VideoCommon: move ApplyDriverBugs for the normal pipeline out into a utility function, add a way to hash the pipeline (using the vertex declaration instead of the native vertex format)

This commit is contained in:
iwubcode
2025-11-23 11:04:24 -06:00
parent 989ecca235
commit 8016e2cfbd
5 changed files with 200 additions and 132 deletions
+2 -132
View File
@@ -15,6 +15,7 @@
#include "VideoCommon/DriverDetails.h"
#include "VideoCommon/FramebufferManager.h"
#include "VideoCommon/FramebufferShaderGen.h"
#include "VideoCommon/PipelineUtils.h"
#include "VideoCommon/Present.h"
#include "VideoCommon/Statistics.h"
#include "VideoCommon/VertexLoaderManager.h"
@@ -604,141 +605,10 @@ AbstractPipelineConfig ShaderCache::GetGXPipelineConfig(
return config;
}
/// Edits the UID based on driver bugs and other special configurations
static GXPipelineUid ApplyDriverBugs(const GXPipelineUid& in)
{
GXPipelineUid out;
// TODO: static_assert(std::is_trivially_copyable_v<GXPipelineUid>);
// GXPipelineUid is not trivially copyable because RasterizationState and BlendingState aren't
// either, but we can pretend it is for now. This will be improved after PR #10848 is finished.
memcpy(static_cast<void*>(&out), static_cast<const void*>(&in), sizeof(out)); // copy padding
pixel_shader_uid_data* ps = out.ps_uid.GetUidData();
BlendingState& blend = out.blending_state;
if (ps->ztest == EmulatedZ::ForcedEarly && !out.depth_state.update_enable)
{
// No need to force early depth test if you're not writing z
ps->ztest = EmulatedZ::Early;
}
// If framebuffer fetch is available, we can emulate logic ops in the fragment shader
// and don't need the below blend approximation
if (blend.logic_op_enable && !g_backend_info.bSupportsLogicOp &&
!g_backend_info.bSupportsFramebufferFetch)
{
if (!blend.LogicOpApproximationIsExact())
WARN_LOG_FMT(VIDEO,
"Approximating logic op with blending, this will produce incorrect rendering.");
if (blend.LogicOpApproximationWantsShaderHelp())
{
ps->emulate_logic_op_with_blend = true;
ps->logic_op_mode = static_cast<u32>(blend.logic_mode.Value());
}
blend.ApproximateLogicOpWithBlending();
}
const bool benefits_from_ps_dual_source_off =
(!g_backend_info.bSupportsDualSourceBlend && g_backend_info.bSupportsFramebufferFetch) ||
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING);
if (benefits_from_ps_dual_source_off && !blend.RequiresDualSrc())
{
// Only use dual-source blending when required on drivers that don't support it very well.
ps->no_dual_src = true;
blend.use_dual_src = false;
}
if (g_backend_info.bSupportsFramebufferFetch)
{
bool fbfetch_blend = false;
if ((DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z) ||
!g_backend_info.bSupportsEarlyZ) &&
ps->ztest == EmulatedZ::ForcedEarly)
{
ps->ztest = EmulatedZ::EarlyWithFBFetch;
fbfetch_blend |= static_cast<bool>(out.blending_state.blend_enable);
ps->no_dual_src = true;
}
fbfetch_blend |= blend.logic_op_enable && !g_backend_info.bSupportsLogicOp;
fbfetch_blend |= blend.use_dual_src && !g_backend_info.bSupportsDualSourceBlend;
if (fbfetch_blend)
{
ps->no_dual_src = true;
if (blend.logic_op_enable)
{
ps->logic_op_enable = true;
ps->logic_op_mode = static_cast<u32>(blend.logic_mode.Value());
blend.logic_op_enable = false;
}
if (blend.blend_enable)
{
ps->blend_enable = true;
ps->blend_src_factor = blend.src_factor;
ps->blend_src_factor_alpha = blend.src_factor_alpha;
ps->blend_dst_factor = blend.dst_factor;
ps->blend_dst_factor_alpha = blend.dst_factor_alpha;
ps->blend_subtract = blend.subtract;
ps->blend_subtract_alpha = blend.subtract_alpha;
blend.blend_enable = false;
}
}
}
// force dual src off if we can't support it
if (!g_backend_info.bSupportsDualSourceBlend)
{
ps->no_dual_src = true;
blend.use_dual_src = false;
}
if (ps->ztest == EmulatedZ::ForcedEarly && !g_backend_info.bSupportsEarlyZ)
{
// These things should be false
ASSERT(!ps->zfreeze);
// ZCOMPLOC HACK:
// The only way to emulate alpha test + early-z is to force early-z in the shader.
// As this isn't available on all drivers and as we can't emulate this feature otherwise,
// we are only able to choose which one we want to respect more.
// Tests seem to have proven that writing depth even when the alpha test fails is more
// important that a reliable alpha test, so we just force the alpha test to always succeed.
// At least this seems to be less buggy.
ps->ztest = EmulatedZ::EarlyWithZComplocHack;
}
if (g_ActiveConfig.UseVSForLinePointExpand() &&
(out.rasterization_state.primitive == PrimitiveType::Points ||
out.rasterization_state.primitive == PrimitiveType::Lines))
{
// All primitives are expanded to triangles in the vertex shader
vertex_shader_uid_data* vs = out.vs_uid.GetUidData();
const PortableVertexDeclaration& decl = out.vertex_format->GetVertexDeclaration();
vs->position_has_3_elems = decl.position.components >= 3;
vs->texcoord_elem_count = 0;
for (int i = 0; i < 8; i++)
{
if (decl.texcoords[i].enable)
{
ASSERT(decl.texcoords[i].components <= 3);
vs->texcoord_elem_count |= decl.texcoords[i].components << (i * 2);
}
}
out.vertex_format = nullptr;
if (out.rasterization_state.primitive == PrimitiveType::Points)
vs->vs_expand = VSExpand::Point;
else
vs->vs_expand = VSExpand::Line;
PrimitiveType prim = g_backend_info.bSupportsPrimitiveRestart ? PrimitiveType::TriangleStrip :
PrimitiveType::Triangles;
out.rasterization_state.primitive = prim;
out.gs_uid.GetUidData()->primitive_type = static_cast<u32>(prim);
}
return out;
}
std::optional<AbstractPipelineConfig>
ShaderCache::GetGXPipelineConfig(const GXPipelineUid& config_in)
{
GXPipelineUid config = ApplyDriverBugs(config_in);
GXPipelineUid config = VideoCommon::ApplyDriverBugs(config_in);
const AbstractShader* vs;
auto vs_iter = m_vs_cache.shader_map.find(config.vs_uid);
if (vs_iter != m_vs_cache.shader_map.end() && !vs_iter->second.pending)