Optimize moe align block size kernel (#7794)

sgl-kernel/csrc/moe/moe_align_kernel.cu

@@ -21,16 +21,10 @@ limitations under the License.
 
 #include "utils.h"
 
-template <typename T, int N, int Alignment = sizeof(T) * N>
-class alignas(Alignment) AlignedArray {
- public:
-  T data[N];
-};
-
 #define WARP_SIZE 32
 
 #define VEC_SIZE 4
-using Vec = AlignedArray<int32_t, VEC_SIZE>;
+using Vec = int4;
 
 template <typename scalar_t>
 __global__ void count_and_sort_expert_tokens_kernel(
@@ -55,73 +49,119 @@ __global__ void moe_align_block_size_kernel(
     int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad,
     int32_t num_experts,
-    int32_t padded_num_experts,
-    int32_t experts_per_warp,
     int32_t block_size,
     size_t numel,
     int32_t* __restrict__ cumsum,
-    bool pad_sorted_token_ids) {
-  extern __shared__ int32_t shared_counts[];
+    bool pad_sorted_token_ids,
+    const int32_t scan_size) {
+  extern __shared__ int32_t smem[];
+  int32_t* shared_counts = smem;                  // [num_experts]
+  int32_t* prefix = shared_counts + num_experts;  // [num_experts + 1]
+  int32_t* scan_buf = prefix + num_experts + 1;   // [scan_size]
+  __shared__ int32_t s_total_tokens_post_pad;
 
-  const int warp_id = threadIdx.x / WARP_SIZE;
-  const int my_expert_start = warp_id * experts_per_warp;
+  const size_t tid = threadIdx.x;
+  const size_t stride = blockDim.x;
 
-  for (int i = 0; i < experts_per_warp; ++i) {
-    if (my_expert_start + i < padded_num_experts) {
-      shared_counts[warp_id * experts_per_warp + i] = 0;
-    }
+  if (tid < num_experts) {
+    shared_counts[tid] = 0;
   }
 
   __syncthreads();
 
-  const size_t tid = threadIdx.x;
-  const size_t stride = blockDim.x;
-
   for (size_t i = tid; i < numel; i += stride) {
     int expert_id = topk_ids[i];
-    int warp_idx = expert_id / experts_per_warp;
-    int expert_offset = expert_id % experts_per_warp;
-    atomicAdd(&shared_counts[warp_idx * experts_per_warp + expert_offset], 1);
+    atomicAdd(&shared_counts[expert_id], 1);
   }
 
   __syncthreads();
 
-  if (threadIdx.x == 0) {
-    cumsum[0] = 0;
-    for (int i = 1; i <= num_experts; ++i) {
-      int expert_count = 0;
-      int warp_idx = (i - 1) / experts_per_warp;
-      int expert_offset = (i - 1) % experts_per_warp;
-      expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
-
-      cumsum[i] = cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
+  int32_t padded_count = 0;
+  if (tid < num_experts) {
+    int32_t count = shared_counts[tid];
+    padded_count = (count + block_size - 1) / block_size * block_size;
+    scan_buf[tid] = padded_count;
   }
-    *total_tokens_post_pad = cumsum[num_experts];
+
+  if (tid >= num_experts && tid < scan_size) {
+    scan_buf[tid] = 0;
   }
 
   __syncthreads();
 
-  if (threadIdx.x < num_experts) {
-    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1]; i += block_size) {
-      expert_ids[i / block_size] = threadIdx.x;
+  // Blelloch scan
+  int offset = 1;
+#pragma unroll
+  for (int d = scan_size >> 1; d > 0; d >>= 1) {
+    if (tid < d) {
+      int ai = offset * (2 * tid + 1) - 1;
+      int bi = offset * (2 * tid + 2) - 1;
+      scan_buf[bi] += scan_buf[ai];
     }
+    offset <<= 1;
+    __syncthreads();
   }
 
-  if (pad_sorted_token_ids) {
-    int32_t fill_val = static_cast<int32_t>(numel);
-    int32_t total = *total_tokens_post_pad;
+  // down-sweep
+  if (tid == 0) {
+    prefix[num_experts] = scan_buf[scan_size - 1];
+    scan_buf[scan_size - 1] = 0;
+  }
+  __syncthreads();
 
-    Vec fill_vec;
 #pragma unroll
-    for (int i = 0; i < VEC_SIZE; ++i) {
-      fill_vec.data[i] = fill_val;
+  for (int d = 1; d < scan_size; d <<= 1) {
+    offset >>= 1;
+    if (tid < d) {
+      int ai = offset * (2 * tid + 1) - 1;
+      int bi = offset * (2 * tid + 2) - 1;
+      if (bi < scan_size) {
+        int temp = scan_buf[ai];
+        scan_buf[ai] = scan_buf[bi];
+        scan_buf[bi] += temp;
+      }
+    }
+    __syncthreads();
   }
 
