fix potential fp16s bf16s conflicts on armv7 vfpv4

Tencent · Jul 11, 2024 · 6aa28af · 6aa28af
1 parent c59885a
commit 6aa28af
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 4 deletions.
diff --git a/src/net.cpp b/src/net.cpp
@@ -622,7 +622,7 @@ int NetPrivate::convert_layout(Mat& bottom_blob, const Layer* layer, const Optio
         // *INDENT-OFF*
 
 #if NCNN_VFPV4
-        if (opt.use_fp16_storage && cpu_support_arm_vfpv4() && layer->support_fp16_storage)
+        if (opt.use_fp16_storage && !opt.use_bf16_storage && cpu_support_arm_vfpv4() && layer->support_fp16_storage)
         {
             Mat bottom_blob_fp16;
             cast_float32_to_float16(bottom_blob, bottom_blob_fp16, opt);
@@ -741,7 +741,7 @@ int NetPrivate::convert_layout(Mat& bottom_blob, const Layer* layer, const Optio
         // *INDENT-OFF*
 
 #if NCNN_VFPV4
-        if (opt.use_fp16_storage && cpu_support_arm_vfpv4() && !layer->support_fp16_storage)
+        if (opt.use_fp16_storage && !opt.use_bf16_storage && cpu_support_arm_vfpv4() && !layer->support_fp16_storage)
         {
             Mat bottom_blob_fp32;
             cast_float16_to_float32(bottom_blob, bottom_blob_fp32, opt);

diff --git a/tests/testutil.cpp b/tests/testutil.cpp
@@ -329,7 +329,7 @@ static int convert_to_optimal_layout(const ncnn::Mat& a, ncnn::Mat& a4, const nc
     // clang-format off
     // *INDENT-OFF*
 #if NCNN_VFPV4
-    if (opt.use_fp16_storage && ncnn::cpu_support_arm_vfpv4() && op->support_fp16_storage && !(flag & TEST_LAYER_DISABLE_AUTO_INPUT_CASTING))
+    if (opt.use_fp16_storage && !opt.use_bf16_storage && ncnn::cpu_support_arm_vfpv4() && op->support_fp16_storage && !(flag & TEST_LAYER_DISABLE_AUTO_INPUT_CASTING))
     {
         ncnn::cast_float32_to_float16(a, a4, opt);
     }
@@ -450,7 +450,7 @@ static int convert_to_vanilla_layout(const ncnn::Mat& c4, ncnn::Mat& c, const nc
     // clang-format off
     // *INDENT-OFF*
 #if NCNN_VFPV4
-    if (opt.use_fp16_storage && ncnn::cpu_support_arm_vfpv4() && op->support_fp16_storage && c4_unpacked.elembits() == 16)
+    if (opt.use_fp16_storage && !opt.use_bf16_storage && ncnn::cpu_support_arm_vfpv4() && op->support_fp16_storage && c4_unpacked.elembits() == 16)
     {
         ncnn::cast_float16_to_float32(c4_unpacked, c, opt);
     }