hdkv/yuzu
1
Fork 0
Code Issues Pull Requests Releases 1 Activity

early-access version 3501

Browse Source
This commit is contained in:
pineappleEA
2023-04-07 02:00:48 +02:00
parent e900ff1ff1
commit 3ef754c1e7
16 changed files with 37 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
View File

@@ -152,7 +152,8 @@ std::string ImageGatherSubpixelOffset(const IR::TextureInstInfo& info, std::stri
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);
return fmt::format("vec3({0}.xy+vec2(0.001953125)/vec2(textureSize({1}, 0)),{0}.z)", coords,
texture);
default:
return std::string{coords};
}

27
src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
View File

@@ -266,30 +266,21 @@ Id ImageGatherSubpixelOffset(EmitContext& ctx, const IR::TextureInstInfo& info,
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 auto calculate_coords{[&](size_t dim) {
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)};
Id offset{dim == 2 ? ctx.ConstantComposite(ctx.F32[dim], nudge, nudge)
: ctx.ConstantComposite(ctx.F32[dim], nudge, nudge, ctx.f32_zero_value)};
offset = ctx.OpFDiv(ctx.F32[dim], offset, ctx.OpConvertUToF(ctx.F32[dim], image_size));
return ctx.OpFAdd(ctx.F32[dim], coords, offset);
}};
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::Color2DRect:
return calculate_coords(2);
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));
}
case TextureType::ColorCube:
return calculate_coords(3);
default:
return coords;
}

6
src/shader_recompiler/profile.h
View File

@@ -52,9 +52,9 @@ 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.
/// Some GPU vendors use a different rounding precision when calculating texture pixel
/// coordinates with the 16.8 format in the ImageGather instruction than the Maxwell
/// architecture. Applying an offset does fix this mismatching rounding behaviour.
bool need_gather_subpixel_offset{};
/// OpFClamp is broken and OpFMax + OpFMin should be used instead