Optimize softmax via flash attention v2 (#2468)

* optimize softmax as flash attention v2
intel · Jan 23, 2024 · f88a7d1 · f88a7d1
1 parent e046f5c
commit f88a7d1
Showing 1 changed file with 18 additions and 47 deletions.
diff --git a/csrc/cpu/aten/kernels/FlashAttentionKrnl.cpp b/csrc/cpu/aten/kernels/FlashAttentionKrnl.cpp
@@ -191,25 +191,25 @@ inline Vectorized<float> exp_u20(Vectorized<float> data) {
 
 // 1) out = exp(a - val)
 // 2) val = sum(out)
-template <typename scalar_t>
+template <typename T1, typename T2>
 inline void _exp_reduce_sum_fusion_kernel(
-    scalar_t* a,
+    T1* a,
     const int& size,
-    scalar_t* out,
-    scalar_t& val) {
-  auto vec_size = at::vec::Vectorized<scalar_t>::size();
-  auto vec_max = at::vec::Vectorized<scalar_t>(val);
-  scalar_t tmp_sum = 0;
-  auto vec_tmp_sum = at::vec::Vectorized<scalar_t>(tmp_sum);
+    T2* out,
+    T1& val) {
+  auto vec_size = at::vec::Vectorized<T1>::size();
+  auto vec_max = at::vec::Vectorized<T1>(val);
+  T1 tmp_sum = 0;
+  auto vec_tmp_sum = at::vec::Vectorized<T1>(tmp_sum);
   for (long i = 0; i < vec_size * (size / vec_size); i += vec_size) {
-    auto tmp0 = at::vec::Vectorized<scalar_t>::loadu(a + i);
+    auto tmp0 = at::vec::Vectorized<T1>::loadu(a + i);
     auto tmp1 = tmp0 - vec_max;
     auto tmp2 = exp_u20(tmp1);
     vec_tmp_sum += tmp2;
     at::native::_store(out + i, tmp2);
   }
-  tmp_sum = at::vec::vec_reduce_all<scalar_t>(
-      [](at::vec::Vectorized<scalar_t>& x, at::vec::Vectorized<scalar_t>& y) {
+  tmp_sum = at::vec::vec_reduce_all<T1>(
+      [](at::vec::Vectorized<T1>& x, at::vec::Vectorized<T1>& y) {
         return x + y;
       },
       vec_tmp_sum);
@@ -223,27 +223,6 @@ inline void _exp_reduce_sum_fusion_kernel(
   val = tmp_sum;
 }
 
-// out = a / sum
-template <typename T1, typename T2>
-inline void _normalization_kernel(
-    const T1* a,
-    const T1& sum,
-    const int& size,
-    T2* out) {
-  auto vec_size = at::vec::Vectorized<T1>::size();
-  auto vec_sum = at::vec::Vectorized<T1>(sum);
-  for (long i = 0; i < vec_size * (size / vec_size); i += vec_size) {
-    auto tmp0 = at::vec::Vectorized<T1>::loadu(a + i);
-    auto tmp1 = tmp0 / vec_sum;
-    at::native::_store(out + i, tmp1);
-  }
-  for (long i = vec_size * (size / vec_size); i < size; i++) {
-    auto tmp0 = a[i];
-    auto tmp1 = tmp0 / sum;
-    out[i] = tmp1;
-  }
-}
-
 // 1) out = a * scale
 // 2) max = max(out)
 template <typename scalar_t>
@@ -767,29 +746,19 @@ void cpu_flash_attention(
               _exp_reduce_sum_fusion_kernel(
                   qk_data + row * kvBlockSize,
                   kvBlockSize,
-                  qk_data + row * kvBlockSize,
+                  conditional_data_ptr(qk_data, qk_reduced_data) +
+                      row * kvBlockSize,
                   tmp_sum);
               // exp_tmp <- exp(max[row] - max)
               exp_tmp = std::exp(qk_max_data[row] - tmp_max);
               // sum[row] <- sum + exp_tmp * sum[row]
               qk_sum_data[row] = tmp_sum + exp_tmp * qk_sum_data[row];
               // max[row] <- max
               qk_max_data[row] = tmp_max;
-              // qk <- qk / sum[row]
-              accum_t sum_new = qk_sum_data[row];
-              _normalization_kernel(
-                  qk_data + row * kvBlockSize,
-                  sum_new,
-                  kvBlockSize,
-                  conditional_data_ptr(qk_data, qk_reduced_data) +
-                      row * kvBlockSize);
-              // dst <- dst * sum_old / sum_new * exp_tmp
+              // dst <- dst * exp_tmp
               if (n > 0) {
-                accum_t sum_cor = sum_old / sum_new;
                 at::vec::map<accum_t>(
-                    [sum_cor, exp_tmp](Vec x) {
-                      return x * Vec(sum_cor) * Vec(exp_tmp);
-                    },
+                    [exp_tmp](Vec x) { return x * Vec(exp_tmp); },
                     dst_data + row * headSize,
                     dst_data + row * headSize,
                     headSize);
@@ -856,10 +825,12 @@ void cpu_flash_attention(
                   headSize);
             }
           }
+          // dst <- dst / sum[row]
           // reorder MHA output with strides
           for (int64_t row = 0; row < qBlockSize; ++row) {
+            accum_t sum_reciprocal = 1 / qk_sum_data[row];
             at::vec::map<scalar_t>(
-                [](Vec x) { return x; },
+                [sum_reciprocal](Vec x) { return x * Vec(sum_reciprocal); },
                 out_data + i * oStrideB + j * oStrideH + m * oStrideM +
                     row * oStrideM,
                 dst_data + row * headSize,