forked from ShuriZma/suyu
1
0
Fork 0

glsl_context_get_set: Add alternative cbuf type for broken drivers

some drivers have a bug bitwise converting floating point cbuf values to uint variables. This adds a workaround for these drivers to make all cbufs uint and convert to floating point as needed.
This commit is contained in:
ameerj 2021-12-27 23:59:32 -05:00
parent 9f34be5a61
commit b84d429c2e
6 changed files with 35 additions and 24 deletions

View File

@ -102,39 +102,46 @@ void GetCbuf16(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, const
void EmitGetCbufU8(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufU8(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
GetCbuf8(ctx, inst, binding, offset, "ftou"); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "" : "ftou"};
GetCbuf8(ctx, inst, binding, offset, cast);
} }
void EmitGetCbufS8(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufS8(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
GetCbuf8(ctx, inst, binding, offset, "ftoi"); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "int" : "ftoi"};
GetCbuf8(ctx, inst, binding, offset, cast);
} }
void EmitGetCbufU16(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufU16(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
GetCbuf16(ctx, inst, binding, offset, "ftou"); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "" : "ftou"};
GetCbuf16(ctx, inst, binding, offset, cast);
} }
void EmitGetCbufS16(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufS16(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
GetCbuf16(ctx, inst, binding, offset, "ftoi"); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "int" : "ftoi"};
GetCbuf16(ctx, inst, binding, offset, cast);
} }
void EmitGetCbufU32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufU32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)}; const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)};
GetCbuf(ctx, ret, binding, offset, 32, "ftou"); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "" : "ftou"};
GetCbuf(ctx, ret, binding, offset, 32, cast);
} }
void EmitGetCbufF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
const auto ret{ctx.var_alloc.Define(inst, GlslVarType::F32)}; const auto ret{ctx.var_alloc.Define(inst, GlslVarType::F32)};
GetCbuf(ctx, ret, binding, offset, 32); const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "utof" : ""};
GetCbuf(ctx, ret, binding, offset, 32, cast);
} }
void EmitGetCbufU32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitGetCbufU32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) { const IR::Value& offset) {
const auto cbuf{fmt::format("{}_cbuf{}", ctx.stage_name, binding.U32())}; const auto cbuf{fmt::format("{}_cbuf{}", ctx.stage_name, binding.U32())};
const auto cast{ctx.profile.has_gl_cbuf_ftou_bug ? "" : "ftou"};
if (offset.IsImmediate()) { if (offset.IsImmediate()) {
static constexpr u32 cbuf_size{0x10000}; static constexpr u32 cbuf_size{0x10000};
const u32 u32_offset{offset.U32()}; const u32 u32_offset{offset.U32()};
@ -145,26 +152,26 @@ void EmitGetCbufU32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding
return; return;
} }
if (u32_offset % 2 == 0) { if (u32_offset % 2 == 0) {
ctx.AddU32x2("{}=ftou({}[{}].{}{});", inst, cbuf, u32_offset / 16, ctx.AddU32x2("{}={}({}[{}].{}{});", inst, cast, cbuf, u32_offset / 16,
OffsetSwizzle(u32_offset), OffsetSwizzle(u32_offset + 4)); OffsetSwizzle(u32_offset), OffsetSwizzle(u32_offset + 4));
} else { } else {
ctx.AddU32x2("{}=uvec2(ftou({}[{}].{}),ftou({}[{}].{}));", inst, cbuf, u32_offset / 16, ctx.AddU32x2("{}=uvec2({}({}[{}].{}),{}({}[{}].{}));", inst, cast, cbuf,
OffsetSwizzle(u32_offset), cbuf, (u32_offset + 4) / 16, u32_offset / 16, OffsetSwizzle(u32_offset), cast, cbuf,
OffsetSwizzle(u32_offset + 4)); (u32_offset + 4) / 16, OffsetSwizzle(u32_offset + 4));
} }
return; return;
} }
const auto offset_var{ctx.var_alloc.Consume(offset)}; const auto offset_var{ctx.var_alloc.Consume(offset)};
if (!ctx.profile.has_gl_component_indexing_bug) { if (!ctx.profile.has_gl_component_indexing_bug) {
ctx.AddU32x2("{}=uvec2(ftou({}[{}>>4][({}>>2)%4]),ftou({}[({}+4)>>4][(({}+4)>>2)%4]));", ctx.AddU32x2("{}=uvec2({}({}[{}>>4][({}>>2)%4]),{}({}[({}+4)>>4][(({}+4)>>2)%4]));", inst,
inst, cbuf, offset_var, offset_var, cbuf, offset_var, offset_var); cast, cbuf, offset_var, offset_var, cast, cbuf, offset_var, offset_var);
return; return;
} }
const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32x2)}; const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32x2)};
const auto cbuf_offset{fmt::format("{}>>2", offset_var)}; const auto cbuf_offset{fmt::format("{}>>2", offset_var)};
for (u32 swizzle = 0; swizzle < 4; ++swizzle) { for (u32 swizzle = 0; swizzle < 4; ++swizzle) {
ctx.Add("if(({}&3)=={}){}=uvec2(ftou({}[{}>>4].{}),ftou({}[({}+4)>>4].{}));", cbuf_offset, ctx.Add("if(({}&3)=={}){}=uvec2({}({}[{}>>4].{}),{}({}[({}+4)>>4].{}));", cbuf_offset,
swizzle, ret, cbuf, offset_var, "xyzw"[swizzle], cbuf, offset_var, swizzle, ret, cast, cbuf, offset_var, "xyzw"[swizzle], cast, cbuf, offset_var,
"xyzw"[(swizzle + 1) % 4]); "xyzw"[(swizzle + 1) % 4]);
} }
} }

