Compile vulkan in ncnn python setup

setup.py

@@ -24,17 +24,18 @@ def find_version():
 
 # Parse parameters from environment
 NCNN_VULKAN = False
+NCNN_CI_VULKAN = False
 for i, arg in enumerate(sys.argv):
     if arg == "-g" or arg == "--gpu":
         NCNN_VULKAN = True
 
 # Parse environment variables
 NCNN_VULKAN_FLAG = os.environ.get("NCNN_VULKAN", "FALSE")
-USE_CI_VULKAN_FLAG = os.environ.get("USE_CI_VULKAN", "FALSE")
+NCNN_CI_VULKAN_FLAG = os.environ.get("NCNN_CI_VULKAN", "FALSE")
 if (NCNN_VULKAN_FLAG == "TRUE"):
     NCNN_VULKAN = True
-if (USE_CI_VULKAN_FLAG == "TRUE"):
-    USE_CI_VULKAN = True
+if (NCNN_CI_VULKAN_FLAG == "TRUE"):
+    NCNN_CI_VULKAN = True
 
 # Convert distutils Windows platform specifiers to CMake -A arguments
 PLAT_TO_CMAKE = {
@@ -87,10 +88,10 @@ def build_extension(self, ext):
             cmake_args.append("-DNCNN_VULKAN=ON")
         else:
             cmake_args.append("-DNCNN_VULKAN=OFF")
-        if USE_CI_VULKAN:
-            cmake_args.append("-DUSE_CI_VULKAN=ON")
+        if NCNN_CI_VULKAN:
+            cmake_args.append("-DNCNN_CI_VULKAN=ON")
         else:
-            cmake_args.append("-DUSE_CI_VULKAN=OFF")
+            cmake_args.append("-DNCNN_CI_VULKAN=OFF")
         build_args = []
 
         if self.compiler.compiler_type == "msvc":

src/layer/arm/unaryop_arm.cpp

@@ -402,10 +402,14 @@ struct unary_op_round
 #endif
         return y;
 #else
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         float y = nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
 #endif
     }
@@ -431,13 +435,17 @@ struct unary_op_round
 #else
         float tmp[4];
         vst1q_f32(tmp, x);
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         tmp[0] = nearbyintf(tmp[0]);
         tmp[1] = nearbyintf(tmp[1]);
         tmp[2] = nearbyintf(tmp[2]);
         tmp[3] = nearbyintf(tmp[3]);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         float32x4_t y = vld1q_f32(tmp);
         return y;
 #endif

src/layer/arm/unaryop_arm_asimdhp.cpp

@@ -471,10 +471,14 @@ struct unary_op_round_fp16s
             :);
         return y;
 #else
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         __fp16 y = (__fp16)nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
 #endif
     }

src/layer/loongarch/unaryop_loongarch.cpp

@@ -383,10 +383,14 @@ struct unary_op_round
             :);
         return y;
 #else
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         float y = nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
 #endif
     }

src/layer/mips/unaryop_mips.cpp

@@ -397,10 +397,14 @@ struct unary_op_round
             :);
         return y;
 #else
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         float y = nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
 #endif
     }

src/layer/unaryop.cpp

@@ -198,10 +198,14 @@ struct unary_op_round
     float operator()(const float& x) const
     {
         // round to nearest even
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         float y = nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
     }
 };

src/layer/vulkan/convolution_vulkan.cpp

@@ -444,7 +444,7 @@ int Convolution_vulkan::create_pipeline(const Option& _opt)
                 specializations[1].i = activation_type;
                 specializations[2].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
                 specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
-                specializations[4 + 0].i = shape_winograd_gemm_packed.c;
+                specializations[4 + 0].i = shape_winograd_gemm_packed.h;
                 specializations[4 + 1].i = shape_winograd_gemm_packed.cstep;
                 specializations[4 + 2].i = block_x;
                 specializations[4 + 3].i = block_y;

src/layer/x86/unaryop_x86.cpp

@@ -578,10 +578,14 @@ struct unary_op_round
     {
         // round to nearest even
         // return (x + 12582912.f) - 12582912.f;
+#ifdef FE_TONEAREST
         int old_rm = fegetround();
         fesetround(FE_TONEAREST);
+#endif
         float y = nearbyintf(x);
+#ifdef FE_TONEAREST
         fesetround(old_rm);
+#endif
         return y;
     }
 #if __SSE2__