Merge pull request #11523 from degasus/OGL_KHR_subgroup

VideoBackend/OGL: Prefer KHR_shader_subgroup over NV_shader_thread.

Source/Core/Common/GL/GLExtensions/GLExtensions.h

@@ -37,6 +37,7 @@
 #include "Common/GL/GLExtensions/EXT_texture_filter_anisotropic.h"
 #include "Common/GL/GLExtensions/HP_occlusion_test.h"
 #include "Common/GL/GLExtensions/KHR_debug.h"
+#include "Common/GL/GLExtensions/KHR_shader_subgroup.h"
 #include "Common/GL/GLExtensions/NV_depth_buffer_float.h"
 #include "Common/GL/GLExtensions/NV_occlusion_query_samples.h"
 #include "Common/GL/GLExtensions/NV_primitive_restart.h"

Source/Core/Common/GL/GLExtensions/KHR_shader_subgroup.h

@@ -0,0 +1,19 @@
+/*
+** Copyright (c) 2013-2015 The Khronos Group Inc.
+** SPDX-License-Identifier: MIT
+*/
+
+#include "Common/GL/GLExtensions/gl_common.h"
+
+#define GL_SUBGROUP_SIZE_KHR 0x9532
+#define GL_SUBGROUP_SUPPORTED_STAGES_KHR 0x9533
+#define GL_SUBGROUP_SUPPORTED_FEATURES_KHR 0x9534
+#define GL_SUBGROUP_QUAD_ALL_STAGES_KHR 0x9535
+#define GL_SUBGROUP_FEATURE_BASIC_BIT_KHR 0x00000001
+#define GL_SUBGROUP_FEATURE_VOTE_BIT_KHR 0x00000002
+#define GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR 0x00000004
+#define GL_SUBGROUP_FEATURE_BALLOT_BIT_KHR 0x00000008
+#define GL_SUBGROUP_FEATURE_SHUFFLE_BIT_KHR 0x00000010
+#define GL_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT_KHR 0x00000020
+#define GL_SUBGROUP_FEATURE_CLUSTERED_BIT_KHR 0x00000040
+#define GL_SUBGROUP_FEATURE_QUAD_BIT_KHR 0x00000080

Source/Core/VideoBackends/Metal/MTLUtil.mm

@@ -386,7 +386,6 @@ fragment float4 fbfetch_test(float4 in [[color(0), raster_order_group(0)]]) {
 #extension GL_KHR_shader_subgroup_ballot : enable
 
 #define SUPPORTS_SUBGROUP_REDUCTION 1
-#define CAN_USE_SUBGROUP_REDUCTION true
 #define IS_HELPER_INVOCATION gl_HelperInvocation
 #define IS_FIRST_ACTIVE_INVOCATION (subgroupElect())
 #define SUBGROUP_MIN(value) value = subgroupMin(value)

Source/Core/VideoBackends/OGL/OGLConfig.cpp

@@ -489,7 +489,14 @@ bool PopulateConfig(GLContext* m_main_gl_context)
     else if (GLExtensions::Version() >= 430)
     {
       // TODO: We should really parse the GL_SHADING_LANGUAGE_VERSION token.
-      g_ogl_config.eSupportedGLSLVersion = Glsl430;
+      if (GLExtensions::Version() >= 450)
+      {
+        g_ogl_config.eSupportedGLSLVersion = Glsl450;
+      }
+      else
+      {
+        g_ogl_config.eSupportedGLSLVersion = Glsl430;
+      }
       g_ogl_config.bSupportsTextureStorage = true;
       g_ogl_config.bSupportsImageLoadStore = true;
       g_Config.backend_info.bSupportsSSAA = true;
@@ -531,8 +538,23 @@ bool PopulateConfig(GLContext* m_main_gl_context)
   if (g_ogl_config.max_samples < 1 || !g_ogl_config.bSupportsMSAA)
     g_ogl_config.max_samples = 1;
 
-  g_ogl_config.bSupportsShaderThreadShuffleNV =
-      GLExtensions::Supports("GL_NV_shader_thread_shuffle");
+  const bool bSupportsIsHelperInvocation = g_ogl_config.bIsES ?
+                                               g_ogl_config.eSupportedGLSLVersion >= GlslEs320 :
+                                               g_ogl_config.eSupportedGLSLVersion >= Glsl450;
+  g_ogl_config.bSupportsKHRShaderSubgroup =
+      GLExtensions::Supports("GL_KHR_shader_subgroup") && bSupportsIsHelperInvocation;
+  if (g_ogl_config.bSupportsKHRShaderSubgroup)
+  {
+    // Check for the features: basic + arithmetic + ballot
+    GLint supported_features = 0;
+    glGetIntegerv(GL_SUBGROUP_SUPPORTED_FEATURES_KHR, &supported_features);
+    if (~supported_features &
+        (GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR |
+         GL_SUBGROUP_FEATURE_BALLOT_BIT_KHR))
+    {
+      g_ogl_config.bSupportsKHRShaderSubgroup = false;
+    }
+  }
 
   // We require texel buffers, image load store, and compute shaders to enable GPU texture decoding.
   // If the driver doesn't expose the extensions, but supports GL4.3/GLES3.1, it will still be