View File

@ -428,9 +428,10 @@ void EmitContext::DefineConstantBuffers(Bindings& bindings) {
return; return;
} }
for (const auto& desc : info.constant_buffer_descriptors) { for (const auto& desc : info.constant_buffer_descriptors) {
header += fmt::format( const auto cbuf_type{profile.has_gl_cbuf_ftou_bug ? "uvec4" : "vec4"};
"layout(std140,binding={}) uniform {}_cbuf_{}{{vec4 {}_cbuf{}[{}];}};", header += fmt::format("layout(std140,binding={}) uniform {}_cbuf_{}{{{} {}_cbuf{}[{}];}};",
bindings.uniform_buffer, stage_name, desc.index, stage_name, desc.index, 4 * 1024); bindings.uniform_buffer, stage_name, desc.index, cbuf_type,
stage_name, desc.index, 4 * 1024);
bindings.uniform_buffer += desc.count; bindings.uniform_buffer += desc.count;
} }
} }

View File

@ -65,6 +65,8 @@ struct Profile {
bool has_gl_component_indexing_bug{}; bool has_gl_component_indexing_bug{};
/// The precise type qualifier is broken in the fragment stage of some drivers /// The precise type qualifier is broken in the fragment stage of some drivers
bool has_gl_precise_bug{}; bool has_gl_precise_bug{};
/// Some drivers do not properly support floatBitsToUint when used on cbufs
bool has_gl_cbuf_ftou_bug{};
/// Ignores SPIR-V ordered vs unordered using GLSL semantics /// Ignores SPIR-V ordered vs unordered using GLSL semantics
bool ignore_nan_fp_comparisons{}; bool ignore_nan_fp_comparisons{};

View File

@ -182,17 +182,12 @@ Device::Device() {
shader_backend = Settings::ShaderBackend::GLSL; shader_backend = Settings::ShaderBackend::GLSL;
} }
if (shader_backend == Settings::ShaderBackend::GLSL && is_nvidia && if (shader_backend == Settings::ShaderBackend::GLSL && is_nvidia) {
!Settings::values.renderer_debug) {
const std::string_view driver_version = version.substr(13); const std::string_view driver_version = version.substr(13);
const int version_major = const int version_major =
std::atoi(driver_version.substr(0, driver_version.find(".")).data()); std::atoi(driver_version.substr(0, driver_version.find(".")).data());
if (version_major >= 495) { if (version_major >= 495) {
LOG_WARNING(Render_OpenGL, "NVIDIA drivers 495 and later causes significant problems " has_cbuf_ftou_bug = true;
"with yuzu. Forcing GLASM as a mitigation.");
shader_backend = Settings::ShaderBackend::GLASM;
use_assembly_shaders = true;
} }
} }

View File

@ -152,6 +152,10 @@ public:
return need_fastmath_off; return need_fastmath_off;
} }
bool HasCbufFtouBug() const {
return has_cbuf_ftou_bug;
}
Settings::ShaderBackend GetShaderBackend() const { Settings::ShaderBackend GetShaderBackend() const {
return shader_backend; return shader_backend;
} }
@ -200,6 +204,7 @@ private:
bool has_sparse_texture_2{}; bool has_sparse_texture_2{};
bool warp_size_potentially_larger_than_guest{}; bool warp_size_potentially_larger_than_guest{};
bool need_fastmath_off{}; bool need_fastmath_off{};
bool has_cbuf_ftou_bug{};
std::string vendor_name; std::string vendor_name;
}; };

View File

@ -214,6 +214,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
.has_broken_fp16_float_controls = false, .has_broken_fp16_float_controls = false,
.has_gl_component_indexing_bug = device.HasComponentIndexingBug(), .has_gl_component_indexing_bug = device.HasComponentIndexingBug(),
.has_gl_precise_bug = device.HasPreciseBug(), .has_gl_precise_bug = device.HasPreciseBug(),
.has_gl_cbuf_ftou_bug = device.HasCbufFtouBug(),
.ignore_nan_fp_comparisons = true, .ignore_nan_fp_comparisons = true,
.gl_max_compute_smem_size = device.GetMaxComputeSharedMemorySize(), .gl_max_compute_smem_size = device.GetMaxComputeSharedMemorySize(),
}, },