177 lines
7.6 KiB
Plaintext
177 lines
7.6 KiB
Plaintext
/*
|
|
* Copyright (c) 2025 by FlashInfer team.
|
|
*
|
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
* you may not use this file except in compliance with the License.
|
|
* You may obtain a copy of the License at
|
|
*
|
|
* http://www.apache.org/licenses/LICENSE-2.0
|
|
*
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
* See the License for the specific language governing permissions and
|
|
* limitations under the License.
|
|
*/
|
|
#include <cub/block/block_scan.cuh>
|
|
|
|
#include "../../utils.cuh"
|
|
|
|
namespace flashinfer {
|
|
|
|
union alignas(8) CostIndex {
|
|
struct {
|
|
int bucket_idx;
|
|
float cost;
|
|
};
|
|
long long packed;
|
|
};
|
|
|
|
__device__ __forceinline__ CostIndex min(CostIndex a, CostIndex b) {
|
|
return a.cost < b.cost || (a.cost == b.cost && a.bucket_idx < b.bucket_idx) ? a : b;
|
|
}
|
|
|
|
__device__ __forceinline__ CostIndex get_min_cost_index(CostIndex* warp_min_cost,
|
|
CostIndex cost_index, int num_buckets) {
|
|
#pragma unroll
|
|
for (int offset = 16; offset > 0; offset >>= 1) {
|
|
CostIndex other;
|
|
other.packed = __shfl_xor_sync(0xffffffff, cost_index.packed, offset);
|
|
cost_index = min(cost_index, other);
|
|
}
|
|
if (static_cast<int>(threadIdx.x) % 32 == 0) {
|
|
warp_min_cost[static_cast<int>(threadIdx.x) / 32] = cost_index;
|
|
}
|
|
__syncthreads();
|
|
if (static_cast<int>(threadIdx.x) < 32) {
|
|
cost_index = static_cast<int>(threadIdx.x) * 32 < num_buckets
|
|
? warp_min_cost[threadIdx.x]
|
|
: CostIndex{static_cast<int>(threadIdx.x) * 32,
|
|
cuda::std::numeric_limits<float>::infinity()};
|
|
#pragma unroll
|
|
for (int offset = 16; offset > 0; offset >>= 1) {
|
|
CostIndex other;
|
|
other.packed = __shfl_xor_sync(0xffffffff, cost_index.packed, offset);
|
|
cost_index = min(cost_index, other);
|
|
}
|
|
if (static_cast<int>(threadIdx.x) == 0) {
|
|
warp_min_cost[0] = cost_index;
|
|
}
|
|
}
|
|
__syncthreads();
|
|
return warp_min_cost[0];
|
|
}
|
|
|
|
__global__ void plan_kernel(int* qo_segment_offsets, int* kv_segment_offsets, int* qo_lens,
|
|
int* kv_lens, int* work_indptr, int* qo_tile_indices, int* head_indices,
|
|
int* batch_indices, int qo_tile_size, int batch_size, int num_heads,
|
|
int num_buckets, bool causal) {
|
|
__shared__ CostIndex warp_min_cost[32];
|
|
constexpr int MAX_BUCKET_SIZE = 256;
|
|
using BlockScan = cub::BlockScan<int, MAX_BUCKET_SIZE>;
|
|
__shared__ typename BlockScan::TempStorage temp_storage;
|
|
// first round, calculate the work count for each bucket
|
|
CostIndex thread_local_cost_index = {static_cast<int>(threadIdx.x), 0.f};
|
|
int thread_local_work_counter = 0;
|
|
if (static_cast<int>(threadIdx.x) >= num_buckets) {
|
|
thread_local_cost_index.cost = cuda::std::numeric_limits<float>::infinity();
|
|
}
|
|
|
|
for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
|
|
for (int batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
|
|
int qo_len = qo_lens ? qo_lens[batch_idx]
|
|
: qo_segment_offsets[batch_idx + 1] - qo_segment_offsets[batch_idx];
|
|
int kv_len = kv_lens ? kv_lens[batch_idx]
|
|
: kv_segment_offsets[batch_idx + 1] - kv_segment_offsets[batch_idx];
|
|
int num_qo_tiles = ceil_div(qo_len, qo_tile_size);
|
|
for (int qo_tile_idx = num_qo_tiles - 1; qo_tile_idx >= 0; --qo_tile_idx) {
|
|
auto min_cost_index =
|
|
get_min_cost_index(warp_min_cost, thread_local_cost_index, num_buckets);
|
|
int bucket_idx = min_cost_index.bucket_idx;
|
|
float cost = min_cost_index.cost;
|
|
if (bucket_idx == threadIdx.x) {
|
|
thread_local_cost_index.cost +=
|
|
