ -152,10 +269,18 @@ struct MoeSortingKernel
k.tokens_per_thread = integer_divide_ceil(h.tokens * h.topk, blocks.x);
k.unit_size_mdiv = mdiv{static_cast<uint32_t>(h.unit_size)};
k.topk_mdiv = mdiv{static_cast<uint32_t>(h.topk)};
k.smem_rows = [&](){
auto [r_, c_] = moe_sorting_get_smem_row_col(h.tokens, h.num_experts);
(void) c_;
return r_;
}();
k.expert_mdiv = mdiv{static_cast<uint32_t>(h.num_experts)};
// k.sub_tokens_mdiv = mdiv{static_cast<uint32_t>(k.smem_rows - 1)};
return k;
}
// [a, b, c, d....] -> [a, a+b, a+b+c, a+b+c+d, ....]
// [a, b, c, d....] -> [a, a+b, a+b+c, a+b+c+d, ....]
// NOTE: wave_size need at least be 16!! dpp 16 is one row
template <typename data_t, int wave_size>
__device__ inline void wave_cumsum(data_t& thread_data) const
{
@@ -196,6 +321,40 @@ struct MoeSortingKernel
bank_mask,
bound_ctrl))); // row_shr:4
}
if constexpr(wave_size == 8) {
// wave-size=8 need one extra shift
thread_data =
reduce_op(thread_data,
__builtin_bit_cast(data_t, __builtin_amdgcn_mov_dpp(__builtin_bit_cast(int, thread_data),
0x118,
row_mask,
bank_mask,
bound_ctrl))); // row_shr:8
#if 0
constexpr int bank_mask_0_7 = 0b1100;
auto reduce_op_r = [&](auto x_, auto y_) { return x_ - y_; };
thread_data = reduce_op_r(thread_data, __builtin_bit_cast(data_t,
__builtin_amdgcn_update_dpp(0, /* old value */
__builtin_bit_cast(int, thread_data),
0x157,
row_mask,
bank_mask_0_7,
bound_ctrl))// row_newbcast:7
);
#else
data_t xxx =__builtin_bit_cast(data_t,
__builtin_amdgcn_mov_dpp(__builtin_bit_cast(int, thread_data),
0x157,
row_mask,
bank_mask,
bound_ctrl)); // row_newbcast:7
data_t yyy = (__lane_id() / 8) % 2 == 0 ? 0 : xxx;
thread_data = thread_data - yyy;
#endif
}
if constexpr(wave_size > 8)
{
thread_data =
@@ -224,6 +383,36 @@ struct MoeSortingKernel
}
}
// reduce single pixel within a wave
template <typename T, typename F, index_t wave_size_ = warpSize>
__device__ static constexpr T wave_reduce(T local, F reduce_f, number<wave_size_> = {})
{
// constexpr int wave_size = 64;
// constexpr int reduce_stage = 6; // 1<<6=64
// clang-format off
constexpr int reduce_stage = [](){
if constexpr(wave_size_ == 2) return 1;
else if constexpr(wave_size_ == 4) return 2;
else if constexpr(wave_size_ == 8) return 3;
else if constexpr(wave_size_ == 16) return 4;
else if constexpr(wave_size_ == 32) return 5;
else if constexpr(wave_size_ == 64) return 6;
else return 0;
}();
// clang-format on
T v_local = local;
#pragma unroll reduce_stage
for(int i_stage = 0; i_stage < reduce_stage; i_stage++)
{
int src_lane = __lane_id() ^ (1 << i_stage);
int32_t v_remote_tmp =
__builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(v_local));
T v_remote = bit_cast<T>(v_remote_tmp);
v_local = reduce_f(v_local, v_remote);
}
return v_local;
}
CK_TILE_DEVICE index_t calc_index(index_t total_col, index_t row, index_t col) const
{
return row * total_col + col;
@@ -257,37 +446,37 @@ struct MoeSortingKernel
index_t* shared_mem = reinterpret_cast<index_t*>(smem);
index_t* tokens_cnts = shared_mem; // 2d: (blockDim.x + 1, num_experts)
index_t* cumsum = shared_mem + (blockDim.x + 1) * (num_experts+1); // 1: (num_experts + 1)
index_t* cumsum = shared_mem + (blockDim.x + 1) * (num_experts + 1); // 1: (num_experts + 1)
for(int i = 0; i < num_experts; ++i)
{
tokens_cnts[calc_index(num_experts+1, tid + 1, i)] = 0;
tokens_cnts[calc_index(num_experts + 1, tid + 1, i)] = 0;
}