Source/Core/VideoBackends/OGL/OGLConfig.h

@@ -15,8 +15,9 @@ enum GlslVersion
   Glsl140,
   Glsl150,
   Glsl330,
-  Glsl400,  // and above
-  Glsl430,
+  Glsl400,    // and above
+  Glsl430,    // 430 - 440
+  Glsl450,    // 450 - xxx
   GlslEs300,  // GLES 3.0
   GlslEs310,  // GLES 3.1
   GlslEs320,  // GLES 3.2
@@ -61,7 +62,7 @@ struct VideoConfig
   bool bSupportsBitfield;
   bool bSupportsTextureSubImage;
   EsFbFetchType SupportedFramebufferFetch;
-  bool bSupportsShaderThreadShuffleNV;
+  bool bSupportsKHRShaderSubgroup;  // basic + arithmetic + ballot
 
   const char* gl_vendor;
   const char* gl_renderer;

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -78,6 +78,8 @@ static std::string GetGLSLVersionString()
     return "#version 400";
   case Glsl430:
     return "#version 430";
+  case Glsl450:
+    return "#version 450";
   default:
     // Shouldn't ever hit this
     return "#version ERROR";
@@ -720,25 +722,18 @@ void ProgramShaderCache::CreateHeader()
   }
 
   std::string shader_shuffle_string;
-  if (g_ogl_config.bSupportsShaderThreadShuffleNV)
+  if (g_ogl_config.bSupportsKHRShaderSubgroup)
   {
     shader_shuffle_string = R"(
-#extension GL_NV_shader_thread_group : enable
-#extension GL_NV_shader_thread_shuffle : enable
-#define SUPPORTS_SUBGROUP_REDUCTION 1
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
 
-// The xor shuffle below produces incorrect results if all threads in a warp are not active.
-#define CAN_USE_SUBGROUP_REDUCTION (ballotThreadNV(true) == 0xFFFFFFFFu)
-
-#define IS_HELPER_INVOCATION gl_HelperThreadNV
-#define IS_FIRST_ACTIVE_INVOCATION (gl_ThreadInWarpNV == findLSB(ballotThreadNV(!gl_HelperThreadNV)))
-#define SUBGROUP_REDUCTION(func, value) value = func(value, shuffleXorNV(value, 16, 32)); \
-                                        value = func(value, shuffleXorNV(value, 8, 32)); \
-                                        value = func(value, shuffleXorNV(value, 4, 32)); \
-                                        value = func(value, shuffleXorNV(value, 2, 32)); \
-                                        value = func(value, shuffleXorNV(value, 1, 32));
-#define SUBGROUP_MIN(value) SUBGROUP_REDUCTION(min, value)
-#define SUBGROUP_MAX(value) SUBGROUP_REDUCTION(max, value)
+#define SUPPORTS_SUBGROUP_REDUCTION 1
+#define IS_HELPER_INVOCATION gl_HelperInvocation
+#define IS_FIRST_ACTIVE_INVOCATION (subgroupElect())
+#define SUBGROUP_MIN(value) value = subgroupMin(value)
+#define SUBGROUP_MAX(value) value = subgroupMax(value)
 )";
   }
 

Source/Core/VideoBackends/Vulkan/ShaderCompiler.cpp

@@ -81,9 +81,8 @@ static const char SUBGROUP_HELPER_HEADER[] = R"(
   #extension GL_KHR_shader_subgroup_ballot : enable
 
   #define SUPPORTS_SUBGROUP_REDUCTION 1
-  #define CAN_USE_SUBGROUP_REDUCTION true
   #define IS_HELPER_INVOCATION gl_HelperInvocation
-  #define IS_FIRST_ACTIVE_INVOCATION (gl_SubgroupInvocationID == subgroupBallotFindLSB(subgroupBallot(!gl_HelperInvocation)))
+  #define IS_FIRST_ACTIVE_INVOCATION (subgroupElect())
   #define SUBGROUP_MIN(value) value = subgroupMin(value)
   #define SUBGROUP_MAX(value) value = subgroupMax(value)
 )";

Source/Core/VideoCommon/PixelShaderGen.cpp

@@ -457,15 +457,12 @@ void UpdateBoundingBox(float2 rawpos) {{
   int2 pos_br = pos | 1;   // round up to odd
 
 #ifdef SUPPORTS_SUBGROUP_REDUCTION
-  if (CAN_USE_SUBGROUP_REDUCTION) {{
-    int2 min_pos = IS_HELPER_INVOCATION ? int2(2147483647, 2147483647) : pos_tl;
-    int2 max_pos = IS_HELPER_INVOCATION ? int2(-2147483648, -2147483648) : pos_br;
-    SUBGROUP_MIN(min_pos);
-    SUBGROUP_MAX(max_pos);
+  if (!IS_HELPER_INVOCATION)
+  {{
+    SUBGROUP_MIN(pos_tl);
+    SUBGROUP_MAX(pos_br);
     if (IS_FIRST_ACTIVE_INVOCATION)
-      UpdateBoundingBoxBuffer(min_pos, max_pos);
-  }} else {{
-    UpdateBoundingBoxBuffer(pos_tl, pos_br);
+      UpdateBoundingBoxBuffer(pos_tl, pos_br);
   }}
 #else
   UpdateBoundingBoxBuffer(pos_tl, pos_br);