video_core, shader_recompiler: Improve Princess Peach: Showtime! support and performance

1. Add geometry shader support for Princess Peach: Showtime!:
   - Implement proper EmitInvocationInfo handling for geometry shaders
   - Support input topology vertex counting in all shader backends (GLASM, GLSL, SPIRV)

2. Performance optimizations:
   - Replace InputTopologyVertices switch statement with a constexpr lookup table
   - Pre-calculate vertex counts and shifts to reduce register pressure
   - Eliminate redundant calculations in shader backends

Signed-off-by: Zephyron <zephyron@citron-emu.org>
This commit is contained in:
Zephyron
2025-05-11 16:28:25 +10:00
parent 9bb2b769a4
commit 2e6bcc9ea6
4 changed files with 34 additions and 0 deletions

View File

@@ -7,6 +7,7 @@
#include "shader_recompiler/backend/glasm/glasm_emit_context.h" #include "shader_recompiler/backend/glasm/glasm_emit_context.h"
#include "shader_recompiler/frontend/ir/value.h" #include "shader_recompiler/frontend/ir/value.h"
#include "shader_recompiler/profile.h" #include "shader_recompiler/profile.h"
#include "shader_recompiler/runtime_info.h"
#include "shader_recompiler/shader_info.h" #include "shader_recompiler/shader_info.h"
namespace Shader::Backend::GLASM { namespace Shader::Backend::GLASM {
@@ -406,6 +407,12 @@ void EmitInvocationInfo(EmitContext& ctx, IR::Inst& inst) {
case Stage::TessellationEval: case Stage::TessellationEval:
ctx.Add("SHL.U {}.x,primitive.vertexcount,16;", inst); ctx.Add("SHL.U {}.x,primitive.vertexcount,16;", inst);
break; break;
case Stage::Geometry: {
// Pre-calculate the vertex count for better performance
const u32 vertex_count = InputTopologyVertices::vertices(ctx.runtime_info.input_topology) << 16;
ctx.Add("MOV.S {}.x,{};", inst, vertex_count);
break;
}
default: default:
LOG_WARNING(Shader, "(STUBBED) called"); LOG_WARNING(Shader, "(STUBBED) called");
ctx.Add("MOV.S {}.x,0x00ff0000;", inst); ctx.Add("MOV.S {}.x,0x00ff0000;", inst);

View File

@@ -426,6 +426,12 @@ void EmitInvocationInfo(EmitContext& ctx, IR::Inst& inst) {
case Stage::TessellationEval: case Stage::TessellationEval:
ctx.AddU32("{}=uint(gl_PatchVerticesIn)<<16;", inst); ctx.AddU32("{}=uint(gl_PatchVerticesIn)<<16;", inst);
break; break;
case Stage::Geometry: {
// Pre-calculate the vertex count for better performance
const u32 vertex_count = InputTopologyVertices::vertices(ctx.runtime_info.input_topology);
ctx.AddU32("{}={}u;", inst, vertex_count << 16);
break;
}
default: default:
LOG_WARNING(Shader, "(STUBBED) called"); LOG_WARNING(Shader, "(STUBBED) called");
ctx.AddU32("{}=uint(0x00ff0000);", inst); ctx.AddU32("{}=uint(0x00ff0000);", inst);

View File

@@ -549,6 +549,11 @@ Id EmitInvocationInfo(EmitContext& ctx) {
case Stage::TessellationEval: case Stage::TessellationEval:
return ctx.OpShiftLeftLogical(ctx.U32[1], ctx.OpLoad(ctx.U32[1], ctx.patch_vertices_in), return ctx.OpShiftLeftLogical(ctx.U32[1], ctx.OpLoad(ctx.U32[1], ctx.patch_vertices_in),
ctx.Const(16u)); ctx.Const(16u));
case Stage::Geometry: {
// Pre-calculate the shifted value for better performance
const u32 shifted_value = InputTopologyVertices::vertices(ctx.runtime_info.input_topology) << 16;
return ctx.Const(shifted_value);
}
default: default:
LOG_WARNING(Shader, "(STUBBED) called"); LOG_WARNING(Shader, "(STUBBED) called");
return ctx.Const(0x00ff0000u); return ctx.Const(0x00ff0000u);

View File

@@ -30,6 +30,22 @@ enum class InputTopology {
TrianglesAdjacency, TrianglesAdjacency,
}; };
namespace InputTopologyVertices {
// Lookup table for vertex counts - faster than switch statement
inline constexpr std::array<u32, 5> vertex_counts = {
1, // Points
2, // Lines
4, // LinesAdjacency
3, // Triangles
6, // TrianglesAdjacency
};
// Force compile-time evaluation when possible
inline constexpr u32 vertices(InputTopology input_topology) {
return vertex_counts[static_cast<std::size_t>(input_topology)];
}
}
enum class CompareFunction { enum class CompareFunction {
Never, Never,
Less, Less,