study log

lite/backends/arm/math/packed_sgemm.cc

@@ -323,7 +323,7 @@ void sgemm_prepack(bool is_transB,
         (has_act == true && act_type == lite_api::ActivationType::kRelu);
     bool has_beta = fabsf(beta) > 1e-8f ? true : false;
     bool a53_sgemm = act_flag && !has_beta;
-    if (a53_sgemm) {
+    if (a53_sgemm) {//无act 无 beta
       sgemm_prepacked_6x8_a53(is_transB,
                               M,
                               N,
@@ -2368,16 +2368,19 @@ void sgemm_prepacked_8x12(bool is_transB,
   //! MBLOCK * x (result) + MBLOCK * k (A) + x * k (B) = l2
   int x_block = (l2_cache - (MBLOCK * K)) / (sizeof(float) * (K + MBLOCK));
   x_block /= NBLOCK;
-  x_block *= NBLOCK;
-  int x_num = (N + (x_block - 1)) / x_block;
-  x_block = (N + x_num - 1) / x_num;
-  x_block = (x_block + NBLOCK - 1) / NBLOCK;
-  x_block *= NBLOCK;
-  x_block = x_block < NBLOCK ? NBLOCK : x_block;
-
+  x_block *= NBLOCK;//一次可以放多少个B的列  为NBLOCK的整数倍
+  int x_num = (N + (x_block - 1)) / x_block; //可以分x_num 进行计算， 一次放x_block列，可以分x_num计算完成。
+  LOG(INFO) << "x_block:"<<x_block<<"  "<<"x_num"<<x_num;
+  x_block = (N + x_num - 1) / x_num; //分x_num次计算完成的话，每次需要计算多少个列 (N方向)，算一个平均值，因为最后一次可能会非常少，
+                                    //计算出x_num次之后，再求一下x_num次读取情况下，每次读取次数的平均值
+                                    //这样如果MP的时候，如果包含x_block个数比较少的情况下，可以使各线程耗时更加平均
+  x_block = (x_block + NBLOCK - 1) / NBLOCK;//算出每次NBLOCK的个数
+  x_block *= NBLOCK;//计算 一次做loadb 总的列数
+  x_block = x_block < NBLOCK ? NBLOCK : x_block;//如果不够NBLOCK，按NBLOCK来计算。
+  LOG(INFO) << "x_block:"<<x_block;
   // unroll 2 loop
-  int tail_pre = (K & (KBLOCK - 1));
-  int k_pre = ((K + KBLOCK - 1) / KBLOCK) - 1;
+  int tail_pre = (K & (KBLOCK - 1));//K方向 KBLOCK的余数
+  int k_pre = ((K + KBLOCK - 1) / KBLOCK) - 1;//K方向 整数倍KBLOCK的个数
 
   bool flag_p_remain = false;
   int remain = 0;
@@ -2390,8 +2393,8 @@ void sgemm_prepacked_8x12(bool is_transB,
     if (xmax > N) {
       xmax = N;
     }
-    int bblocks = (xmax - x0 + NBLOCK - 1) / NBLOCK;
-    remain = xmax - x0 - (bblocks - 1) * NBLOCK;
+    int bblocks = (xmax - x0 + NBLOCK - 1) / NBLOCK;//B 有多少个NBLOCK
+    remain = xmax - x0 - (bblocks - 1) * NBLOCK;//不够NBLOCK，的余数
     if (remain > 0) {
       flag_p_remain = true;
     }
@@ -2402,7 +2405,7 @@ void sgemm_prepacked_8x12(bool is_transB,
     } else {
       loadb(b_pannel, B, ldb, 0, K, x0, xmax);
     }
-#pragma omp parallel for num_threads(threads)
+#pragma omp parallel for num_threads(threads)//在A的M方向，按照MBLOCK进行MP
     for (unsigned int y = 0; y < M; y += MBLOCK) {
       unsigned int ymax = y + MBLOCK;
       if (ymax > M) {
@@ -2421,7 +2424,7 @@ void sgemm_prepacked_8x12(bool is_transB,
         bias_local[7] = bias[y + 7];
       }
 
-      float cout0[NBLOCK];
+      float cout0[NBLOCK];//C 输出  8*12
       float cout1[NBLOCK];
       float cout2[NBLOCK];
       float cout3[NBLOCK];