VideoBackends:Multiple: Split up BUG_BROKEN_SUBGROUP_OPS

We now use subgroup ops for more than just a minor performance optimization

Source/Core/VideoBackends/Metal/MTLUtil.mm

@@ -245,8 +245,6 @@ void Metal::Util::PopulateBackendInfoFeatures(VideoConfig* config, id<MTLDevice>
         [device supportsFamily:MTLGPUFamilyMac2] || [device supportsFamily:MTLGPUFamilyApple6];
     config->backend_info.bSupportsFramebufferFetch = [device supportsFamily:MTLGPUFamilyApple1];
   }
-  if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS))
-    g_features.subgroup_ops = false;
 #if TARGET_OS_OSX
   if (@available(macOS 11, *))
     if (vendor == DriverDetails::VENDOR_INTEL)
@@ -434,6 +432,13 @@ std::optional<std::string> Metal::Util::TranslateShaderToMSL(ShaderStage stage,
   full_source.append(header);
   if (Metal::g_features.subgroup_ops)
     full_source.append(SUBGROUP_HELPER_HEADER);
+  if (DriverDetails::HasBug(DriverDetails::BUG_INVERTED_IS_HELPER))
+  {
+    full_source.append("#define gl_HelperInvocation !gl_HelperInvocation "
+                       "// Work around broken AMD Metal driver\n");
+  }
+  if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD))
+    full_source.append("#define BROKEN_SUBGROUP_WITH_DISCARD 1\n");
   full_source.append(source);
 
   std::optional<SPIRV::CodeVector> code;

Source/Core/VideoBackends/Vulkan/ShaderCompiler.cpp

@@ -7,6 +7,7 @@
 #include <string>
 
 #include "VideoBackends/Vulkan/VulkanContext.h"
+#include "VideoCommon/DriverDetails.h"
 #include "VideoCommon/Spirv.h"
 
 namespace Vulkan::ShaderCompiler
@@ -98,6 +99,13 @@ static std::string GetShaderCode(std::string_view source, std::string_view heade
     full_source_code.append(header);
     if (g_vulkan_context->SupportsShaderSubgroupOperations())
       full_source_code.append(SUBGROUP_HELPER_HEADER, subgroup_helper_header_length);
+    if (DriverDetails::HasBug(DriverDetails::BUG_INVERTED_IS_HELPER))
+    {
+      full_source_code.append("#define gl_HelperInvocation !gl_HelperInvocation "
+                              "// Work around broken AMD Metal driver\n");
+    }
+    if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD))
+      full_source_code.append("#define BROKEN_SUBGROUP_WITH_DISCARD 1\n");
     full_source_code.append(source);
   }
 

Source/Core/VideoBackends/Vulkan/VulkanContext.cpp

@@ -1000,8 +1000,7 @@ void VulkanContext::PopulateShaderSubgroupSupport()
       VK_SUBGROUP_FEATURE_BALLOT_BIT | VK_SUBGROUP_FEATURE_SHUFFLE_BIT;
   m_supports_shader_subgroup_operations =
       (subgroup_properties.supportedOperations & required_operations) == required_operations &&
-      subgroup_properties.supportedStages & VK_SHADER_STAGE_FRAGMENT_BIT &&
-      !DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS);
+      subgroup_properties.supportedStages & VK_SHADER_STAGE_FRAGMENT_BIT;
 }
 
 bool VulkanContext::SupportsExclusiveFullscreen(const WindowSystemInfo& wsi, VkSurfaceKHR surface)

Source/Core/VideoCommon/DriverDetails.cpp

@@ -130,14 +130,14 @@ constexpr BugInfo m_known_bugs[] = {
      -1.0, -1.0, true},
     {API_VULKAN, OS_ALL, VENDOR_ARM, DRIVER_ARM, Family::UNKNOWN, BUG_BROKEN_VECTOR_BITWISE_AND,
      -1.0, -1.0, true},
-    {API_VULKAN, OS_OSX, VENDOR_ATI, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS,
+    {API_VULKAN, OS_OSX, VENDOR_ATI, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_INVERTED_IS_HELPER,
      -1.0, -1.0, true},
-    {API_VULKAN, OS_OSX, VENDOR_INTEL, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS,
-     -1.0, -1.0, true},
-    {API_METAL, OS_OSX, VENDOR_ATI, DRIVER_APPLE, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS, -1.0,
-     -1.0, true},
-    {API_METAL, OS_OSX, VENDOR_INTEL, DRIVER_APPLE, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS, -1.0,
+    {API_METAL, OS_OSX, VENDOR_ATI, DRIVER_APPLE, Family::UNKNOWN, BUG_INVERTED_IS_HELPER, -1.0,
      -1.0, true},
+    {API_VULKAN, OS_OSX, VENDOR_INTEL, DRIVER_PORTABILITY, Family::UNKNOWN,
+     BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD, -1.0, -1.0, true},
+    {API_METAL, OS_OSX, VENDOR_INTEL, DRIVER_APPLE, Family::UNKNOWN,
+     BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD, -1.0, -1.0, true},
     {API_OPENGL, OS_ANDROID, VENDOR_ALL, DRIVER_ALL, Family::UNKNOWN,
      BUG_BROKEN_MULTITHREADED_SHADER_PRECOMPILATION, -1.0, -1.0, true},
     {API_VULKAN, OS_ANDROID, VENDOR_ALL, DRIVER_ALL, Family::UNKNOWN,

Source/Core/VideoCommon/DriverDetails.h

@@ -299,14 +299,23 @@ enum Bug
 
   // BUG: Accessing gl_SubgroupInvocationID causes the Metal shader compiler to crash.
   //      Affected devices: AMD (older macOS)
-  // BUG: gl_HelperInvocation always returns true, even for non-helper invocations
-  //      Affected devices: AMD (newer macOS)
+  // Started version: ???
+  // Ended version: ???
+  // (Workaround currently disabled, will put it back when someone hits the issue and we can
+  //  find out what devices and OSes it actually affects)
+
   // BUG: Using subgroupMax in a shader that can discard results in garbage data
   //      (For some reason, this only happens at 4x+ IR on Metal, but 2x+ IR on MoltenVK)
   //      Affected devices: Intel (macOS)
   // Started version: -1
   // Ended version: -1
-  BUG_BROKEN_SUBGROUP_OPS,
+  BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD,
+
+  // BUG: gl_HelperInvocation is actually !gl_HelperInvocation
+  //      Affected devices: AMD (macOS)
+  // Started version: -1
+  // Ended version: -1
+  BUG_INVERTED_IS_HELPER,
 
   // BUG: Multi-threaded shader pre-compilation sometimes crashes
   // Used primarily in Videoconfig.cpp's GetNumAutoShaderPreCompilerThreads()

Source/Core/VideoCommon/PixelShaderGen.cpp

@@ -485,7 +485,7 @@ void UpdateBoundingBox(float2 rawpos) {{
   int2 pos_tl = pos & ~1;  // round down to even
   int2 pos_br = pos | 1;   // round up to odd
 
-#ifdef SUPPORTS_SUBGROUP_REDUCTION
+#if defined(SUPPORTS_SUBGROUP_REDUCTION) && !defined(BROKEN_SUBGROUP_WITH_DISCARD)
   if (!IS_HELPER_INVOCATION)
   {{
     SUBGROUP_MIN(pos_tl);