#pragma unroll Problem_::InternalLoadUnroll
for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
{
++tokens_cnts[calc_index(num_experts+1, tid + 1, topk_id[i])];
++tokens_cnts[calc_index(num_experts + 1, tid + 1, topk_id[i])];
}
__syncthreads();
#if 1
if(tid < num_experts)
{
tokens_cnts[calc_index(num_experts+1, 0, tid)] = 0;
tokens_cnts[calc_index(num_experts + 1, 0, tid)] = 0;
index_t local_c[8];
index_t prev_c = 0;
// TODO: manually unroll. pragma unroll does not work well when we have dependency
for(int i = 1; i <= static_cast<index_t>(blockDim.x); i+= 8)
for(int i = 1; i <= static_cast<index_t>(blockDim.x); i += 8)
{
local_c[0] = tokens_cnts[calc_index(num_experts+1, i + 0, tid)];
local_c[1] = tokens_cnts[calc_index(num_experts+1, i + 1, tid)];
local_c[2] = tokens_cnts[calc_index(num_experts+1, i + 2, tid)];
local_c[3] = tokens_cnts[calc_index(num_experts+1, i + 3, tid)];
local_c[4] = tokens_cnts[calc_index(num_experts+1, i + 4, tid)];
local_c[5] = tokens_cnts[calc_index(num_experts+1, i + 5, tid)];
local_c[6] = tokens_cnts[calc_index(num_experts+1, i + 6, tid)];
local_c[7] = tokens_cnts[calc_index(num_experts+1, i + 7, tid)];
local_c[0] = tokens_cnts[calc_index(num_experts + 1, i + 0, tid)];
local_c[1] = tokens_cnts[calc_index(num_experts + 1, i + 1, tid)];
local_c[2] = tokens_cnts[calc_index(num_experts + 1, i + 2, tid)];
local_c[3] = tokens_cnts[calc_index(num_experts + 1, i + 3, tid)];
local_c[4] = tokens_cnts[calc_index(num_experts + 1, i + 4, tid)];
local_c[5] = tokens_cnts[calc_index(num_experts + 1, i + 5, tid)];
local_c[6] = tokens_cnts[calc_index(num_experts + 1, i + 6, tid)];
local_c[7] = tokens_cnts[calc_index(num_experts + 1, i + 7, tid)];
local_c[0] += prev_c;
local_c[1] += local_c[0];
@@ -299,51 +488,57 @@ struct MoeSortingKernel
local_c[7] += local_c[6];
prev_c = local_c[7];
tokens_cnts[calc_index(num_experts+1, i + 0, tid)] = local_c[0];
tokens_cnts[calc_index(num_experts+1, i + 1, tid)] = local_c[1];
tokens_cnts[calc_index(num_experts+1, i + 2, tid)] = local_c[2];
tokens_cnts[calc_index(num_experts+1, i + 3, tid)] = local_c[3];
tokens_cnts[calc_index(num_experts+1, i + 4, tid)] = local_c[4];
tokens_cnts[calc_index(num_experts+1, i + 5, tid)] = local_c[5];
tokens_cnts[calc_index(num_experts+1, i + 6, tid)] = local_c[6];
tokens_cnts[calc_index(num_experts+1, i + 7, tid)] = local_c[7];
tokens_cnts[calc_index(num_experts + 1, i + 0, tid)] = local_c[0];
tokens_cnts[calc_index(num_experts + 1, i + 1, tid)] = local_c[1];
tokens_cnts[calc_index(num_experts + 1, i + 2, tid)] = local_c[2];
tokens_cnts[calc_index(num_experts + 1, i + 3, tid)] = local_c[3];
tokens_cnts[calc_index(num_experts + 1, i + 4, tid)] = local_c[4];
tokens_cnts[calc_index(num_experts + 1, i + 5, tid)] = local_c[5];
tokens_cnts[calc_index(num_experts + 1, i + 6, tid)] = local_c[6];
tokens_cnts[calc_index(num_experts + 1, i + 7, tid)] = local_c[7];
}
}
#else
// TODO: below code still working, but slow in expert=32/topk=5 case. Put here for future heuristic
// TODO: below code still working, but slow in expert=32/topk=5 case. Put here for future
// heuristic
{
if(tid < num_experts)
tokens_cnts[calc_index(num_experts+1, 0, tid)] = 0;
for(int i = 0; i < num_experts; i+=8) {
tokens_cnts[calc_index(num_experts + 1, 0, tid)] = 0;
for(int i = 0; i < num_experts; i += 8)
{
index_t local_c[8];
#pragma unroll
for(int j = 0; j < 8; j++) {
local_c[j] = tokens_cnts[calc_index(num_experts+1, tid+1, i+j)];
