shader_recompiler: Add subpixel offset for correct rounding at `ImageGather`

On AMD a subpixel offset of 1/512 of the texel size is applied to the texture coordinates at a ImageGather call to ensure the rounding at the texel centers is done the same way as in Maxwell or other Nvidia architectures.
See https://www.reedbeta.com/blog/texture-gathers-and-coordinate-precision/ for more details why this might be necessary.

This should fix shadow artifacts at object edges in Zelda: Breath of the Wild (#9957, #6956).

9 files changed, 86 insertions, 0 deletions

diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
index f335c8af0..418505475 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
@@ -143,6 +143,21 @@ IR::Inst* PrepareSparse(IR::Inst& inst) {
     }
     return sparse_inst;
 }
+
+std::string ImageGatherSubpixelOffset(const IR::TextureInstInfo& info, std::string_view texture,
+                                      std::string_view coords) {
+    switch (info.type) {
+    case TextureType::Color2D:
+    case TextureType::Color2DRect:
+        return fmt::format("{}+vec2(0.001953125)/vec2(textureSize({}, 0))", coords, texture);
+    case TextureType::ColorArray2D:
+    case TextureType::ColorCube:
+        return fmt::format("vec3({0}.xy+vec2(0.001953125)/vec2(textureSize({1}, 0)),{0}.z)", coords,
+                           texture);
+    default:
+        return std::string{coords};
+    }
+}
 } // Anonymous namespace
 
 void EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
@@ -340,6 +355,13 @@ void EmitImageGather(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
         LOG_WARNING(Shader_GLSL, "Device does not support sparse texture queries. STUBBING");
         ctx.AddU1("{}=true;", *sparse_inst);
     }
+    std::string coords_with_subpixel_offset;
+    if (ctx.profile.need_gather_subpixel_offset) {
+        // Apply a subpixel offset of 1/512 the texel size of the texture to ensure same rounding on
+        // AMD hardware as on Maxwell or other Nvidia architectures.
+        coords_with_subpixel_offset = ImageGatherSubpixelOffset(info, texture, coords);
+        coords = coords_with_subpixel_offset;
+    }
     if (!sparse_inst || !supports_sparse) {
         if (offset.IsEmpty()) {
             ctx.Add("{}=textureGather({},{},int({}));", texel, texture, coords,
@@ -387,6 +409,13 @@ void EmitImageGatherDref(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
         LOG_WARNING(Shader_GLSL, "Device does not support sparse texture queries. STUBBING");
         ctx.AddU1("{}=true;", *sparse_inst);
     }
+    std::string coords_with_subpixel_offset;
+    if (ctx.profile.need_gather_subpixel_offset) {
+        // Apply a subpixel offset of 1/512 the texel size of the texture to ensure same rounding on
+        // AMD hardware as on Maxwell or other Nvidia architectures.
+        coords_with_subpixel_offset = ImageGatherSubpixelOffset(info, texture, coords);
+        coords = coords_with_subpixel_offset;
+    }
     if (!sparse_inst || !supports_sparse) {
         if (offset.IsEmpty()) {
             ctx.Add("{}=textureGather({},{},{});", texel, texture, coords, dref);
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
index 02073c420..968901d42 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
@@ -261,6 +261,39 @@ Id BitTest(EmitContext& ctx, Id mask, Id bit) {
     const Id bit_value{ctx.OpBitwiseAnd(ctx.U32[1], shifted, ctx.Const(1u))};
     return ctx.OpINotEqual(ctx.U1, bit_value, ctx.u32_zero_value);
 }
+
+Id ImageGatherSubpixelOffset(EmitContext& ctx, const IR::TextureInstInfo& info, Id texture,
+                             Id coords) {
+    // Apply a subpixel offset of 1/512 the texel size of the texture to ensure same rounding on
+    // AMD hardware as on Maxwell or other Nvidia architectures.
+    const auto calculate_offset{[&](size_t dim) -> std::array<Id, 2> {
+        const Id nudge{ctx.Const(0x1p-9f)};
+        const Id image_size{ctx.OpImageQuerySizeLod(ctx.U32[dim], texture, ctx.u32_zero_value)};
+        const Id offset_x{ctx.OpFDiv(
+            ctx.F32[1], nudge,
+            ctx.OpConvertUToF(ctx.F32[1], ctx.OpCompositeExtract(ctx.U32[1], image_size, 0)))};
+        const Id offset_y{ctx.OpFDiv(
+            ctx.F32[1], nudge,
+            ctx.OpConvertUToF(ctx.F32[1], ctx.OpCompositeExtract(ctx.U32[1], image_size, 1)))};
+        return {ctx.OpFAdd(ctx.F32[1], ctx.OpCompositeExtract(ctx.F32[1], coords, 0), offset_x),
+                ctx.OpFAdd(ctx.F32[1], ctx.OpCompositeExtract(ctx.F32[1], coords, 1), offset_y)};
+    }};
+    switch (info.type) {
+    case TextureType::Color2D:
+    case TextureType::Color2DRect: {
+        const auto offset{calculate_offset(2)};
+        return ctx.OpCompositeConstruct(ctx.F32[2], offset[0], offset[1]);
+    }
+    case TextureType::ColorArray2D:
+    case TextureType::ColorCube: {
+        const auto offset{calculate_offset(3)};
+        return ctx.OpCompositeConstruct(ctx.F32[3], offset[0], offset[1],
+                                        ctx.OpCompositeExtract(ctx.F32[1], coords, 2));
+    }
+    default:
+        return coords;
+    }
+}
 } // Anonymous namespace
 
