archive/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-07-16 03:38:38 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions

rsx-capture: reduce capture size

Browse source 
* Dont bother capturing 'destination' blocks with no data. instead premap all main memory to ensure allocated
* Capture zcull and tile state as their compressed gcm forms
* Fix index array capturing, ignore empty sets
* hle gcm: Fix byteswaping in cellGcmSetZcull
This commit is contained in:
eladash 2018-11-30 21:06:48 -08:00 • committed by kd-11
parent 32059bfaa2
commit b48a4b6459
6 changed files with 100 additions and 400 deletions
Download patch file Download diff file Expand all files Collapse all files

276
rpcs3/Emu/RSX/Capture/rsx_capture.cpp
View file

@ -13,40 +13,11 @@ namespace rsx
{ {
namespace capture namespace capture
{ {
u32 get_io_offset(u32 offset, u32 location)
{
switch (location)
{
case CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER:
case CELL_GCM_LOCATION_MAIN:
{
if (u32 result = RSXIOMem.RealAddr(offset))
{
return offset;
}
}
case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN:
{
if (u32 result = RSXIOMem.RealAddr(0x0e000000 + offset))
{
return 0x0e000000 + offset;
}
}
default: return 0xFFFFFFFFu;
}
}
void insert_mem_block_in_map(std::unordered_set<u64>& mem_changes, frame_capture_data::memory_block&& block, frame_capture_data::memory_block_data&& data) void insert_mem_block_in_map(std::unordered_set<u64>& mem_changes, frame_capture_data::memory_block&& block, frame_capture_data::memory_block_data&& data)
{ {
u64 data_hash = 0;
if (data.data.size() > 0) if (data.data.size() > 0)
{ {
data_hash = XXH64(data.data.data(), data.data.size(), 0); u64 data_hash = XXH64(data.data.data(), data.data.size(), 0);
// using 0 to signify no block in use, so this one is 'reserved'
if (data_hash == 0)
fmt::throw_exception("Memory block data hash equal to 0");
block.size = data.data.size();
block.data_state = data_hash; block.data_state = data_hash;
auto it = frame_capture.memory_data_map.find(data_hash); auto it = frame_capture.memory_data_map.find(data_hash);
@ -58,13 +29,13 @@ namespace rsx
} }
else else
frame_capture.memory_data_map.insert(std::make_pair(data_hash, std::move(data))); frame_capture.memory_data_map.insert(std::make_pair(data_hash, std::move(data)));
}
u64 block_hash = XXH64(&block, sizeof(frame_capture_data::memory_block), 0); u64 block_hash = XXH64(&block, sizeof(frame_capture_data::memory_block), 0);
mem_changes.insert(block_hash); mem_changes.insert(block_hash);
if (frame_capture.memory_map.find(block_hash) == frame_capture.memory_map.end()) if (frame_capture.memory_map.find(block_hash) == frame_capture.memory_map.end())
frame_capture.memory_map.insert(std::make_pair(block_hash, std::move(block))); frame_capture.memory_map.insert(std::make_pair(block_hash, std::move(block)));
} }
}
void capture_draw_memory(thread* rsx) void capture_draw_memory(thread* rsx)
{ {
@ -91,7 +62,7 @@ namespace rsx
const u32 ucode_size = program_info.program_ucode_length; const u32 ucode_size = program_info.program_ucode_length;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = program_offset, block.offset = program_offset;
block.location = program_location; block.location = program_location;
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(ucode_size + program_start); block_data.data.resize(ucode_size + program_start);
@ -118,9 +89,8 @@ namespace rsx
continue; continue;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = tex.offset(); block.offset = tex.offset();
block.location = tex.location(); block.location = tex.location();