#pragma unroll
for(int j = 0; j < 8; j++)
{
local_c[j] = tokens_cnts[calc_index(num_experts + 1, tid + 1, i + j)];
}
#pragma unroll
for(int j = 0; j < 8; j++) {
#pragma unroll
for(int j = 0; j < 8; j++)
{
wave_cumsum<int, 64>(local_c[j]);
}
#pragma unroll
for(int j = 0; j < 8; j++) {
tokens_cnts[calc_index(num_experts+1, tid+1, i+j)] = local_c[j];
#pragma unroll
for(int j = 0; j < 8; j++)
{
tokens_cnts[calc_index(num_experts + 1, tid + 1, i + j)] = local_c[j];
}
}
}
#endif
__syncthreads();
if constexpr (Problem::ExpertTile == 0) {
if constexpr(Problem::ExpertTile == 0)
{
if(tid == 0)
{
cumsum[0] = 0;
for(int i = 1; i <= num_experts; ++i)
{
auto current_units = [&]() {
index_t x_ = tokens_cnts[calc_index(num_experts+1, blockDim.x, i - 1)] +
unit_size_mdiv.divisor - 1;
index_t x_ = tokens_cnts[calc_index(num_experts + 1, blockDim.x, i - 1)] +
unit_size_mdiv.divisor - 1;
index_t y_ = unit_size_mdiv.div(x_);
return max(y_, 1) * unit_size_mdiv.divisor;
}();
@@ -351,20 +546,24 @@ struct MoeSortingKernel
}
*p_total_tokens_post_pad = cumsum[num_experts];
}
} else {
// TODO: we have out-of-bound read here. But result is still OK (will ignore tid >= expert)
// for simplicity, not check experts here.
int local_cnt = tokens_cnts[calc_index(num_experts+1, blockDim.x, tid)];
}
else
{
// TODO: we have out-of-bound read here. But result is still OK (will ignore tid >=
// expert) for simplicity, not check experts here.
int local_cnt = tokens_cnts[calc_index(num_experts + 1, blockDim.x, tid)];
int blocks_pers_expert = unit_size_mdiv.div(local_cnt + unit_size_mdiv.divisor - 1);
int padded_tokens_per_expert = max(blocks_pers_expert, 1) * unit_size_mdiv.divisor;
int local_cumsum = padded_tokens_per_expert;
int local_cumsum = padded_tokens_per_expert;
wave_cumsum<int, 64>(local_cumsum);
if(tid == (num_experts - 1)) {
cumsum[0] = 0;
if(tid == (num_experts - 1))
{
cumsum[0] = 0;
*p_total_tokens_post_pad = local_cumsum;
}
if(tid < num_experts) {
if(tid < num_experts)
{
cumsum[tid + 1] = local_cumsum;
}
}
@@ -373,7 +572,7 @@ struct MoeSortingKernel
if(tid < num_experts)
{
int e_start = cumsum[tid];
int e_end = cumsum[tid + 1];
int e_end = cumsum[tid + 1];
for(int i = e_start; i < e_end; i += unit_size_mdiv.divisor)
{
p_sorted_expert_ids[unit_size_mdiv.div(i)] = tid;
@@ -383,8 +582,8 @@ struct MoeSortingKernel
#pragma unroll Problem_::InternalLoadUnroll
for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
{
index_t expert_id = topk_id[i];
index_t local_cnt = tokens_cnts[calc_index(num_experts+1, tid, expert_id)];
index_t expert_id = topk_id[i];
index_t local_cnt = tokens_cnts[calc_index(num_experts + 1, tid, expert_id)];
index_t rank_post_pad = local_cnt + cumsum[expert_id];
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
uint32_t curr_token_id, curr_topk_id;
@@ -393,16 +592,17 @@ struct MoeSortingKernel
#else
p_sorted_token_ids[rank_post_pad] = topk_mdiv.div(i);
#endif
p_sorted_weights[rank_post_pad] = weights[i];
tokens_cnts[calc_index(num_experts+1, tid, expert_id)] = local_cnt+1;
p_sorted_weights[rank_post_pad] = weights[i];
tokens_cnts[calc_index(num_experts + 1, tid, expert_id)] = local_cnt + 1;
}
if constexpr (Problem::ExpertTile == 0) {
if constexpr(Problem::ExpertTile == 0)
{
const index_t prefill_token = topk_mdiv.div(numel);
if(tid < num_experts)
{
index_t expert_offset =
cumsum[tid] + tokens_cnts[calc_index(num_experts+1, blockDim.x, tid)];
