raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-28 13:56:59 +01:00
Code Issues Packages Projects Releases Wiki Activity

Translate NGG mesh prolog with guest wave32 width

Browse Source
This commit is contained in:
nmzik
2026-09-23 20:39:46 +04:00
parent 2aa0a64f63
commit 7d7267cc56
2 changed files with 15 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/frontend/translate/Translate.cpp
+10 -7
View File
@@ -1063,8 +1063,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
} }
} else if (options.stage == ShaderType::Mesh) { } else if (options.stage == ShaderType::Mesh) {
const auto& mesh = options.input_info.vertex->mesh; const auto& mesh = options.input_info.vertex->mesh;
EXIT_NOT_IMPLEMENTED(options.wave_size != 64u || mesh.primitives_per_group == 0u || EXIT_NOT_IMPLEMENTED(mesh.primitives_per_group == 0u || mesh.vertices_per_group > 64u ||
mesh.vertices_per_group > 64u || total_threads > 15u * 64u); total_threads > 15u * options.wave_size);
const auto u32 = [](uint32_t value) { return IR::U32(IR::Value(value)); }; const auto u32 = [](uint32_t value) { return IR::U32(IR::Value(value)); };
const auto draw = [&](uint32_t index) { const auto draw = [&](uint32_t index) {
return IR::U32( return IR::U32(
@@ -1089,12 +1089,15 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
entry_ir.IAdd(IR::U32(entry_ir.Emit(IR::ValueOpcode::UDiv32, entry_ir.IAdd(IR::U32(entry_ir.Emit(IR::ValueOpcode::UDiv32,
{subtract_saturate(vertices, size), step})), {subtract_saturate(vertices, size), step})),
u32(1))); u32(1)));
const auto wave = entry_ir.ShiftRightLogical(local, u32(6)); const auto wave = entry_ir.ShiftRightLogical(local, u32(options.wave_size == 32u ? 5u : 6u));
const auto wave_base = entry_ir.BitwiseAnd(local, u32(~63u)); const auto wave_base = entry_ir.BitwiseAnd(local, u32(~(options.wave_size - 1u)));
const auto vertex_count = minimum(subtract_saturate(vertices, wave_base), u32(64)); const auto vertex_count =
const auto primitive_count = minimum(subtract_saturate(primitives, wave_base), u32(64)); minimum(subtract_saturate(vertices, wave_base), u32(options.wave_size));
const auto primitive_count =
minimum(subtract_saturate(primitives, wave_base), u32(options.wave_size));
const auto wave_info = entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(wave, u32(24)), const auto wave_info = entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(wave, u32(24)),
u32(((total_threads + 63u) / 64u) << 28u)); u32(((total_threads + options.wave_size - 1u) /
options.wave_size) << 28u));
entry_ir.SetScalarReg( entry_ir.SetScalarReg(
static_cast<IR::ScalarReg>(3), static_cast<IR::ScalarReg>(3),
entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical( entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
tests/shaderCfgTests.cpp
+5 -1
View File
@@ -9885,6 +9885,7 @@ void TestMeshInputAssembly() {
uint32_t capacity, count, group, lane, width, address_low, base_vertex; uint32_t capacity, count, group, lane, width, address_low, base_vertex;
uint32_t wave_info, first, second, third, byte_offset, vertex_id; uint32_t wave_info, first, second, third, byte_offset, vertex_id;
bool fetch; bool fetch;
uint32_t wave_size = 64;
}; };
const Case cases[] = { const Case cases[] = {
{Prospero::PrimitiveType::kTriList, 14, 177, 14, 2, 2, 0x1002, UINT32_MAX, {Prospero::PrimitiveType::kTriList, 14, 177, 14, 2, 2, 0x1002, UINT32_MAX,
@@ -9931,11 +9932,14 @@ void TestMeshInputAssembly() {
0x40000c0c, 1, 0, 0, 52, 0xabcd0128, true}, 0x40000c0c, 1, 0, 0, 52, 0xabcd0128, true},
{Prospero::PrimitiveType::kPointList, 12, 265, 1, 64, 4, 0x1000, 0, {Prospero::PrimitiveType::kPointList, 12, 265, 1, 64, 4, 0x1000, 0,
0x41000000, 64, 0, 0, 304, 0, false}, 0x41000000, 64, 0, 0, 304, 0, false},
{Prospero::PrimitiveType::kTriStrip, 40, 40, 0, 32, 0, 0, 11,
0x81000608, 32, 33, 34, 0, 43, false, 32},
}; };
for (const auto &test : cases) { for (const auto &test : cases) {
ShaderVertexInputInfo input{}; ShaderVertexInputInfo input{};
auto &mesh = input.mesh; auto &mesh = input.mesh;
mesh.input_primitive = static_cast<uint32_t>(test.topology); mesh.input_primitive = static_cast<uint32_t>(test.topology);
mesh.wave_size = test.wave_size;
mesh.primitives_per_group = mesh.InputPrimitiveCount(test.capacity); mesh.primitives_per_group = mesh.InputPrimitiveCount(test.capacity);
mesh.vertices_per_group = mesh.InputVertexCount(mesh.primitives_per_group); mesh.vertices_per_group = mesh.InputVertexCount(mesh.primitives_per_group);
mesh.threads_num[0] = 256; mesh.threads_num[0] = 256;
@@ -9949,7 +9953,7 @@ void TestMeshInputAssembly() {
graph.entry_block = 0; graph.entry_block = 0;
Frontend::TranslateOptions options{}; Frontend::TranslateOptions options{};
options.stage = ShaderType::Mesh; options.stage = ShaderType::Mesh;
options.wave_size = 64; options.wave_size = test.wave_size;
options.user_data_count = 0; options.user_data_count = 0;
options.input_info.vertex = &input; options.input_info.vertex = &input;
auto program = Frontend::TranslateProgram(decoded, graph, options); auto program = Frontend::TranslateProgram(decoded, graph, options);