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(texSize); block_data.data.resize(texSize);
std::memcpy(block_data.data.data(), vm::base(texaddr), texSize); std::memcpy(block_data.data.data(), vm::base(texaddr), texSize);
@ -145,7 +115,7 @@ namespace rsx
continue; continue;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = tex.offset(); block.offset = tex.offset();
block.location = tex.location(); block.location = tex.location();
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(texSize); block_data.data.resize(texSize);
@ -183,12 +153,11 @@ namespace rsx
const size_t bufferSize = vertCount * vertStride + vertSize; const size_t bufferSize = vertCount * vertStride + vertSize;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = base_address; block.offset = base_address + (range.first * vertStride);
block.location = memory_location; block.location = memory_location;
block.offset = (range.first * vertStride);
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(bufferSize); block_data.data.resize(bufferSize);
std::memcpy(block_data.data.data(), vm::base(addr + block.offset), bufferSize); std::memcpy(block_data.data.data(), vm::base(addr + (range.first * vertStride)), bufferSize);
insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data)); insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
} }
while (method_registers.current_draw_clause.next()); while (method_registers.current_draw_clause.next());
@ -207,9 +176,7 @@ namespace rsx
const auto index_type = method_registers.index_type(); const auto index_type = method_registers.index_type();
// manually parse index buffer and copy vertex buffer // manually parse index buffer and copy vertex buffer
u32 min_index = 0xFFFF, max_index = 0; u32 min_index = 0xFFFFFFFF, max_index = 0;
if (index_type == index_array_type::u32)
min_index = 0xFFFFFFFF;
const bool is_primitive_restart_enabled = method_registers.restart_index_enabled(); const bool is_primitive_restart_enabled = method_registers.restart_index_enabled();
const u32 primitive_restart_index = method_registers.restart_index(); const u32 primitive_restart_index = method_registers.restart_index();
@ -225,10 +192,8 @@ namespace rsx
const size_t bufferSize = idxCount * type_size; const size_t bufferSize = idxCount * type_size;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = base_address; block.offset = base_address + (idxFirst * type_size);
block.location = memory_location; block.location = memory_location;
block.offset = (idxFirst * type_size);
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(bufferSize); block_data.data.resize(bufferSize);
std::memcpy(block_data.data.data(), vm::base(idxAddr), bufferSize); std::memcpy(block_data.data.data(), vm::base(idxAddr), bufferSize);
@ -268,11 +233,8 @@ namespace rsx
} }
while (method_registers.current_draw_clause.next()); while (method_registers.current_draw_clause.next());
if (min_index > max_index) if (min_index <= max_index)
{ {
// ignore?
}
for (u8 index = 0; index < limits::vertex_count; ++index) for (u8 index = 0; index < limits::vertex_count; ++index)
{ {
const bool enabled = !!(input_mask & (1 << index)); const bool enabled = !!(input_mask & (1 << index));
@ -285,28 +247,25 @@ namespace rsx
// vert buffer // vert buffer
const u32 vertStride = info.stride(); const u32 vertStride = info.stride();
const u32 base_address = get_vertex_offset_from_base(method_registers.vertex_data_base_offset(), info.offset() & 0x7fffffff); const u32 base_address = get_vertex_offset_from_base(method_registers.vertex_data_base_offset(), (info.offset() & 0x7fffffff));
const u32 memory_location = info.offset() >> 31; const u32 memory_location = info.offset() >> 31;