cumsum[tid] + tokens_cnts[calc_index(num_experts + 1, blockDim.x, tid)];
index_t expert_end = cumsum[tid + 1];
while(expert_offset < expert_end)
{
@@ -417,16 +617,19 @@ struct MoeSortingKernel
}
}
}
else {
else
{
const index_t prefill_token = topk_mdiv.div(numel);
// TODO: only support expert-tile like 8, 16, 32
static constexpr index_t experts_per_wave = warpSize / Problem::ExpertTile;
{
index_t eid = tid / experts_per_wave;
index_t expert_offset =
cumsum[eid] + tokens_cnts[calc_index(num_experts+1, blockDim.x, eid)] + tid % experts_per_wave;
index_t eid = tid / experts_per_wave;
index_t expert_offset = cumsum[eid] +
tokens_cnts[calc_index(num_experts + 1, blockDim.x, eid)] +
tid % experts_per_wave;
index_t expert_end = cumsum[eid + 1];
if(eid < num_experts) {
if(eid < num_experts)
{
while(expert_offset < expert_end)
{
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
@@ -436,10 +639,363 @@ struct MoeSortingKernel
p_sorted_token_ids[expert_offset] = prefill_token;
#endif
p_sorted_weights[expert_offset] = static_cast<WeightType>(0.0);
expert_offset+=experts_per_wave;
expert_offset += experts_per_wave;
}
}
}
}
}
}
// only support index_t, and single pixel access
struct simple_smem_indexer
{
index_t* smem;
index_t row_stride;
// this is 2D
CK_TILE_DEVICE simple_smem_indexer(index_t* smem_, index_t row_stride_)
: smem(smem_), row_stride(row_stride_)
{
}
CK_TILE_DEVICE const index_t& operator()(index_t i_row, index_t i_col) const
{
return smem[i_row * row_stride + i_col];
}
CK_TILE_DEVICE index_t& operator()(index_t i_row, index_t i_col)
{
return smem[i_row * row_stride + i_col];
}
// this is 1D or linear
CK_TILE_DEVICE simple_smem_indexer(index_t* smem_) : smem(smem_), row_stride(0) {}
CK_TILE_DEVICE const index_t& operator()(index_t idx) const { return smem[idx]; }
CK_TILE_DEVICE index_t& operator()(index_t idx) { return smem[idx]; }
};
CK_TILE_DEVICE void
moe_align_block_size_kernel_ex(const IndexType* __restrict__ topk_id,
const WeightType* __restrict__ weights,
const IndexType* __restrict__ local_expert_mask,
index_t* p_sorted_token_ids,
WeightType* p_sorted_weights,
index_t* p_sorted_expert_ids,
index_t* p_total_tokens_post_pad,
const index_t num_experts,
const index_t tokens,
const mdiv unit_size_mdiv,
const mdiv topk_mdiv,
const mdiv expert_mdiv,
const index_t smem_rows,
void* smem) const
{
const index_t tid = static_cast<index_t>(threadIdx.x);
const index_t wid = __builtin_amdgcn_readfirstlane(tid / warpSize);
const index_t lid = __lane_id();
constexpr index_t block_size = 256; // blockDim.x;
const index_t sub_tokens = smem_rows - 2; // sub_tokens_mdiv.divisor;
const index_t topk = topk_mdiv.divisor;
auto f_sum = [](auto x_, auto y_) { return x_ + y_; };
const index_t smem_cols = num_experts + 1;
simple_smem_indexer smem_cumsum{reinterpret_cast<index_t*>(smem) + 0};
simple_smem_indexer smem_cumdup{reinterpret_cast<index_t*>(smem) + smem_cols};
simple_smem_indexer smem_tokens{reinterpret_cast<index_t*>(smem) + 2 * smem_cols,
smem_cols};
// #pragma unroll 8
for(int i = tid; i < (sub_tokens * num_experts); i += block_size)
{
uint32_t curr_token_id, curr_expert_id;
expert_mdiv.divmod(i, curr_token_id, curr_expert_id);
smem_tokens(curr_token_id, curr_expert_id) = 0;
}
__syncthreads();
for(int i_token = 0; i_token < tokens; i_token += sub_tokens)
{
// NOTE: below for loop can't have barrier inside!!