-    int32_t total_vec_count = (total + VEC_SIZE - 1) / VEC_SIZE;
-    Vec* out_ptr = reinterpret_cast<Vec*>(sorted_token_ids);
+  if (tid < num_experts) {
+    prefix[tid] = scan_buf[tid];
+  }
+
+  if (tid == 0) {
+    s_total_tokens_post_pad = prefix[num_experts];
+    *total_tokens_post_pad = s_total_tokens_post_pad;
+  }
+
+  __syncthreads();
+
+  if (tid <= num_experts) {
+    cumsum[tid] = prefix[tid];
+  }
 
-    for (int32_t idx = tid; idx < total_vec_count; idx += stride) {
-      out_ptr[idx] = fill_vec;
+  // fill expert_ids
+  const int32_t num_blocks = s_total_tokens_post_pad / block_size;
+  for (int32_t i = tid; i < num_blocks; i += stride) {
+    int32_t block_start = i * block_size;
+    int left = 0, right = num_experts;
+    while (left < right) {
+      int mid = (left + right) >> 1;
+      if (prefix[mid] <= block_start) {
+        left = mid + 1;
+      } else {
+        right = mid;
+      }
+    }
+    expert_ids[i] = left - 1;
+  }
+
+  if (pad_sorted_token_ids) {
+    Vec fill_vec;
+    fill_vec.x = fill_vec.y = fill_vec.z = fill_vec.w = numel;
+    int32_t total_vecs = (s_total_tokens_post_pad + VEC_SIZE - 1) / VEC_SIZE;
+    Vec* out_ptr = reinterpret_cast<Vec*>(sorted_token_ids);
+    for (int32_t i = tid; i < total_vecs; i += stride) {
+      out_ptr[i] = fill_vec;
     }
   }
 }
@@ -179,20 +219,12 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
   }
 
   if (pad_sorted_token_ids) {
-    int32_t fill_val = static_cast<int32_t>(numel);
-    int32_t total = *total_tokens_post_pad;
-
     Vec fill_vec;
-#pragma unroll
-    for (int i = 0; i < VEC_SIZE; ++i) {
-      fill_vec.data[i] = fill_val;
-    }
-
-    int32_t total_vec_count = (total + VEC_SIZE - 1) / VEC_SIZE;
+    fill_vec.x = fill_vec.y = fill_vec.z = fill_vec.w = numel;
+    int32_t total_vecs = (*total_tokens_post_pad + VEC_SIZE - 1) / VEC_SIZE;
     Vec* out_ptr = reinterpret_cast<Vec*>(sorted_token_ids);
-
-    for (int32_t idx = tid; idx < total_vec_count; idx += stride) {
-      out_ptr[idx] = fill_vec;
+    for (int32_t i = tid; i < total_vecs; i += stride) {
+      out_ptr[i] = fill_vec;
     }
   }
 
@@ -245,8 +277,8 @@ void moe_align_block_size(
     } else {
       auto align_kernel = moe_align_block_size_kernel<scalar_t>;
 
-      size_t num_warps = CEILDIV(padded_num_experts, experts_per_warp);
-      size_t shared_mem_size = num_warps * experts_per_warp * sizeof(int32_t);
+      const size_t scan_size = next_pow2(num_experts);
+      const size_t shared_mem_size = (num_experts + (num_experts + 1) + scan_size) * sizeof(int32_t);
 
       align_kernel<<<1, threads, shared_mem_size, stream>>>(
           topk_ids.data_ptr<scalar_t>(),
@@ -254,12 +286,11 @@ void moe_align_block_size(
           experts_ids.data_ptr<int32_t>(),
           num_tokens_post_pad.data_ptr<int32_t>(),
           num_experts,
-          padded_num_experts,
-          experts_per_warp,
           block_size,
           topk_ids.numel(),
           cumsum_buffer.data_ptr<int32_t>(),
-          pad_sorted_token_ids);
+          pad_sorted_token_ids,
+          scan_size);
 
       const int block_threads = std::min(256, (int)threads);
       const int num_blocks = (topk_ids.numel() + block_threads - 1) / block_threads;

sgl-kernel/include/utils.h

@@ -363,3 +363,9 @@ inline torch::Tensor pad_tensor(const torch::Tensor& tensor, int64_t alignment =
   }
   return tensor_padded;
 }
+
+// Get the next power of 2 of a number
+inline uint32_t next_pow2(uint32_t x) noexcept {
+  if (x <= 1) return 1;
+  return 1u << (32 - __builtin_clz(x - 1));
+}