const u32 addr = get_address(base_address, memory_location); const u32 addr = get_address(base_address, memory_location);
const u32 vertSize = get_vertex_type_size_on_host(info.type(), info.size()); const u32 vertSize = get_vertex_type_size_on_host(info.type(), info.size());
const u32 bufferSize = vertStride * (max_index - min_index + 1) + vertSize; const u32 bufferSize = vertStride * (max_index - min_index + 1) + vertSize;
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = base_address; block.offset = base_address + (min_index * vertStride);
block.location = memory_location; block.location = memory_location;
block.offset = (min_index * vertStride);
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(bufferSize); block_data.data.resize(bufferSize);
std::memcpy(block_data.data.data(), vm::base(addr + block.offset), bufferSize); std::memcpy(block_data.data.data(), vm::base(addr + (min_index * vertStride)), bufferSize);
insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data)); insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
} }
} }
}
capture_display_tile_state(rsx, frame_capture.replay_commands.back()); capture_display_tile_state(rsx, frame_capture.replay_commands.back());
capture_surface_state(rsx, frame_capture.replay_commands.back());
} }
// i realize these are a slight copy pasta of the rsx_method implementations but its kinda unavoidable currently // i realize these are a slight copy pasta of the rsx_method implementations but its kinda unavoidable currently
@ -314,8 +273,6 @@ namespace rsx
{ {
const rsx::blit_engine::transfer_operation operation = method_registers.blit_engine_operation(); const rsx::blit_engine::transfer_operation operation = method_registers.blit_engine_operation();
const u16 out_x = method_registers.blit_engine_output_x();
const u16 out_y = method_registers.blit_engine_output_y();
const u16 out_w = method_registers.blit_engine_output_width(); const u16 out_w = method_registers.blit_engine_output_width();
const u16 out_h = method_registers.blit_engine_output_height(); const u16 out_h = method_registers.blit_engine_output_height();
@ -344,7 +301,7 @@ namespace rsx
const tiled_region src_region = rsx->get_tiled_address(src_offset + in_offset, src_dma & 0xf); const tiled_region src_region = rsx->get_tiled_address(src_offset + in_offset, src_dma & 0xf);
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = src_region.tile ? src_region.base : src_offset + in_offset; block.offset = src_region.tile ? src_region.base : src_offset + in_offset;
block.location = src_dma & 0xf; block.location = src_dma & 0xf;
u8* pixels_src = src_region.tile ? src_region.ptr + src_region.base : src_region.ptr; u8* pixels_src = src_region.tile ? src_region.ptr + src_region.base : src_region.ptr;
@ -361,48 +318,6 @@ namespace rsx
std::memcpy(block_data.data.data(), pixels_src, in_pitch * in_h); std::memcpy(block_data.data.data(), pixels_src, in_pitch * in_h);
insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data)); insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
// 'capture' destination to ensure memory is alloc'd and usable in replay
u32 dst_offset = 0;
u32 dst_dma = 0;
rsx::blit_engine::transfer_destination_format dst_color_format;
u32 out_pitch = 0;
u32 out_alignment = 64;
switch (method_registers.blit_engine_context_surface())
{
case blit_engine::context_surface::surface2d:
dst_dma = method_registers.blit_engine_output_location_nv3062();
dst_offset = method_registers.blit_engine_output_offset_nv3062();
dst_color_format = method_registers.blit_engine_nv3062_color_format();