for(int i = tid; i < (sub_tokens * topk); i += block_size)
{
uint32_t curr_token_id, curr_topk_id;
topk_mdiv.divmod(i, curr_token_id, curr_topk_id);
int i_t = i_token + curr_token_id;
if(i_t < tokens)
{
int eid = topk_id[i_t * topk + curr_topk_id];
if constexpr(Problem::SubTokenOneShot)
smem_tokens(curr_token_id, eid) = curr_topk_id + 1;
else
smem_tokens(curr_token_id, eid)++;
}
__builtin_amdgcn_s_waitcnt(0xc07f);
}
__syncthreads(); // make sure different i_token iteration not overlap by different wave
}
// counting
if(tid == 0)
{
smem_cumsum(0) = 0;
// smem_cumdup(0) = 0;
}
{
constexpr int lane_group_sz = 8;
int lane_group_id = tid / lane_group_sz;
int lane_group_os = tid % lane_group_sz;
constexpr int lane_group_nm = block_size / lane_group_sz;
for(int i_e = lane_group_id; i_e < num_experts; i_e += lane_group_nm)
{
index_t local_c[Problem::SubTokenTile];
index_t cnt = 0;
for(int i = 0; i < sub_tokens; i += 8 * Problem::SubTokenTile)
{
#pragma unroll Problem::SubTokenTile
for(int j = 0; j < Problem::SubTokenTile; j++)
{
local_c[j] = smem_tokens(i + j * 8 + lane_group_os, i_e);
if constexpr(Problem::SubTokenOneShot)
{
local_c[j] = local_c[j] != 0 ? 1 : 0;
}
}
#pragma unroll Problem::SubTokenTile
for(int j = 0; j < Problem::SubTokenTile; j++)
{
cnt += wave_reduce(local_c[j], f_sum, number<8>{});
}
}
if(lane_group_os == 0)
smem_cumsum(i_e + 1) = cnt;
}
}
if constexpr(Problem::LocalExpertMasking)
{
smem_cumdup(0) = 0;
for(int i_e = tid; i_e < num_experts; i_e += block_size)
{
// reuse this buffer
smem_cumdup(i_e + 1) = local_expert_mask[i_e];
}
}
__syncthreads();
{
if(wid == 0)
{
// NOTE: under this block can never use __syncthreads!
int i_e_ = 0;
int local_cumsum_ = 0;
for(; i_e_ < num_experts; i_e_ += warpSize)
{
int pre_cumsum_ = smem_cumsum(lid == 0 ? i_e_ : 0);
int local_cnt = smem_cumsum(i_e_ + lid + 1);
int blocks_pers_expert =
unit_size_mdiv.div(local_cnt + unit_size_mdiv.divisor - 1);
int pre_cumsum_masking = [&]() {
if constexpr(Problem::LocalExpertMasking)
return smem_cumdup(lid == 0 ? i_e_ : 0);
else
return 0; // not used
}();
int local_masking = [&]() {
if constexpr(Problem::LocalExpertMasking)
return smem_cumdup(i_e_ + lid + 1);
else
return 0; // not used
}();
int padded_tokens_per_expert = [&]() {
int x_ = [&]() {
if constexpr(Problem::SkipExpertsWithZeroTokens)
{
// if local_cnt is zero, blocks_pers_expert will be zero
// this is what we want to achieve
return blocks_pers_expert * unit_size_mdiv.divisor;
}
else
{
return max(blocks_pers_expert, 1) * unit_size_mdiv.divisor;
}
}();
if constexpr(Problem::LocalExpertMasking)
{
return local_masking ? x_ : 0;
}
else
return x_;
}();
local_cumsum_ = padded_tokens_per_expert;
local_cumsum_ += pre_cumsum_; // note pre_cumsum must be added after local
// cumsum padded in case local cumsum is zero, but
// pre_sumsum has value, which will result int
// zero local cumsum(but we want at least padded)
wave_cumsum<int, warpSize>(local_cumsum_);
if((i_e_ + lid) < num_experts)
smem_cumsum(i_e_ + lid + 1) = local_cumsum_;
if constexpr(Problem::LocalExpertMasking)
{
local_masking += pre_cumsum_masking;
wave_cumsum<int, warpSize>(local_masking);
if((i_e_ + lid) < num_experts)
