4 files changed, 22 insertions, 17 deletions
diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp
index bf08a6d93..5838fc02f 100644
--- a/src/video_core/renderer_opengl/gl_device.cpp
+++ b/src/video_core/renderer_opengl/gl_device.cpp
@@ -162,6 +162,7 @@ Device::Device() {
     has_amd_shader_half_float = GLAD_GL_AMD_gpu_shader_half_float;
     has_sparse_texture_2 = GLAD_GL_ARB_sparse_texture2;
     warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
+    need_fastmath_off = is_nvidia;
 
     // At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
     // uniform buffers as "push constants"
diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h
index 0b59c9df0..0c9d6fe31 100644
--- a/src/video_core/renderer_opengl/gl_device.h
+++ b/src/video_core/renderer_opengl/gl_device.h
@@ -136,6 +136,10 @@ public:
         return warp_size_potentially_larger_than_guest;
     }
 
+    bool NeedsFastmathOff() const {
+        return need_fastmath_off;
+    }
+
 private:
     static bool TestVariableAoffi();
     static bool TestPreciseBug();
@@ -171,6 +175,7 @@ private:
     bool has_amd_shader_half_float{};
     bool has_sparse_texture_2{};
     bool warp_size_potentially_larger_than_guest{};
+    bool need_fastmath_off{};
 };
 
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
index d27a3cf46..8d11fbc55 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
@@ -132,28 +132,23 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
     std::ranges::transform(infos, stage_infos.begin(),
                            [](const Shader::Info* info) { return info ? *info : Shader::Info{}; });
     auto func{[this, device, sources, shader_notify, xfb_state](ShaderContext::Context*) mutable {
-        if (device.UseAssemblyShaders()) {
-            for (size_t stage = 0; stage < 5; ++stage) {
-                const auto code{sources[stage]};
-                if (code.empty()) {
-                    continue;
-                }
+        if (!device.UseAssemblyShaders()) {
+            program.handle = glCreateProgram();
+        }
+        for (size_t stage = 0; stage < 5; ++stage) {
+            const auto code{sources[stage]};
+            if (code.empty()) {
+                continue;
+            }
+            if (device.UseAssemblyShaders()) {
                 assembly_programs[stage] = CompileProgram(code, AssemblyStage(stage));
                 enabled_stages_mask |= (assembly_programs[stage].handle != 0 ? 1 : 0) << stage;
-            }
-        } else {
-            program.handle = glCreateProgram();
-            for (size_t stage = 0; stage < 5; ++stage) {
-                const auto code{sources[stage]};
-                if (code.empty()) {
-                    continue;
-                }
+            } else {
                 AttachShader(Stage(stage), program.handle, code);
             }
-            LinkProgram(program.handle);
         }
-        if (shader_notify) {
-            shader_notify->MarkShaderComplete();
+        if (!device.UseAssemblyShaders()) {
+            LinkProgram(program.handle);
         }
         u32 num_textures{};
         u32 num_images{};
@@ -198,6 +193,9 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
         if (assembly_shaders && xfb_state) {
             GenerateTransformFeedbackState(*xfb_state);
         }
+        if (shader_notify) {
+            shader_notify->MarkShaderComplete();
+        }
         is_built.store(true, std::memory_order_relaxed);
     }};
     if (thread_worker) {
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp
index fedbce2f0..620666622 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp
@@ -193,6 +193,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
 
           .lower_left_origin_mode = true,
           .need_declared_frag_colors = true,
+          .need_fastmath_off = device.NeedsFastmathOff(),
 
           .has_broken_spirv_clamp = true,
           .has_broken_unsigned_image_offsets = true,