out_pitch = method_registers.blit_engine_output_pitch_nv3062();
out_alignment = method_registers.blit_engine_output_alignment_nv3062();
break;
case blit_engine::context_surface::swizzle2d:
dst_dma = method_registers.blit_engine_nv309E_location();
dst_offset = method_registers.blit_engine_nv309E_offset();
dst_color_format = method_registers.blit_engine_output_format_nv309E();
break;
default: break;
}
const u32 out_bpp = (dst_color_format == rsx::blit_engine::transfer_destination_format::r5g6b5) ? 2 : 4;
const s32 out_offset = out_x * out_bpp + out_pitch * out_y;
const tiled_region dst_region = rsx->get_tiled_address(dst_offset + out_offset, dst_dma & 0xf);
frame_capture_data::memory_block blockDst;
blockDst.ioOffset = dst_region.tile ? dst_region.base : dst_offset + out_offset;
blockDst.location = dst_dma & 0xf;
if (get_io_offset(blockDst.ioOffset, blockDst.location) != -1)
{
u32 blockSize = method_registers.blit_engine_context_surface() != blit_engine::context_surface::swizzle2d ? out_pitch * out_h : out_bpp * next_pow2(out_w) * next_pow2(out_h);
blockDst.size = blockSize;
frame_capture_data::memory_block_data block_data;
insert_mem_block_in_map(replay_command.memory_state, std::move(blockDst), std::move(block_data));
}
capture_display_tile_state(rsx, replay_command); capture_display_tile_state(rsx, replay_command);
} }
@ -427,9 +342,8 @@ namespace rsx
const u8* src = (u8*)vm::base(src_addr); const u8* src = (u8*)vm::base(src_addr);
frame_capture_data::memory_block block; frame_capture_data::memory_block block;
block.ioOffset = src_offset; block.offset = src_offset;
block.location = src_dma; block.location = src_dma;
frame_capture_data::memory_block_data block_data; frame_capture_data::memory_block_data block_data;
block_data.data.resize(in_pitch * line_count); block_data.data.resize(in_pitch * line_count);
@ -441,29 +355,6 @@ namespace rsx
insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data)); insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
// we 'capture' destination mostly to ensure that the location is allocated when replaying
u32 dst_offset = method_registers.nv0039_output_offset();
u32 dst_dma = method_registers.nv0039_output_location();
u32 dst_addr = get_address(dst_offset, dst_dma);
s32 out_pitch = method_registers.nv0039_output_pitch();
if (!out_pitch)
{
out_pitch = line_length;
}
frame_capture_data::memory_block blockDst;
blockDst.ioOffset = dst_offset;
blockDst.location = dst_dma;
// check if allocated
if (get_io_offset(blockDst.ioOffset, blockDst.location) != -1)
{
frame_capture_data::memory_block_data block_data;
blockDst.size = out_pitch * line_count;
insert_mem_block_in_map(replay_command.memory_state, std::move(blockDst), std::move(block_data));
}
capture_display_tile_state(rsx, replay_command); capture_display_tile_state(rsx, replay_command);
} }
@ -489,34 +380,26 @@ namespace rsx
frame_capture_data::tile_state tilestate; frame_capture_data::tile_state tilestate;
for (u32 i = 0; i < limits::tiles_count; ++i) for (u32 i = 0; i < limits::tiles_count; ++i)
{ {
const auto& tile = rsx->tiles[i]; // Avoid byteswapping
auto tile = rsx->tiles[i].pack();
auto& tstate = tilestate.tiles[i]; auto& tstate = tilestate.tiles[i];
tstate.bank = tile.bank; tstate.tile = *reinterpret_cast<u32*>(&tile.tile);
tstate.base = tile.base; tstate.limit = *reinterpret_cast<u32*>(&tile.limit);
tstate.binded = tile.binded; tstate.pitch = rsx->tiles[i].binded ? *reinterpret_cast<u32*>(&tile.pitch) : 0;
tstate.comp = tile.comp; tstate.format = rsx->tiles[i].binded ? *reinterpret_cast<u32*>(&tile.format) : 0;
tstate.location = tile.location;
tstate.offset = tile.offset;
