raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 19:07:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

graphics: fix PS5 NGG vertex and instance draw offsets

Browse Source 
Refs: https://github.com/KytyPS5/KytyPS5/issues/130
This commit is contained in:
nmzik
2026-09-03 21:52:02 +02:00
parent 5ffcd28ca4
commit fb4d97e8a9
10 changed files with 147 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/guest_gpu/command_processor/commandProcessor.h
+2 -2
View File
@@ -76,7 +76,7 @@ public:
void DrawIndex(uint32_t index_count, const void* index_addr, uint32_t flags, uint32_t type,
uint32_t instance_count = 0, const void* object_ids = nullptr,
uint32_t render_target_slice_offset = 0, int32_t vertex_offset_add = 0,
uint32_t first_instance = 0);
uint32_t first_instance = 0, bool offsets_from_indirect_args = false);
void DrawIndexOffset(uint32_t index_offset, uint32_t index_count, uint32_t flags);
void DrawIndexAuto(uint32_t index_count, uint32_t flags,
uint32_t render_target_slice_offset = 0);
@@ -151,7 +151,7 @@ private:
void SuspendPm4();
void SubmitNonIndexedDraw(uint32_t vertex_count, uint32_t flags,
uint32_t render_target_slice_offset, uint32_t first_vertex,
uint32_t first_instance);
uint32_t first_instance, bool offsets_from_indirect_args = false);
CommandScheduler& GetScheduler() const { return m_renderer.GetCommandScheduler(); }
CommandBuffer& CurrentBuffer() { return GetScheduler().Current(); }
src/graphics/guest_gpu/graphicsRun.cpp
+15 -12
View File
@@ -850,7 +850,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
void CommandProcessor::DrawIndex(uint32_t index_count, const void* index_addr, uint32_t flags,
uint32_t type, uint32_t instance_count, const void* object_ids,
uint32_t render_target_slice_offset, int32_t vertex_offset_add,
uint32_t first_instance) {
uint32_t first_instance, bool offsets_from_indirect_args) {
CheckBuffer();
if (instance_count == 0) {
@@ -868,9 +868,10 @@ void CommandProcessor::DrawIndex(uint32_t index_count, const void* index_addr, u
"\n",
vertex_offset_add, first_instance);
}
m_renderer.GetRenderExecutor().DrawIndex(
m_submit_id, CurrentBuffer(), m_index_type_and_size, index_count, index_addr, flags, type,
instance_count, render_target_slice_offset, vertex_offset_add, first_instance);
m_renderer.GetRenderExecutor().DrawIndex(m_submit_id, CurrentBuffer(), m_index_type_and_size,
index_count, index_addr, flags, type, instance_count,
render_target_slice_offset, vertex_offset_add,
first_instance, offsets_from_indirect_args);
}
void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_count,
@@ -929,7 +930,7 @@ void CommandProcessor::DrawIndirect(uint32_t data_offset, uint32_t draw_initiato
}
m_num_instances = args.instance_count;
SubmitNonIndexedDraw(args.vertex_count_per_instance, 0, 0, args.start_vertex_location,
args.start_instance_location);
args.start_instance_location, true);
return;
}
@@ -971,7 +972,7 @@ void CommandProcessor::DrawIndirect(uint32_t data_offset, uint32_t draw_initiato
m_num_instances = args.instance_count;
DrawIndex(index_count, index_addr, 0, 1, args.instance_count, nullptr, 0,
static_cast<int32_t>(args.base_vertex_location), args.start_instance_location);
static_cast<int32_t>(args.base_vertex_location), args.start_instance_location, true);
}
void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
@@ -1029,7 +1030,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
}
m_num_instances = args->instance_count;
SubmitNonIndexedDraw(args->vertex_count_per_instance, 0, 0, args->start_vertex_location,
args->start_instance_location);
args->start_instance_location, true);
continue;
}
@@ -1072,7 +1073,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
m_num_instances = args->instance_count;
DrawIndex(index_count, index_addr, 0, 1, args->instance_count, nullptr, 0,
static_cast<int32_t>(args->base_vertex_location), args->start_instance_location);
static_cast<int32_t>(args->base_vertex_location), args->start_instance_location,
true);
}
}
@@ -1170,12 +1172,13 @@ void CommandProcessor::DrawIndexAuto(uint32_t index_count, uint32_t flags,
void CommandProcessor::SubmitNonIndexedDraw(uint32_t vertex_count, uint32_t flags,
uint32_t render_target_slice_offset,
uint32_t first_vertex, uint32_t first_instance) {
uint32_t first_vertex, uint32_t first_instance,