causal ? kv_len - (num_qo_tiles - qo_tile_idx - 1) * qo_tile_size : kv_len;
|
|
thread_local_work_counter++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
__syncthreads();
|
|
// compute exclusive prefix sum of
|
|
int thread_local_work_indptr = 0;
|
|
BlockScan(temp_storage).ExclusiveSum(thread_local_work_counter, thread_local_work_indptr);
|
|
__syncthreads();
|
|
if (static_cast<int>(threadIdx.x) < num_buckets) {
|
|
work_indptr[threadIdx.x] = thread_local_work_indptr;
|
|
}
|
|
if (static_cast<int>(threadIdx.x) + 1 == num_buckets) {
|
|
work_indptr[num_buckets] = thread_local_work_indptr + thread_local_work_counter;
|
|
}
|
|
|
|
// second round, write qo_tile_idx, head_idx, batch_idx to the output
|
|
thread_local_work_counter = 0;
|
|
if (static_cast<int>(threadIdx.x) >= num_buckets) {
|
|
thread_local_cost_index.cost = cuda::std::numeric_limits<float>::infinity();
|
|
} else {
|
|
thread_local_cost_index.cost = 0.f;
|
|
}
|
|
for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
|
|
for (int batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
|
|
int qo_len = qo_lens ? qo_lens[batch_idx]
|
|
: qo_segment_offsets[batch_idx + 1] - qo_segment_offsets[batch_idx];
|
|
int kv_len = kv_lens ? kv_lens[batch_idx]
|
|
: kv_segment_offsets[batch_idx + 1] - kv_segment_offsets[batch_idx];
|
|
int num_qo_tiles = ceil_div(qo_len, qo_tile_size);
|
|
for (int qo_tile_idx = num_qo_tiles - 1; qo_tile_idx >= 0; --qo_tile_idx) {
|
|
auto min_cost_index =
|
|
get_min_cost_index(warp_min_cost, thread_local_cost_index, num_buckets);
|
|
int bucket_idx = min_cost_index.bucket_idx;
|
|
float cost = min_cost_index.cost;
|
|
if (bucket_idx == threadIdx.x) {
|
|
thread_local_cost_index.cost +=
|
|
causal ? kv_len - (num_qo_tiles - qo_tile_idx - 1) * qo_tile_size : kv_len;
|
|
qo_tile_indices[thread_local_work_indptr + thread_local_work_counter] = qo_tile_idx;
|
|
head_indices[thread_local_work_indptr + thread_local_work_counter] = head_idx;
|
|
batch_indices[thread_local_work_indptr + thread_local_work_counter] = batch_idx;
|
|
thread_local_work_counter++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
#if (__CUDACC_VER_MAJOR__ >= 12 && defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
|
|
asm volatile("griddepcontrol.launch_dependents;");
|
|
#endif
|
|
}
|
|
|
|
cudaError_t plan_kernel_wrapper(int* qo_segment_offsets, int* kv_segment_offsets, int* qo_lens,
|
|
int* kv_lens, int* work_indptr, int* qo_tile_indices,
|
|
int* head_indices, int* batch_indices, int qo_tile_size,
|
|
int batch_size, int num_heads, int num_buckets, bool causal,
|
|
bool enable_pdl, cudaStream_t stream) {
|
|
if (enable_pdl) {
|
|
cudaLaunchConfig_t config;
|
|
config.gridDim = 1;
|
|
config.blockDim = 256;
|
|
config.dynamicSmemBytes = 0;
|
|
config.stream = stream;
|
|
cudaLaunchAttribute attrs[1];
|
|
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
|
|
attrs[0].val.programmaticStreamSerializationAllowed = true;
|
|
config.numAttrs = 1;
|
|
config.attrs = attrs;
|
|
FLASHINFER_CUDA_CALL(
|
|
cudaLaunchKernelEx(&config, plan_kernel, qo_segment_offsets, kv_segment_offsets, qo_lens,
|
|
kv_lens, work_indptr, qo_tile_indices, head_indices, batch_indices,
|
|
qo_tile_size, batch_size, num_heads, num_buckets, causal));
|
|
} else {
|
|
plan_kernel<<<1, 256, 0, stream>>>(qo_segment_offsets, kv_segment_offsets, qo_lens, kv_lens,
|
|
work_indptr, qo_tile_indices, head_indices, batch_indices,
|
|
qo_tile_size, batch_size, num_heads, num_buckets, causal);
|
|
FLASHINFER_CUDA_CALL(cudaGetLastError());
|
|
}
|
|
return cudaSuccess;
|
|
}
|
|
|
|
} // namespace flashinfer
|