tstate.pitch = tile.pitch;
tstate.size = tile.size;
} }
for (u32 i = 0; i < limits::zculls_count; ++i) for (u32 i = 0; i < limits::zculls_count; ++i)
{ {
const auto& zc = rsx->zculls[i]; // Avoid byteswapping
auto zc = rsx->zculls[i].pack();
auto& zcstate = tilestate.zculls[i]; auto& zcstate = tilestate.zculls[i];
zcstate.aaFormat = zc.aaFormat; zcstate.region = *reinterpret_cast<u32*>(&zc.region);
zcstate.binded = zc.binded; zcstate.size = *reinterpret_cast<u32*>(&zc.size);
zcstate.cullStart = zc.cullStart; zcstate.start = *reinterpret_cast<u32*>(&zc.start);
zcstate.height = zc.height; zcstate.offset = *reinterpret_cast<u32*>(&zc.offset);
zcstate.offset = zc.offset; zcstate.status0 = rsx->zculls[i].binded ? *reinterpret_cast<u32*>(&zc.status0) : 0;
zcstate.sFunc = zc.sFunc; zcstate.status1 = rsx->zculls[i].binded ? *reinterpret_cast<u32*>(&zc.status1) : 0;
zcstate.sMask = zc.sMask;
zcstate.sRef = zc.sRef;
zcstate.width = zc.width;
zcstate.zcullDir = zc.zcullDir;
zcstate.zcullFormat = zc.zcullFormat;
zcstate.zFormat = zc.zFormat;
} }
const u64 tsnum = XXH64(&tilestate, sizeof(frame_capture_data::tile_state), 0); const u64 tsnum = XXH64(&tilestate, sizeof(frame_capture_data::tile_state), 0);
@ -527,104 +410,5 @@ namespace rsx
replay_command.display_buffer_state = dbnum; replay_command.display_buffer_state = dbnum;
replay_command.tile_state = tsnum; replay_command.tile_state = tsnum;
} }
// for the most part capturing this is just to make sure the iomemory is recorded/allocated and accounted for in replay
void capture_get_report(thread* rsx, frame_capture_data::replay_command& replay_command, u32 arg)
{
u32 location = 0;
u32 offset = arg & 0xffffff;
blit_engine::context_dma report_dma = method_registers.context_dma_report();
switch (report_dma)
{
// ignore regular report location
case blit_engine::context_dma::to_memory_get_report: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL; return;
case blit_engine::context_dma::report_location_main: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN; break;
case blit_engine::context_dma::memory_host_buffer: location = CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER; break;
default: return;
}
u32 addr = get_address(offset, location);
frame_capture_data::memory_block block;
block.ioOffset = offset;
block.location = location;
block.size = 16;
frame_capture_data::memory_block_data block_data;
insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
}
// This just checks the color/depth addresses and makes sure they are accounted for in io allocations
// Hopefully this works without fully having to calculate actual size
void capture_surface_state(thread* rsx, frame_capture_data::replay_command& replay_command)
{
const auto target = rsx::method_registers.surface_color_target();
u32 offset_color[] =
{
rsx::method_registers.surface_a_offset(),
rsx::method_registers.surface_b_offset(),
rsx::method_registers.surface_c_offset(),
rsx::method_registers.surface_d_offset(),
};
u32 context_dma_color[] =
{
rsx::method_registers.surface_a_dma(),
rsx::method_registers.surface_b_dma(),
rsx::method_registers.surface_c_dma(),
rsx::method_registers.surface_d_dma(),
};
auto check_add = [&replay_command](u32 offset, u32 dma) -> void
{
u32 ioOffset = get_io_offset(offset, dma);
if (ioOffset == -1)
return;
frame_capture_data::memory_block block;
block.ioOffset = offset;
block.location = dma;
block.size = 64;
frame_capture_data::memory_block_data block_data;
insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
};
for (const auto& index : utility::get_rtt_indexes(target))
{
check_add(offset_color[index], context_dma_color[index]);
}