bool offsets_from_indirect_args) {
CheckBuffer();
m_renderer.GetRenderExecutor().DrawAuto(m_submit_id, CurrentBuffer(), vertex_count, flags,
render_target_slice_offset, m_num_instances,
first_vertex, first_instance);
m_renderer.GetRenderExecutor().DrawAuto(
m_submit_id, CurrentBuffer(), vertex_count, flags, render_target_slice_offset,
m_num_instances, first_vertex, first_instance, offsets_from_indirect_args);
}
void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
src/graphics/host_gpu/renderer/render.h
+4 -2
View File
@@ -132,10 +132,12 @@ public:
void DrawIndex(uint64_t submit_id, CommandBuffer& buffer, uint32_t index_type_and_size,
uint32_t index_count, const void* index_addr, uint32_t flags, uint32_t type,
uint32_t instance_count = 1, uint32_t render_target_slice_offset = 0,
int32_t vertex_offset_add = 0, uint32_t first_instance = 0);
int32_t vertex_offset_add = 0, uint32_t first_instance = 0,
bool offsets_from_indirect_args = false);
void DrawAuto(uint64_t submit_id, CommandBuffer& buffer, uint32_t index_count, uint32_t flags,
uint32_t render_target_slice_offset = 0, uint32_t instance_count = 1,
uint32_t first_vertex = 0, uint32_t first_instance = 0);
uint32_t first_vertex = 0, uint32_t first_instance = 0,
bool offsets_from_indirect_args = false);
void DispatchDirect(uint64_t submit_id, CommandBuffer& buffer, uint32_t thread_group_x,
uint32_t thread_group_y, uint32_t thread_group_z, uint32_t mode);
src/graphics/host_gpu/renderer/renderDraw.cpp
+50 -15
View File
@@ -64,6 +64,25 @@ int32_t ResolveVertexOffset(uint32_t index_offset, const ShaderVertexInputInfo&
return 0;
}
uint32_t ResolveInstanceOffset(const ShaderVertexInputInfo& vs_input_info) {
if (!vs_input_info.fetch_embedded) {
return 0;
}
EXIT_IF(!vs_input_info.stage);
const auto& program = *vs_input_info.stage.program;
const auto& resources = vs_input_info.stage.resources;
if (program.info.instance_offset_sgpr >= static_cast<int32_t>(program.user_data_base)) {
const auto index =
static_cast<uint32_t>(program.info.instance_offset_sgpr) - program.user_data_base;
if (index < resources.user_data.size()) {
return resources.user_data[index];
}
}
return 0;
}
static std::atomic<uint32_t> g_draw_state_log_count = 0;
static std::atomic<uint32_t> g_draw_input_log_count = 0;
static std::atomic<uint32_t> g_mrt_state_log_count = 0;
@@ -508,6 +527,7 @@ struct DrawCallInfo {
uint32_t flags = 0;
uint32_t instance_count = 0;
uint32_t first_instance = 0;
bool offsets_from_indirect_args = false;
};
static bool ResolveDccAttachmentClear(TextureCache& cache, const RenderColorInfo& target,
@@ -747,6 +767,7 @@ struct DrawEmitInfo {
bool indexed = false;
int32_t vertex_offset = 0;
uint32_t first_vertex = 0;
uint32_t first_instance = 0;
};
struct DrawIndexBufferSource {
@@ -1124,10 +1145,10 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
case Prospero::PrimitiveType::kRectList:
if (emit.indexed) {
vk_buffer.drawIndexed(draw.index_count, draw.instance_count, 0, emit.vertex_offset,
draw.first_instance);
emit.first_instance);
} else {
vk_buffer.draw(draw.index_count, draw.instance_count, emit.first_vertex,
draw.first_instance);
emit.first_instance);
}
break;
case Prospero::PrimitiveType::kRectListLegacy:
@@ -1136,17 +1157,17 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
}
// Sarah
EXIT_NOT_IMPLEMENTED(!(draw.index_count == 3 && vs_input_info.buffers_num == 0));
vk_buffer.draw(4, draw.instance_count, emit.first_vertex, draw.first_instance);
vk_buffer.draw(4, draw.instance_count, emit.first_vertex, emit.first_instance);
break;
case Prospero::PrimitiveType::kQuadListLegacy:
EXIT_NOT_IMPLEMENTED((draw.index_count & 0x3u) != 0);
for (uint32_t i = 0; i < draw.index_count; i += 4) {
if (emit.indexed) {
vk_buffer.drawIndexed(4, draw.instance_count, i, emit.vertex_offset,
draw.first_instance);
emit.first_instance);
} else {
vk_buffer.draw(4, draw.instance_count, i + emit.first_vertex,
draw.first_instance);
emit.first_instance);
}
}
break;
@@ -1242,7 +1263,8 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
uint32_t index_type_and_size, uint32_t index_count,
const void* index_addr, uint32_t flags, uint32_t type,
uint32_t instance_count, uint32_t render_target_slice_offset,
int32_t vertex_offset_add, uint32_t first_instance) {
