glsl: Implement VOTE for subgroup size potentially larger

src/shader_recompiler/backend/glsl/emit_context.cpp

@@ -122,9 +122,11 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
 
 void EmitContext::SetupExtensions(std::string&) {
     header += "#extension GL_ARB_separate_shader_objects : enable\n";
-    header += "#extension GL_ARB_sparse_texture2 : enable\n";
-    header += "#extension GL_EXT_texture_shadow_lod : enable\n";
-    // header += "#extension GL_ARB_texture_cube_map_array : enable\n";
+    if (stage != Stage::Compute) {
+        // TODO: track this usage
+        header += "#extension GL_ARB_sparse_texture2 : enable\n";
+        header += "#extension GL_EXT_texture_shadow_lod : enable\n";
+    }
     if (info.uses_int64) {
         header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
     }
@@ -149,6 +151,10 @@ void EmitContext::SetupExtensions(std::string&) {
         info.uses_subgroup_shuffles || info.uses_fswzadd) {
         header += "#extension GL_ARB_shader_ballot : enable\n";
         header += "#extension GL_ARB_shader_group_vote : enable\n";
+        header += "#extension GL_KHR_shader_subgroup_basic : enable\n";
+        if (!info.uses_int64) {
+            header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
+        }
     }
 }
 

src/shader_recompiler/backend/glsl/emit_glsl_warp.cpp

@@ -42,31 +42,42 @@ void EmitLaneId([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& in
 }
 
 void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
-    ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
-    // TODO:
-    // if (ctx.profile.warp_size_potentially_larger_than_guest) {
-    // }
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
+    } else {
+        const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
+        const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
+        ctx.AddU1("{}=({}&{})=={};", inst, ballot, active_mask, active_mask);
+    }
 }
 
 void EmitVoteAny(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
-    ctx.AddU1("{}=anyInvocationARB({});", inst, pred);
-    // TODO:
-    // if (ctx.profile.warp_size_potentially_larger_than_guest) {
-    // }
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        ctx.AddU1("{}=anyInvocationARB({});", inst, pred);
+    } else {
+        const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
+        const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
+        ctx.AddU1("{}=({}&{})!=0u;", inst, ballot, active_mask, active_mask);
+    }
 }
 
 void EmitVoteEqual(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
-    ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
-    // TODO:
-    // if (ctx.profile.warp_size_potentially_larger_than_guest) {
-    // }
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
+    } else {
+        const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
+        const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
+        const auto value{fmt::format("({}^{})", ballot, active_mask)};
+        ctx.AddU1("{}=({}==0)||({}=={});", inst, value, value, active_mask);
+    }
 }
 
 void EmitSubgroupBallot(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
-    ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred);
-    // TODO:
-    // if (ctx.profile.warp_size_potentially_larger_than_guest) {
-    // }
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred);
+    } else {
+        ctx.AddU32("{}=uvec2(ballotARB({}))[gl_SubgroupInvocationID];", inst, pred);
+    }
 }
 
 void EmitSubgroupEqMask(EmitContext& ctx, IR::Inst& inst) {

src/video_core/renderer_opengl/gl_device.cpp

@@ -160,6 +160,7 @@ Device::Device() {
     has_depth_buffer_float = HasExtension(extensions, "GL_NV_depth_buffer_float");
     has_nv_gpu_shader_5 = GLAD_GL_NV_gpu_shader5;
     has_amd_shader_half_float = GLAD_GL_AMD_gpu_shader_half_float;
+    warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
 
     // At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
     // uniform buffers as "push constants"

src/video_core/renderer_opengl/gl_device.h

@@ -128,6 +128,10 @@ public:
         return has_amd_shader_half_float;
     }
 
+    bool IsWarpSizePotentiallyLargerThanGuest() const {
+        return warp_size_potentially_larger_than_guest;
+    }
+
 private:
     static bool TestVariableAoffi();
     static bool TestPreciseBug();
@@ -161,6 +165,7 @@ private:
     bool has_depth_buffer_float{};
     bool has_nv_gpu_shader_5{};
     bool has_amd_shader_half_float{};
+    bool warp_size_potentially_larger_than_guest{};
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -220,7 +220,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
           .support_gl_nv_gpu_shader_5 = device.HasNvGpuShader5(),
           .support_gl_amd_gpu_shader_half_float = device.HasAmdShaderHalfFloat(),
 
-          .warp_size_potentially_larger_than_guest = true,
+          .warp_size_potentially_larger_than_guest = device.IsWarpSizePotentiallyLargerThanGuest(),
 
           .lower_left_origin_mode = true,
           .need_declared_frag_colors = true,