smem_cumdup(i_e_ + lid + 1) = local_masking;
}
// NOTE: this waitcnt is a must, compiler will not generate waitcnt lgkmcnt()
// for above write however __syncthreads will cause barrier with waves other
// than 0(which is not we want)
__builtin_amdgcn_s_waitcnt(0xc07f);
}
if((lid + i_e_ - warpSize) == (num_experts - 1))
{
*p_total_tokens_post_pad = local_cumsum_;
}
}
__syncthreads();
}
for(int i_e = tid; i_e < num_experts; i_e += block_size)
{
int e_start = smem_cumsum(i_e);
int e_end = smem_cumsum(i_e + 1);
int expert_id = [&]() {
if constexpr(Problem::LocalExpertMasking)
{
// local expert id from cumsum
return smem_cumdup(i_e);
}
else
return i_e;
}();
smem_cumdup(i_e) = e_start; // duplicate cumsum for later use
if constexpr(Problem::SkipExpertsWithZeroTokens)
{
if(e_start == e_end) // skip zero token expert
continue;
}
if constexpr(Problem::LocalExpertMasking)
{
if(local_expert_mask[i_e] == 0)
continue;
}
for(int i = e_start; i < e_end; i += unit_size_mdiv.divisor)
{
p_sorted_expert_ids[unit_size_mdiv.div(i)] = expert_id;
}
}
smem_cumdup(num_experts) = smem_cumsum(num_experts);
// fill the p_sorted_token_ids/p_sorted_weights
for(int i_token = 0; i_token < tokens; i_token += sub_tokens)
{
if constexpr(!Problem::SubTokenOneShot)
{
// clear every time
for(int i = tid; i < (sub_tokens * num_experts); i += block_size)
{
uint32_t curr_token_id, curr_expert_id;
expert_mdiv.divmod(i, curr_token_id, curr_expert_id);
smem_tokens(curr_token_id, curr_expert_id) = 0;
}
__syncthreads();
// load again
for(int i = tid; i < (sub_tokens * topk); i += block_size)
{
uint32_t curr_token_id_, curr_topk_id_;
topk_mdiv.divmod(i, curr_token_id_, curr_topk_id_);
int curr_token_id = static_cast<int>(curr_token_id_);
int curr_topk_id = static_cast<int>(curr_topk_id_);
int i_t = i_token + curr_token_id;
if(i_t < tokens)
{
int eid = topk_id[i_t * topk + curr_topk_id];
smem_tokens(curr_token_id, eid) = curr_topk_id + 1; // at least 1
}
}
__syncthreads();
}
{
constexpr int lane_group_sz = 8;
int lane_group_id = tid / lane_group_sz;
int lane_group_os = tid % lane_group_sz;
constexpr int lane_group_nm = block_size / lane_group_sz;
for(int eid = lane_group_id; eid < num_experts; eid += lane_group_nm)
{
if constexpr(Problem::LocalExpertMasking)
{
if(local_expert_mask[eid] == 0)
continue;
}
int position = smem_cumsum(eid);
for(int i_sub_token = lane_group_os; i_sub_token < sub_tokens;
i_sub_token += lane_group_sz)
{
auto x = smem_tokens(i_sub_token, eid);
int local_cnt_cache = x != 0 ? 1 : 0;
int local_cnt = local_cnt_cache;
wave_cumsum<int, lane_group_sz>(local_cnt);
if(x != 0)
{
// now x is topk value
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
p_sorted_token_ids[position + local_cnt - 1] =
MOE_SORTING_MOCK_ID(i_token + i_sub_token, x - 1);
#else
p_sorted_token_ids[position + local_cnt - 1] = i_token + i_sub_token;
#endif
p_sorted_weights[position + local_cnt - 1] =
weights[(i_token + i_sub_token) * topk + x - 1];
}
int remote_cnt = __builtin_amdgcn_ds_bpermute(
(lane_group_sz * (lane_group_id + 1) - 1) << 2, local_cnt);
position += remote_cnt;
}
smem_cumsum(eid) = position;
}
}
__syncthreads();
}
// add the skip number