check_add(rsx::method_registers.surface_z_offset(), rsx::method_registers.surface_z_dma());
}
void capture_inline_transfer(thread* rsx, frame_capture_data::replay_command& replay_command, u32 idx, u32 arg)
{
const u16 x = method_registers.nv308a_x();
const u16 y = method_registers.nv308a_y();
const u32 pixel_offset = (method_registers.blit_engine_output_pitch_nv3062() * y) + (x << 2);
const u32 addr_offset = method_registers.blit_engine_output_offset_nv3062() + pixel_offset + idx * 4;
// just need to capture dst for allocation later if in iomem
const u32 memory_location = method_registers.blit_engine_output_location_nv3062();
const u32 ioOffset = get_io_offset(addr_offset, memory_location);
if (ioOffset == -1)
return;
frame_capture_data::memory_block block;
block.ioOffset = addr_offset;
block.location = memory_location;
block.size = 4;
frame_capture_data::memory_block_data block_data;
insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
}
} }
} }

3
rpcs3/Emu/RSX/Capture/rsx_capture.h
View file

@ -10,8 +10,5 @@ namespace rsx
void capture_image_in(thread* rsx, frame_capture_data::replay_command& replay_command); void capture_image_in(thread* rsx, frame_capture_data::replay_command& replay_command);
void capture_buffer_notify(thread* rsx, frame_capture_data::replay_command& replay_command); void capture_buffer_notify(thread* rsx, frame_capture_data::replay_command& replay_command);
void capture_display_tile_state(thread* rsx, frame_capture_data::replay_command& replay_command); void capture_display_tile_state(thread* rsx, frame_capture_data::replay_command& replay_command);
void capture_surface_state(thread* rsx, frame_capture_data::replay_command& replay_command);
void capture_get_report(thread* rsx, frame_capture_data::replay_command& replay_command, u32 arg);
void capture_inline_transfer(thread* rsx, frame_capture_data::replay_command& replay_command, u32 idx, u32 arg);
} }
} }

68
rpcs3/Emu/RSX/Capture/rsx_replay.cpp
View file

@ -113,16 +113,13 @@ namespace rsx
if (it == frame->memory_map.end()) if (it == frame->memory_map.end())
fmt::throw_exception("requested memory state for command not found in memory_map"); fmt::throw_exception("requested memory state for command not found in memory_map");
if (it->second.data_state != 0)
{
const auto& memblock = it->second; const auto& memblock = it->second;
auto it_data = frame->memory_data_map.find(it->second.data_state); auto it_data = frame->memory_data_map.find(it->second.data_state);
if (it_data == frame->memory_data_map.end()) if (it_data == frame->memory_data_map.end())
fmt::throw_exception("requested memory data state for command not found in memory_data_map"); fmt::throw_exception("requested memory data state for command not found in memory_data_map");
const auto& data_block = it_data->second; const auto& data_block = it_data->second;
std::memcpy(vm::base(get_address(memblock.ioOffset + memblock.offset, memblock.location)), data_block.data.data(), data_block.data.size()); std::memcpy(vm::base(get_address(memblock.offset, memblock.location)), data_block.data.data(), data_block.data.size());
}
} }
if (replay_cmd.display_buffer_state != 0 && replay_cmd.display_buffer_state != cs.display_buffer_hash) if (replay_cmd.display_buffer_state != 0 && replay_cmd.display_buffer_state != cs.display_buffer_hash)
@ -154,50 +151,22 @@ namespace rsx
const auto& tstate = it->second; const auto& tstate = it->second;
for (u32 i = 0; i < limits::tiles_count; ++i) for (u32 i = 0; i < limits::tiles_count; ++i)
{ {
const auto& tstile = tstate.tiles[i]; const auto& ti = tstate.tiles[i];