 Id EmitBindlessImageSampleImplicitLod(EmitContext&) {
@@ -423,6 +456,9 @@ Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id
                    const IR::Value& offset, const IR::Value& offset2) {
     const auto info{inst->Flags<IR::TextureInstInfo>()};
     const ImageOperands operands(ctx, offset, offset2);
+    if (ctx.profile.need_gather_subpixel_offset) {
+        coords = ImageGatherSubpixelOffset(ctx, info, TextureImage(ctx, info, index), coords);
+    }
     return Emit(&EmitContext::OpImageSparseGather, &EmitContext::OpImageGather, ctx, inst,
                 ctx.F32[4], Texture(ctx, info, index), coords, ctx.Const(info.gather_component),
                 operands.MaskOptional(), operands.Span());
@@ -432,6 +468,9 @@ Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, const IR::Value& index,
                        const IR::Value& offset, const IR::Value& offset2, Id dref) {
     const auto info{inst->Flags<IR::TextureInstInfo>()};
     const ImageOperands operands(ctx, offset, offset2);
+    if (ctx.profile.need_gather_subpixel_offset) {
+        coords = ImageGatherSubpixelOffset(ctx, info, TextureImage(ctx, info, index), coords);
+    }
     return Emit(&EmitContext::OpImageSparseDrefGather, &EmitContext::OpImageDrefGather, ctx, inst,
                 ctx.F32[4], Texture(ctx, info, index), coords, dref, operands.MaskOptional(),
                 operands.Span());
diff --git a/src/shader_recompiler/profile.h b/src/shader_recompiler/profile.h
index 253e0d0bd..31390e869 100644
--- a/src/shader_recompiler/profile.h
+++ b/src/shader_recompiler/profile.h
@@ -52,6 +52,10 @@ struct Profile {
     bool need_declared_frag_colors{};
     /// Prevents fast math optimizations that may cause inaccuracies
     bool need_fastmath_off{};
+    /// Some GPU vendors use a lower fixed point format of 16.8 when calculating pixel coordinates
+    /// in the ImageGather instruction than the Maxwell architecture does. Applying an offset does
+    /// fix this mismatching rounding behaviour.
+    bool need_gather_subpixel_offset{};
 