int32_t vertex_offset_add, uint32_t first_instance,
bool offsets_from_indirect_args) {
KYTY_PROFILER_FUNCTION();
EXIT_IF(buffer.IsInvalid());
@@ -1328,7 +1350,8 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
EXIT_NOT_IMPLEMENTED(type != 1);
const DrawCallInfo draw {"DrawIndex", CommandBufferDebugOp::DrawIndex,
index_count, flags,
instance_count, first_instance};
instance_count, first_instance,
offsets_from_indirect_args};
std::vector<uint16_t> expanded_indices;
if (expand_index8_to_u16) {
EXIT_NOT_IMPLEMENTED(index_addr == nullptr);
@@ -1359,11 +1382,16 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
LogDrawStateIfNeeded(buffer, draw, state, true, false, index_type_and_size, index_addr);
const auto vertex_offset =
ResolveVertexOffset(ucfg.GetIndexOffset(), state.vs_input_info) + vertex_offset_add;
draw.offsets_from_indirect_args
? vertex_offset_add
: ResolveVertexOffset(ucfg.GetIndexOffset(), state.vs_input_info) + vertex_offset_add;
DrawEmitInfo emit {};
emit.indexed = true;
emit.vertex_offset = vertex_offset;
emit.indexed = true;
emit.vertex_offset = vertex_offset;
emit.first_instance = draw.offsets_from_indirect_args
? draw.first_instance
: ResolveInstanceOffset(state.vs_input_info);
ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source,
primitive_restart, true, true, false);
@@ -1374,7 +1402,7 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, uint32_t index_count,
uint32_t flags, uint32_t render_target_slice_offset,
uint32_t instance_count, uint32_t first_vertex,
uint32_t first_instance) {
uint32_t first_instance, bool offsets_from_indirect_args) {
KYTY_PROFILER_FUNCTION();
EXIT_IF(buffer.IsInvalid());
@@ -1426,7 +1454,8 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, uint32_
EXIT_NOT_IMPLEMENTED(flags != 0);
const DrawCallInfo draw {"DrawIndexAuto", CommandBufferDebugOp::DrawIndexAuto,
index_count, flags,
instance_count, first_instance};
instance_count, first_instance,
offsets_from_indirect_args};
DrawRenderState state {};
if (!PrepareDrawRenderState(submit_id, buffer, draw, render_target_slice_offset, false,
@@ -1460,10 +1489,16 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, uint32_
ucfg.GetPrimType() == Prospero::PrimitiveType::kRectListLegacy, 0,
nullptr);
const auto vertex_offset = ResolveVertexOffset(ucfg.GetIndexOffset(), state.vs_input_info) +
static_cast<int32_t>(first_vertex);
const auto vertex_offset =
draw.offsets_from_indirect_args
? static_cast<int32_t>(first_vertex)
: ResolveVertexOffset(ucfg.GetIndexOffset(), state.vs_input_info) +
static_cast<int32_t>(first_vertex);
DrawEmitInfo emit {};
emit.first_vertex = static_cast<uint32_t>(vertex_offset);
emit.first_vertex = static_cast<uint32_t>(vertex_offset);
emit.first_instance = draw.offsets_from_indirect_args
? draw.first_instance
: ResolveInstanceOffset(state.vs_input_info);
DrawIndexBufferSource index_source {};
ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source, false, false,
src/graphics/host_gpu/renderer/renderDraw.h
+3 -2
View File
@@ -7,8 +7,9 @@ namespace Libs::Graphics {
struct ShaderVertexInputInfo;
[[nodiscard]] int32_t ResolveVertexOffset(uint32_t index_offset,
const ShaderVertexInputInfo& vs_input_info);
[[nodiscard]] int32_t ResolveVertexOffset(uint32_t index_offset,
const ShaderVertexInputInfo& vs_input_info);
[[nodiscard]] uint32_t ResolveInstanceOffset(const ShaderVertexInputInfo& vs_input_info);
} // namespace Libs::Graphics
src/graphics/shader/recompiler/ShaderRecompiler.cpp
+29 -13
View File
@@ -234,8 +234,10 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program& decoded
uint32_t wave_size) {
EmbeddedFetchData data;
data.loads.reserve(input_info->resources_num);
int32_t offset_candidate = -1;
bool offset_conflict = false;
int32_t vertex_offset_candidate = -1;
int32_t instance_offset_candidate = -1;
bool vertex_offset_conflict = false;
bool instance_offset_conflict = false;
const int shift_regs = 8;
const int attrib_reg = input_info->fetch_attrib_reg + shift_regs;
@@ -263,26 +265,34 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program& decoded
for (const auto& inst: decoded.instructions) {
// Fetch shaders accumulate the draw's vertex offset in v0. The PS5 NGG ABI
// seeds S_NGG_VERTEX_INDEX in v5 and applies the same offset there before fetching.
