From af6cdb2004157a8567db33e6a4e64c97525c4f40 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Sat, 7 May 2022 18:50:29 +0800
Subject: [PATCH] feat(fallback): fix ci

GitOrigin-RevId: b6e4e59553a7afe0d7c8dcb2db480f3d266dbacd
---
 CMakeLists.txt                                      | 10 ++++++++++
 dnn/src/fallback/conv_bias/gi/fp32/strategy_5x4.cpp | 13 +++++++++++++
 dnn/src/fallback/conv_bias/gi/fp32/strategy_6x3.cpp | 13 +++++++++++++
 dnn/test/fallback/gi.cpp                            |  6 +++---
 4 files changed, 39 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 55c40bdc..8af8648f 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -453,6 +453,16 @@ if(MSVC OR WIN32)
   # FIXME: fix MegRay on windows
   message(STATUS "Disable distributed build on windows host build...")
   set(MGE_WITH_DISTRIBUTED OFF)
+  if(${MGE_ARCH} STREQUAL "i386" AND ${CMAKE_BUILD_TYPE} STREQUAL "Debug")
+    # https://docs.microsoft.com/en-us/cpp/build/reference/z7-zi-zi-debug-information-format?view=msvc-170
+    # Workround for error LNK1318
+    message(
+      STATUS
+        "force use full symbolic debugging with build for 32bit for Windows with Debug mode"
+    )
+    set(CMAKE_C_FLAGS_DEBUG "/Z7")
+    set(CMAKE_CXX_FLAGS_DEBUG "/Z7")
+  endif()
 else()
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra")
 
diff --git a/dnn/src/fallback/conv_bias/gi/fp32/strategy_5x4.cpp b/dnn/src/fallback/conv_bias/gi/fp32/strategy_5x4.cpp
index 77bd62c8..c26c9a5e 100644
--- a/dnn/src/fallback/conv_bias/gi/fp32/strategy_5x4.cpp
+++ b/dnn/src/fallback/conv_bias/gi/fp32/strategy_5x4.cpp
@@ -290,6 +290,13 @@ struct InputTransform5X4 {
         CONCAT(s, 4).mla(m1addm2, 0.0625f).add(m3addm4).mla(m5addm6, 16.0f); \
     } while (0)
 
+#if defined(__GNUC__) && !defined(__llvm__) && !defined(_MSC_VER)
+#define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
+#if GCC_VERSION < 80000
+#pragma GCC push_options
+#pragma GCC optimize("O0")
+#endif
+#endif
 template <BiasMode bmode, typename Op>
 struct OutputTransform5X4 {
     static void transform(
@@ -393,6 +400,12 @@ struct OutputTransform5X4 {
         }
     }
 };
+#if defined(__GNUC__) && !defined(__llvm__) && !defined(_MSC_VER)
+#define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
+#if GCC_VERSION < 80000
+#pragma GCC pop_options
+#endif
+#endif
 #undef OUTPUT_TRANSFORM
 #undef GET_VECTOR_HIGH_ELEM
 #undef GET_VECTOR_LOW_ELEM
diff --git a/dnn/src/fallback/conv_bias/gi/fp32/strategy_6x3.cpp b/dnn/src/fallback/conv_bias/gi/fp32/strategy_6x3.cpp
index 387794ed..9481d7bd 100644
--- a/dnn/src/fallback/conv_bias/gi/fp32/strategy_6x3.cpp
+++ b/dnn/src/fallback/conv_bias/gi/fp32/strategy_6x3.cpp
@@ -207,6 +207,13 @@ struct InputTransform6X3 {
         CONCAT(s, 5).mla(m3subm4, 32.f).add(m5subm6).add(m##7);    \
     } while (0);
 
+#if defined(__GNUC__) && !defined(__llvm__) && !defined(_MSC_VER)
+#define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
+#if GCC_VERSION < 80000
+#pragma GCC push_options
+#pragma GCC optimize("O0")
+#endif
+#endif
 template <BiasMode bmode, typename Op>
 struct OutputTransform6X3 {
     static void transform(
@@ -318,6 +325,12 @@ struct OutputTransform6X3 {
         }
     }
 };
+#if defined(__GNUC__) && !defined(__llvm__) && !defined(_MSC_VER)
+#define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
+#if GCC_VERSION < 80000
+#pragma GCC pop_options
+#endif
+#endif
 
 #undef GET_VECTOR_HIGH_ELEM
 #undef GET_VECTOR_LOW_ELEM
diff --git a/dnn/test/fallback/gi.cpp b/dnn/test/fallback/gi.cpp
index e37a872e..d49e1685 100644
--- a/dnn/test/fallback/gi.cpp
+++ b/dnn/test/fallback/gi.cpp
@@ -905,7 +905,7 @@ TEST_F(FALLBACK, GiMultiplyAddFloat32) {
         naive.push_back(s1[i] * s2[i] + s0[i]);
     }
 
-    assert_eq((float*)&ret, naive);
+    assert_lt((float*)&ret, naive, 1e-3);
 }
 
 TEST_F(FALLBACK, GiMultiplyAddScalarFloat32) {
@@ -2793,7 +2793,7 @@ TEST_F(FALLBACK, GiMaximumInt8) {
     for (size_t i = 0; i < SIMD_LEN_8; i++) {
         s2.push_back(s1[i] < s0[i] ? 0xFF : 0);
     }
-    s2.resize(SIMD_LEN);
+    s2.resize(SIMD_LEN_8);
     init((int8_t*)&src2, s2, SIMD_LEN_8);
     ret = GiMaximumInt8(src0, src1);
 
@@ -2853,7 +2853,7 @@ TEST_F(FALLBACK, GiMinimumInt8) {
     for (size_t i = 0; i < SIMD_LEN_8; i++) {
         s2.push_back(s1[i] > s0[i] ? 0xFF : 0);
     }
-    s2.resize(SIMD_LEN);
+    s2.resize(SIMD_LEN_8);
     init((int8_t*)&src2, s2, SIMD_LEN_8);
     ret = GiMinimumInt8(src0, src1);