if (cs.tile_hash != 0 && memcmp(&cs.tile_state.tiles[i], &tstile, sizeof(rsx::frame_capture_data::tile_info)) == 0) if (cs.tile_hash != 0 && memcmp(&cs.tile_state.tiles[i], &ti, sizeof(rsx::frame_capture_data::tile_info)) == 0)
continue; continue;
cs.tile_state.tiles[i] = tstile; cs.tile_state.tiles[i] = ti;
sys_rsx_context_attribute(context_id, 0x300, i, (u64)ti.tile << 32 | ti.limit, (u64)ti.pitch << 32 | ti.format, 0);
GcmTileInfo t;
t.bank = tstile.bank;
t.base = tstile.base;
t.binded = tstile.binded;
t.comp = tstile.comp;
t.location = tstile.location;
t.offset = tstile.offset;
t.pitch = tstile.pitch;
t.size = tstile.size;
const auto& ti = t.pack();
sys_rsx_context_attribute(context_id, 0x300, i, (u64)ti.tile << 32 | ti.limit, t.binded ? (u64)ti.pitch << 32 | ti.format : 0, 0);
} }
for (u32 i = 0; i < limits::zculls_count; ++i) for (u32 i = 0; i < limits::zculls_count; ++i)
{ {
const auto& zctile = tstate.zculls[i]; const auto& zci = tstate.zculls[i];
if (cs.tile_hash != 0 && memcmp(&cs.tile_state.zculls[i], &zctile, sizeof(rsx::frame_capture_data::zcull_info)) == 0) if (cs.tile_hash != 0 && memcmp(&cs.tile_state.zculls[i], &zci, sizeof(rsx::frame_capture_data::zcull_info)) == 0)
continue; continue;
cs.tile_state.zculls[i] = zctile; cs.tile_state.zculls[i] = zci;
sys_rsx_context_attribute(context_id, 0x301, i, (u64)zci.region << 32 | zci.size, (u64)zci.start << 32 | zci.offset, (u64)zci.status0 << 32 | zci.status1);
GcmZcullInfo zc;
zc.aaFormat = zctile.aaFormat;
zc.binded = zctile.binded;
zc.cullStart = zctile.cullStart;
zc.height = zctile.height;
zc.offset = zctile.offset;
zc.sFunc = zctile.sFunc;
zc.sMask = zctile.sMask;
zc.sRef = zctile.sRef;
zc.width = zctile.width;
zc.zcullDir = zctile.zcullDir;
zc.zcullFormat = zctile.zcullFormat;
zc.zFormat = zctile.zFormat;
const auto& zci = zc.pack();
sys_rsx_context_attribute(context_id, 0x301, i, (u64)zci.region << 32 | zci.size, (u64)zci.start << 32 | zci.offset, zc.binded ? (u64)zci.status0 << 32 | zci.status1 : 0);
} }
cs.tile_hash = replay_cmd.tile_state; cs.tile_hash = replay_cmd.tile_state;
@ -211,23 +180,8 @@ namespace rsx
auto fifo_stops = alloc_write_fifo(context_id); auto fifo_stops = alloc_write_fifo(context_id);
// map game io // map game io
for (const auto it : frame->memory_map) if (sys_rsx_context_iomap(context_id, 0x0, user_mem_addr, 0x10000000, 0) != CELL_OK)
{ fmt::throw_exception("rsx io map failed");
const auto& memblock = it.second;
if (memblock.location == CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN)
{
// Special area for reports
if (sys_rsx_context_iomap(context_id, (memblock.ioOffset & ~0xFFFFF) + 0x0e000000, (memblock.ioOffset & ~0xFFFFF) + user_mem_addr + 0x0e000000, 0x100000, 0) != CELL_OK)
fmt::throw_exception("rsx io map failed for block");
continue;
}
if (const u32 location = memblock.location; location != CELL_GCM_LOCATION_MAIN && location != CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER)
continue;
if (sys_rsx_context_iomap(context_id, memblock.ioOffset & ~0xFFFFF, user_mem_addr + (memblock.ioOffset & ~0xFFFFF), ::align<u32>(memblock.size + memblock.offset, 0x100000), 0) != CELL_OK)
fmt::throw_exception("rsx io map failed for block");
}
while (!Emu.IsStopped()) while (!Emu.IsStopped())
{ {

73
rpcs3/Emu/RSX/Capture/rsx_replay.h
View file

@ -4,9 +4,7 @@
#include "Emu/IdManager.h" #include "Emu/IdManager.h"
#include "Emu/Cell/PPUModule.h" #include "Emu/Cell/PPUModule.h"
#include "Emu/Cell/lv2/sys_sync.h" #include "Emu/Cell/lv2/sys_sync.h"
#include "Emu/Cell/lv2/sys_ppu_thread.h"
#include "Emu/RSX/rsx_methods.h" #include "Emu/RSX/rsx_methods.h"
#include "Emu/RSX/gcm_enums.h"