for(int eid = tid; eid < num_experts; eid += block_size)
{
int e_start = smem_cumsum(eid);
int e_end = smem_cumdup(eid + 1);
if constexpr(Problem::SkipExpertsWithZeroTokens)
{
if(e_start == e_end) // skip zero token expert
continue;
}
while(e_start < e_end)
{
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
p_sorted_token_ids[e_start] = MOE_SORTING_MOCK_ID(tokens, topk);
#else
p_sorted_token_ids[e_start] = tokens;
#endif
p_sorted_weights[e_start] = static_cast<WeightType>(0.0);
e_start++;
}
}
}
@@ -456,6 +1012,24 @@ struct MoeSortingKernel
}
const size_t numel = kargs.tokens * kargs.topk_mdiv.divisor;
extern __shared__ char smem[];
#if MOE_SORTING_USE_EX_KERNEL
(void)numel;
return moe_align_block_size_kernel_ex(
static_cast<const IndexType*>(kargs.p_topk_ids),
static_cast<const WeightType*>(kargs.p_weights),
static_cast<const IndexType*>(kargs.p_local_expert_mask),
static_cast<IndexType*>(kargs.p_sorted_token_ids),
static_cast<WeightType*>(kargs.p_sorted_weights),
static_cast<IndexType*>(kargs.p_sorted_expert_ids),
static_cast<IndexType*>(kargs.p_total_tokens_post_pad),
kargs.num_experts,
kargs.tokens,
kargs.unit_size_mdiv,
kargs.topk_mdiv,
kargs.expert_mdiv,
kargs.smem_rows,
smem);
#else
return moe_align_block_size_kernel(static_cast<const IndexType*>(kargs.p_topk_ids),
static_cast<const WeightType*>(kargs.p_weights),
static_cast<IndexType*>(kargs.p_sorted_token_ids),
@@ -468,6 +1042,7 @@ struct MoeSortingKernel
kargs.unit_size_mdiv,
kargs.topk_mdiv,
smem);
#endif
}
};

52
include/ck_tile/ops/fused_moe/kernel/moe_sorting_problem.hpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename IndexType_,
typename WeightType_,
index_t InternalLoadUnroll_,
index_t ExpertTile_ = 0>
struct MoeSortingProblem
{
// TODO: this kernel only support warp per row
using WeightType = remove_cvref_t<WeightType_>;
using IndexType = remove_cvref_t<IndexType_>;
static constexpr index_t WarpSize = get_warp_size();
static constexpr index_t WarpsPerBlock = 1;
static constexpr index_t InternalLoadUnroll =
InternalLoadUnroll_; // TODO: need better design(like tile size)
static constexpr index_t ExpertTile = ExpertTile_; // TODO: only used in store out
};
template <typename IndexType_,
typename WeightType_,
index_t SubTokenTile_, // 1,2,4,8, or 0 in the future
bool SubTokenOneShot_, // if we only loop over once or not
bool LocalExpertMasking_, // used in EP case
bool SkipExpertsWithZeroTokens_ = true,
index_t ExpertTile_ = 0>
struct MoeSortingProblemEx
{
// TODO: this kernel only support warp per row
using WeightType = remove_cvref_t<WeightType_>;
using IndexType = remove_cvref_t<IndexType_>;
static constexpr index_t WarpSize = get_warp_size();
static constexpr index_t WarpsPerBlock = 1;
static constexpr index_t SubTokenTile = SubTokenTile_;
static constexpr bool SubTokenOneShot = SubTokenOneShot_;
static constexpr bool LocalExpertMasking = LocalExpertMasking_;
static constexpr bool SkipExpertsWithZeroTokens = SkipExpertsWithZeroTokens_;
static_assert(SubTokenTile == 1 || SubTokenTile == 2 || SubTokenTile == 4 || SubTokenTile == 8);
static constexpr index_t ExpertTile = ExpertTile_; // TODO: only used in store out
};
} // namespace ck_tile