YouHaveTrouble/Kyty
1
0
Fork 0
mirror of https://github.com/InoriRus/Kyty.git synced 2026-08-28 05:06:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

PS5 graphics (#42)

Browse Source
This commit is contained in:
InoriRus
2022-08-16 13:38:38 +03:00
parent e110bd0f63
commit 90acd3d364
46 changed files with 12642 additions and 3962 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/emulator/src/Config.cpp
+12 -1
View File
@@ -13,6 +13,7 @@ struct Config
uint32_t screen_width = 1280;
uint32_t screen_height = 720;
bool neo = true;
bool next_gen = false;
bool vulkan_validation_enabled = false;
bool shader_validation_enabled = false;
ShaderOptimizationType shader_optimization_type = ShaderOptimizationType::None;
@@ -115,7 +116,7 @@ uint32_t GetScreenHeight()
bool IsNeo()
{
return g_config->neo;
return g_config->neo || g_config->next_gen;
}
bool VulkanValidationEnabled()
@@ -193,6 +194,16 @@ String GetPipelineDumpFolder()
return g_config->pipeline_dump_folder;
}
void SetNextGen(bool mode)
{
g_config->next_gen = mode;
}
bool IsNextGen()
{
return g_config->next_gen;
}
} // namespace Kyty::Config
#endif // KYTY_EMU_ENABLED
source/emulator/src/Graphics/Graphics.cpp
+64 -102
View File
@@ -12,6 +12,7 @@
#include "Emulator/Graphics/Objects/IndexBuffer.h"
#include "Emulator/Graphics/Objects/Label.h"
#include "Emulator/Graphics/Pm4.h"
#include "Emulator/Graphics/Shader.h"
#include "Emulator/Graphics/Tile.h"
#include "Emulator/Graphics/VideoOut.h"
#include "Emulator/Graphics/Window.h"
@@ -40,6 +41,7 @@ KYTY_SUBSYSTEM_INIT(Graphics)
LabelInit();
TileInit();
IndexBufferInit();
ShaderInit();
}
KYTY_SUBSYSTEM_UNEXPECTED_SHUTDOWN(Graphics) {}
@@ -755,12 +757,6 @@ namespace Gen5 {
LIB_NAME("Graphics5", "Graphics5");
struct ShaderRegister
{
uint32_t offset;
uint32_t value;
};
struct RegisterDefaultInfo
{
uint32_t type;
@@ -778,93 +774,6 @@ struct RegisterDefaults
uint32_t count = 0;
};
struct ShaderSharp
{
uint16_t offset_dw : 15;
uint16_t size : 1;
};
struct ShaderUserData
{
uint16_t* direct_resource_offset;
ShaderSharp* sharp_resource_offset[4];
uint16_t eud_size_dw;
uint16_t srt_size_dw;
uint16_t direct_resource_count;
uint16_t sharp_resource_count[4];
};
struct ShaderRegisterRange
{
uint16_t start;
uint16_t end;
};
struct ShaderDrawModifier
{
uint32_t enbl_start_vertex_offset : 1;
uint32_t enbl_start_index_offset : 1;
uint32_t enbl_start_instance_offset : 1;
uint32_t enbl_draw_index : 1;
uint32_t enbl_user_vgprs : 1;
uint32_t render_target_slice_offset : 3;
uint32_t fuse_draws : 1;
uint32_t compiler_flags : 23;
uint32_t is_default : 1;
uint32_t reserved : 31;
};
struct ShaderSpecialRegs
{
ShaderRegister ge_cntl;
ShaderRegister vgt_shader_stages_en;
uint32_t dispatch_modifier;
ShaderRegisterRange user_data_range;
ShaderDrawModifier draw_modifier;
ShaderRegister vgt_gs_out_prim_type;
ShaderRegister ge_user_vgpr_en;
};
struct ShaderSemantic
{
uint32_t semantic : 8;
uint32_t hardware_mapping : 8;
uint32_t size_in_elements : 4;
uint32_t is_f16 : 2;
uint32_t is_flat_shaded : 1;
uint32_t is_linear : 1;
uint32_t is_custom : 1;
uint32_t static_vb_index : 1;
uint32_t static_attribute : 1;
uint32_t reserved : 1;
uint32_t default_value : 2;
uint32_t default_value_hi : 2;
};
struct Shader
{
uint32_t file_header;
uint32_t version;
ShaderUserData* user_data;
const volatile void* code;
ShaderRegister* cx_registers;
ShaderRegister* sh_registers;
ShaderSpecialRegs* specials;
ShaderSemantic* input_semantics;
ShaderSemantic* output_semantics;
uint32_t header_size;
uint32_t shader_size;
uint32_t embedded_constant_buffer_size_dqw;
uint32_t target;
uint32_t num_input_semantics;
uint16_t scratch_size_dw_per_thread;
uint16_t num_output_semantics;
uint16_t special_sizes_bytes;
uint8_t type;
uint8_t num_cx_registers;
uint8_t num_sh_registers;
};
struct CommandBuffer
{
using Callback = KYTY_SYSV_ABI bool (*)(CommandBuffer*, uint32_t, void*);
