fp16 pack8 output, cc

Tencent · Oct 11, 2024 · 545f075 · 545f075
1 parent 5e4fcfa
commit 545f075
Show file tree

Hide file tree

Showing 2 changed files with 3,848 additions and 6,847 deletions.
diff --git a/src/layer/arm/gemm_arm.cpp b/src/layer/arm/gemm_arm.cpp
@@ -6201,7 +6201,15 @@ int Gemm_arm::forward_int8(const std::vector<Mat>& bottom_blobs, std::vector<Mat
     if (opt.use_packing_layout)
     {
         int outh = output_transpose ? N : M;
-        out_elempack = outh % 4 == 0 ? 4 : 1;
+        if (support_fp16_storage && opt.use_fp16_arithmetic)
+        {
+            // TODO use output_elemtype
+            out_elempack = outh % 8 == 0 ? 8 : outh % 4 == 0 ? 4 : 1;
+        }
+        else
+        {
+            out_elempack = outh % 4 == 0 ? 4 : 1;
+        }
     }
 #endif // __ARM_NEON