#include <cereal/types/vector.hpp> #include <cereal/types/vector.hpp>
#include <cereal/types/array.hpp> #include <cereal/types/array.hpp>
@ -17,10 +15,9 @@
namespace rsx namespace rsx
{ {
constexpr u32 FRAME_CAPTURE_MAGIC = 0x52524300; // ascii 'RRC/0' constexpr u32 FRAME_CAPTURE_MAGIC = 0x52524300; // ascii 'RRC/0'
constexpr u32 FRAME_CAPTURE_VERSION = 0x3; constexpr u32 FRAME_CAPTURE_VERSION = 0x4;
struct frame_capture_data struct frame_capture_data
{ {
struct memory_block_data struct memory_block_data
{ {
std::vector<u8> data; std::vector<u8> data;
@ -34,17 +31,14 @@ namespace rsx
// simple block to hold ps3 address and data // simple block to hold ps3 address and data
struct memory_block struct memory_block
{ {
u32 ioOffset{0xFFFFFFFF}; // rsx ioOffset, -1 signifies unused u32 offset; // Offset in rsx address space
u32 offset{0}; // offset into addr/ioOffset to copy state into u32 location; // rsx memory location of the block
u32 size{0}; // size of block needed u64 data_state;
u32 location{0xFFFFFFFF}; // Location of the block in RSX memory space
u64 data_state{0}; // this can be 0, in which case its just needed as an alloc
template<typename Archive> template<typename Archive>
void serialize(Archive & ar) void serialize(Archive & ar)
{ {
ar(ioOffset);
ar(offset); ar(offset);
ar(size);
ar(location); ar(location);
ar(data_state); ar(data_state);
} }
@ -67,62 +61,41 @@ namespace rsx
} }
}; };
// same thing as gcmtileinfo
struct tile_info struct tile_info
{ {
u32 location{0}; u32 tile;
u32 offset{0}; u32 limit;
u32 size{0}; u32 pitch;
u32 pitch{0}; u32 format;
u32 comp{0};
u32 base{0};
u32 bank{0};
bool binded{false};
template<typename Archive> template<typename Archive>
void serialize(Archive & ar) void serialize(Archive & ar)
{ {
ar(location); ar(tile);
ar(offset); ar(limit);
ar(size);
ar(pitch); ar(pitch);
ar(comp); ar(format);
ar(base);
ar(bank);
ar(binded);
} }
}; };
// same thing as gcmzcullinfo
struct zcull_info struct zcull_info
{ {
u32 offset{0}; u32 region;
u32 width{0}; u32 size;
u32 height{0}; u32 start;
u32 cullStart{0}; u32 offset;
u32 zFormat{0}; u32 status0;
u32 aaFormat{0}; u32 status1;
u32 zcullDir{0};
u32 zcullFormat{0};
u32 sFunc{0};
u32 sRef{0};
u32 sMask{0};
bool binded{false};
template<typename Archive> template<typename Archive>
void serialize(Archive & ar) void serialize(Archive & ar)
{ {
ar(region);
ar(size);
ar(start);
ar(offset); ar(offset);
ar(width); ar(status0);
ar(cullStart); ar(status1);
ar(zFormat);
ar(aaFormat);
ar(zcullDir);
ar(zcullFormat);
ar(sFunc);
ar(sRef);
ar(sMask);
ar(binded);
} }
}; };

12
rpcs3/Emu/RSX/GCM.h
View file

@ -73,12 +73,12 @@ struct CellGcmReportData
struct CellGcmZcullInfo struct CellGcmZcullInfo
{ {
u32 region; be_t<u32> region;
u32 size; be_t<u32> size;
u32 start; be_t<u32> start;
u32 offset; be_t<u32> offset;
u32 status0; be_t<u32> status0;
u32 status1; be_t<u32> status1;
}; };
struct CellGcmDisplayInfo struct CellGcmDisplayInfo

8
rpcs3/Emu/RSX/RSXFIFO.cpp
View file

@ -523,21 +523,13 @@ namespace rsx
switch (reg) switch (reg)
{ {
case NV4097_GET_REPORT:
capture::capture_get_report(this, *it, value);
break;
case NV3089_IMAGE_IN: case NV3089_IMAGE_IN:
capture::capture_image_in(this, *it); capture::capture_image_in(this, *it);
break; break;
case NV0039_BUFFER_NOTIFY: case NV0039_BUFFER_NOTIFY:
capture::capture_buffer_notify(this, *it); capture::capture_buffer_notify(this, *it);
break; break;
case NV4097_CLEAR_SURFACE:
capture::capture_surface_state(this, *it);
break;
default: default:
if (reg >= NV308A_COLOR && reg < NV3089_SET_OBJECT)
capture::capture_inline_transfer(this, *it, reg - NV308A_COLOR, value);
break; break;
} }
} }