@@ -1561,6 +1470,17 @@ int KYTY_SYSV_ABI GraphicsCreateShader(Shader** dst, void* header, const volatil
auto base = reinterpret_cast<uint64_t>(code);
printf("\t base = 0x%016" PRIx64 "\n", base);
ShaderMappedData map;
map.user_data = h->user_data;
map.input_semantics = h->input_semantics;
map.num_input_semantics = h->num_input_semantics;
ShaderMapUserData(base, map);
EXIT_NOT_IMPLEMENTED((base & 0xFFFF0000000000FFull) != 0);
if (h->type == 2 && h->num_sh_registers >= 2 && h->sh_registers[0].offset == Pm4::SPI_SHADER_PGM_LO_ES &&
h->sh_registers[1].offset == Pm4::SPI_SHADER_PGM_HI_ES)
{
@@ -1716,6 +1636,7 @@ int KYTY_SYSV_ABI GraphicsCreatePrimState(ShaderRegister* cx_regs, ShaderRegiste
return OK;
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
int KYTY_SYSV_ABI GraphicsCreateInterpolantMapping(ShaderRegister* regs, const Shader* gs, const Shader* ps)
{
PRINT_NAME();
@@ -1733,20 +1654,42 @@ int KYTY_SYSV_ABI GraphicsCreateInterpolantMapping(ShaderRegister* regs, const S
EXIT_NOT_IMPLEMENTED(sizeof(ShaderSemantic) != 4);
EXIT_NOT_IMPLEMENTED(ps != nullptr && gs->num_output_semantics != ps->num_input_semantics);
EXIT_NOT_IMPLEMENTED(ps != nullptr && ps->num_output_semantics != 0);
auto* out32 = reinterpret_cast<uint32_t*>(gs->output_semantics);
auto* in32 = (ps != nullptr ? reinterpret_cast<uint32_t*>(ps->input_semantics) : nullptr);
for (int i = 0; i < 32; i++)
for (uint16_t i = 0; i < 32; i++)
{
regs[i].offset = Pm4::SPI_PS_INPUT_CNTL_0 + i;
regs[i].value = 0;
if (i < static_cast<int>(gs->num_output_semantics))
{
EXIT_NOT_IMPLEMENTED(out32[i] != 0x0000000f);
EXIT_NOT_IMPLEMENTED(in32 != nullptr && out32[i] != in32[i]);
regs[i].value = i;
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].semantic != 15 + i);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].hardware_mapping != i);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].size_in_elements != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].is_f16 != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].is_flat_shaded != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].is_linear != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].is_custom != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].static_vb_index != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].static_attribute != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].default_value != 0);
EXIT_NOT_IMPLEMENTED(gs->output_semantics[i].default_value_hi != 0);
bool flat = false;
if (ps != nullptr)
{
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].semantic != gs->output_semantics[i].semantic);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].hardware_mapping != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].size_in_elements != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].is_f16 != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].is_linear != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].is_custom != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].static_vb_index != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].static_attribute != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].default_value != 0);
EXIT_NOT_IMPLEMENTED(ps->input_semantics[i].default_value_hi != 0);
flat = ps->input_semantics[i].is_flat_shaded != 0;
}
regs[i].value = i | (flat ? 0x400u : 0u);
}
}
@@ -1837,7 +1780,7 @@ uint32_t* KYTY_SYSV_ABI GraphicsCbReleaseMem(CommandBuffer* buf, uint8_t action,
EXIT_NOT_IMPLEMENTED(buf == nullptr);
EXIT_NOT_IMPLEMENTED(dst != 1);
EXIT_NOT_IMPLEMENTED(data_sel != 2);
EXIT_NOT_IMPLEMENTED(data_sel != 2 && data_sel != 3);
EXIT_NOT_IMPLEMENTED(gds_offset != 0);
EXIT_NOT_IMPLEMENTED(gds_size != 1);
EXIT_NOT_IMPLEMENTED(interrupt != 0);
@@ -1851,7 +1794,7 @@ uint32_t* KYTY_SYSV_ABI GraphicsCbReleaseMem(CommandBuffer* buf, uint8_t action,