     /// OpFClamp is broken and OpFMax + OpFMin should be used instead
     bool has_broken_spirv_clamp{};
diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp
index 22ed16ebf..d36a0a7a1 100644
--- a/src/video_core/renderer_opengl/gl_device.cpp
+++ b/src/video_core/renderer_opengl/gl_device.cpp
@@ -169,6 +169,7 @@ Device::Device(Core::Frontend::EmuWindow& emu_window) {
     has_draw_texture = GLAD_GL_NV_draw_texture;
     warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
     need_fastmath_off = is_nvidia;
+    need_gather_subpixel_offset = is_amd;
     can_report_memory = GLAD_GL_NVX_gpu_memory_info;
 
     // At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h
index 3ff8cad83..e8104c4de 100644
--- a/src/video_core/renderer_opengl/gl_device.h
+++ b/src/video_core/renderer_opengl/gl_device.h
@@ -160,6 +160,10 @@ public:
         return need_fastmath_off;
     }
 
+    bool NeedsGatherSubpixelOffset() const {
+        return need_gather_subpixel_offset;
+    }
+
     bool HasCbufFtouBug() const {
         return has_cbuf_ftou_bug;
     }
@@ -225,6 +229,7 @@ private:
     bool has_draw_texture{};
     bool warp_size_potentially_larger_than_guest{};
     bool need_fastmath_off{};
+    bool need_gather_subpixel_offset{};
     bool has_cbuf_ftou_bug{};
     bool has_bool_ref_bug{};
     bool can_report_memory{};
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp
index 479bb8ba3..b40aa6f5e 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp
@@ -218,6 +218,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
           .lower_left_origin_mode = true,
           .need_declared_frag_colors = true,
           .need_fastmath_off = device.NeedsFastmathOff(),
+          .need_gather_subpixel_offset = device.NeedsGatherSubpixelOffset(),
 
           .has_broken_spirv_clamp = true,
           .has_broken_unsigned_image_offsets = true,
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index 0684cceed..f51257267 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -329,6 +329,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
 
         .lower_left_origin_mode = false,
         .need_declared_frag_colors = false,
+        .need_gather_subpixel_offset = device.NeedsGatherSubpixelOffset(),
 
         .has_broken_spirv_clamp = driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS,
         .has_broken_spirv_position_input = driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY,
diff --git a/src/video_core/vulkan_common/vulkan_device.cpp b/src/video_core/vulkan_common/vulkan_device.cpp
index 6f288b3f8..0939b62c9 100644
--- a/src/video_core/vulkan_common/vulkan_device.cpp
+++ b/src/video_core/vulkan_common/vulkan_device.cpp
@@ -431,6 +431,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
                         "AMD GCN4 and earlier have broken VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT");
             has_broken_cube_compatibility = true;
         }
+        need_gather_subpixel_offset = true;
     }
     if (extensions.sampler_filter_minmax && is_amd) {
         // Disable ext_sampler_filter_minmax on AMD GCN4 and lower as it is broken.
diff --git a/src/video_core/vulkan_common/vulkan_device.h b/src/video_core/vulkan_common/vulkan_device.h
index 41b5da18a..50e95bcca 100644
--- a/src/video_core/vulkan_common/vulkan_device.h
+++ b/src/video_core/vulkan_common/vulkan_device.h
@@ -554,6 +554,10 @@ public:
         return features.robustness2.nullDescriptor;
     }
 
+    bool NeedsGatherSubpixelOffset() const {
+        return need_gather_subpixel_offset;
+    }
+
     u32 GetMaxVertexInputAttributes() const {
         return properties.properties.limits.maxVertexInputAttributes;
     }
@@ -664,6 +668,7 @@ private:
     bool must_emulate_bgr565{};             ///< Emulates BGR565 by swizzling RGB565 format.
     bool dynamic_state3_blending{};         ///< Has all blending features of dynamic_state3.
     bool dynamic_state3_enables{};          ///< Has all enables features of dynamic_state3.
+    bool need_gather_subpixel_offset{};     ///< Needs offset at ImageGather for correct rounding.
     u64 device_access_memory{};             ///< Total size of device local memory in bytes.
     u32 sets_per_pool{};                    ///< Sets per Description Pool