// seeds S_NGG_VERTEX_INDEX in v5 and S_NGG_INSTANCE_INDEX in v8, then applies the
// corresponding direct-draw offsets before fetching.
const bool vertex_index_accumulator =
IsDecodedVgpr(inst.dst) &&
(inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == 5));
const bool instance_index_accumulator =
IsDecodedVgpr(inst.dst) && (inst.dst.reg == (user_data_base == 8 ? 8u : 3u));
uint32_t sad_zero = 0;
const bool vertex_offset_add =
vertex_index_accumulator && IsDecodedSgpr(inst.src0) &&
const bool index_offset_add =
(vertex_index_accumulator || instance_index_accumulator) &&
IsDecodedSgpr(inst.src0) &&
((inst.opcode == Decoder::Opcode::V_ADD_I32 && IsDecodedVgpr(inst.src1) &&
inst.src1.reg == inst.dst.reg) ||
(user_data_base == 8 && inst.dst.reg == 5 &&
(user_data_base == 8 && (inst.dst.reg == 5 || inst.dst.reg == 8) &&
inst.opcode == Decoder::Opcode::V_SAD_U32 && IsDecodedVgpr(inst.src2) &&
inst.src2.reg == inst.dst.reg &&
TryDecodedOperandConstant(sgprs, inst.src1, sad_zero) && sad_zero == 0));
if (data.loads.empty() && vertex_offset_add) {
if (data.loads.empty() && index_offset_add) {
const auto reg = DecodedSgprReg(inst.src0);
if (reg >= user_data_base && reg - user_data_base < user_data_count) {
if (offset_candidate >= 0 && offset_candidate != static_cast<int32_t>(reg)) {
offset_conflict = true;
auto& candidate = vertex_index_accumulator ? vertex_offset_candidate
: instance_offset_candidate;
auto& conflict = vertex_index_accumulator ? vertex_offset_conflict
: instance_offset_conflict;
if (candidate >= 0 && candidate != static_cast<int32_t>(reg)) {
conflict = true;
} else {
offset_candidate = static_cast<int32_t>(reg);
candidate = static_cast<int32_t>(reg);
}
}
}
@@ -441,8 +451,13 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program& decoded
load.attrib_id = buffer.attrib_id;
load.components = DecodedDstSize(inst);
load.prolog_loads = buffer.prolog_loads;
if (data.loads.empty() && !offset_conflict) {
data.vertex_offset_sgpr = offset_candidate;
if (data.loads.empty()) {
if (!vertex_offset_conflict) {
data.vertex_offset_sgpr = vertex_offset_candidate;
}
if (!instance_offset_conflict) {
data.instance_offset_sgpr = instance_offset_candidate;
}
}
data.loads.push_back(load);
}
@@ -615,7 +630,8 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
IR::TrackResources(ir);
IR::EliminateDeadCode(ir.blocks);
if (options.stage == ShaderType::Vertex) {
ir.info.vertex_offset_sgpr = embedded_fetch.vertex_offset_sgpr;
ir.info.vertex_offset_sgpr = embedded_fetch.vertex_offset_sgpr;
ir.info.instance_offset_sgpr = embedded_fetch.instance_offset_sgpr;
}
TranslateResult result;
src/graphics/shader/recompiler/frontend/translate/Translate.cpp
+2 -1
View File
@@ -950,7 +950,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
result.cfg_failure_kind = options.cfg_failure_kind;
result.fallback_reason = options.fallback_reason;
if (options.embedded_fetch != nullptr) {
result.info.vertex_offset_sgpr = options.embedded_fetch->vertex_offset_sgpr;
result.info.vertex_offset_sgpr = options.embedded_fetch->vertex_offset_sgpr;
result.info.instance_offset_sgpr = options.embedded_fetch->instance_offset_sgpr;
}
uint32_t vector_limit = 1u;
src/graphics/shader/recompiler/frontend/translate/Translate.h
+2 -1
View File
@@ -19,7 +19,8 @@ struct EmbeddedFetchLoad {
struct EmbeddedFetchPlan {
std::vector<EmbeddedFetchLoad> loads;
int32_t vertex_offset_sgpr = -1;
int32_t vertex_offset_sgpr = -1;
int32_t instance_offset_sgpr = -1;
};
struct TranslateOptions {
src/graphics/shader/recompiler/ir/ShaderIR.h
+1
View File
@@ -418,6 +418,7 @@ struct ShaderInfo {
std::vector<StageOutput> outputs;
std::array<uint8_t, 32> vertex_fetch_components {};
int32_t vertex_offset_sgpr = -1;
int32_t instance_offset_sgpr = -1;
bool has_bitwise_xor = false;
bool uses_dma = false;