cmd[0] = KYTY_PM4(7, Pm4::IT_NOP, Pm4::R_RELEASE_MEM);
cmd[1] = action | (static_cast<uint32_t>(cache_policy) << 8u);
cmd[2] = gcr_cntl;
cmd[2] = gcr_cntl | (static_cast<uint32_t>(data_sel) << 16u);
cmd[3] = static_cast<uint32_t>(reinterpret_cast<uint64_t>(address) & 0xffffffffu);
cmd[4] = static_cast<uint32_t>((reinterpret_cast<uint64_t>(address) >> 32u) & 0xffffffffu);
cmd[5] = static_cast<uint32_t>(data & 0xffffffffu);
@@ -1909,6 +1852,25 @@ uint32_t* KYTY_SYSV_ABI GraphicsDcbWaitUntilSafeForRendering(CommandBuffer* buf,
return cmd;
}
uint32_t* KYTY_SYSV_ABI GraphicsDcbSetShRegisterDirect(CommandBuffer* buf, ShaderRegister reg)
{
PRINT_NAME();
EXIT_NOT_IMPLEMENTED(buf == nullptr);
buf->DbgDump();
auto* cmd = buf->AllocateDW(3);
EXIT_NOT_IMPLEMENTED(cmd == nullptr);
cmd[0] = KYTY_PM4(3, Pm4::IT_SET_SH_REG, 0u);
cmd[1] = reg.offset;
cmd[2] = reg.value;
return cmd;
}
uint32_t* KYTY_SYSV_ABI GraphicsDcbSetCxRegistersIndirect(CommandBuffer* buf, const volatile ShaderRegister* regs, uint32_t num_regs)
{
PRINT_NAME();
source/emulator/src/Graphics/GraphicsRender.cpp
+385 -177
View File
@@ -471,6 +471,7 @@ static void uc_print(const char* func, const HW::UserConfig& uc)
printf("\t GetPrimType() = 0x%08" PRIx32 "\n", uc.GetPrimType());
printf("\t primitive_group_size = 0x%04" PRIx16 "\n", ge_cntl.primitive_group_size);
printf("\t vertex_group_size = 0x%04" PRIx16 "\n", ge_cntl.vertex_group_size);
printf("\t en_user_vgpr1 = %s\n", user_en.vgpr1 ? "true" : "false");
printf("\t en_user_vgpr2 = %s\n", user_en.vgpr2 ? "true" : "false");
printf("\t en_user_vgpr3 = %s\n", user_en.vgpr3 ? "true" : "false");
@@ -481,8 +482,8 @@ static void uc_check(const HW::UserConfig& uc)
const auto& ge_cntl = uc.GetGeControl();
const auto& user_en = uc.GetGeUserVgprEn();
EXIT_NOT_IMPLEMENTED(ge_cntl.primitive_group_size != 0x0000);
EXIT_NOT_IMPLEMENTED(ge_cntl.vertex_group_size != 0x0000);
EXIT_NOT_IMPLEMENTED(ge_cntl.primitive_group_size != 0x0000 && ge_cntl.primitive_group_size != 0x0040);
EXIT_NOT_IMPLEMENTED(ge_cntl.vertex_group_size != 0x0000 && ge_cntl.vertex_group_size != 0x0040);
EXIT_NOT_IMPLEMENTED(user_en.vgpr1 != false);
EXIT_NOT_IMPLEMENTED(user_en.vgpr2 != false);
EXIT_NOT_IMPLEMENTED(user_en.vgpr3 != false);
@@ -565,11 +566,15 @@ static void rt_check(const HW::RenderTarget& rt)
{
if (rt.base.addr != 0)
{
bool render_to_texture = (rt.attrib.tile_mode == 0x0d);
// EXIT_NOT_IMPLEMENTED(rt.base_addr == 0);
// EXIT_NOT_IMPLEMENTED(rt.pitch_div8_minus1 != 0x000000ef);
// EXIT_NOT_IMPLEMENTED(rt.fmask_pitch_div8_minus1 != 0x000000ef);
// EXIT_NOT_IMPLEMENTED(rt.slice_div64_minus1 != 0x000086ff);
bool ps5 = Config::IsNextGen();
// bool render_to_texture = (rt.attrib.tile_mode == 0x0d);
// EXIT_NOT_IMPLEMENTED(rt.base_addr == 0);
if (ps5)
{
EXIT_NOT_IMPLEMENTED(rt.pitch.pitch_div8_minus1 != 0);
EXIT_NOT_IMPLEMENTED(rt.pitch.fmask_pitch_div8_minus1 != 0);
EXIT_NOT_IMPLEMENTED(rt.slice.slice_div64_minus1 != 0);
}
EXIT_NOT_IMPLEMENTED(rt.view.base_array_slice_index != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.view.last_array_slice_index != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.view.current_mip_level != 0x00000000);
@@ -581,7 +586,7 @@ static void rt_check(const HW::RenderTarget& rt)
EXIT_NOT_IMPLEMENTED(rt.info.cmask_fast_clear_enable != false);
EXIT_NOT_IMPLEMENTED(rt.info.dcc_compression_enable != false);
EXIT_NOT_IMPLEMENTED(!render_to_texture && rt.info.neo_mode != Config::IsNeo());
EXIT_NOT_IMPLEMENTED(!(rt.attrib.tile_mode == 0x0d) && rt.info.neo_mode != Config::IsNeo());
EXIT_NOT_IMPLEMENTED(rt.info.cmask_tile_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.info.cmask_tile_mode_neo != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.info.blend_bypass != false);
@@ -595,14 +600,27 @@ static void rt_check(const HW::RenderTarget& rt)
// EXIT_NOT_IMPLEMENTED(rt.fmask_tile_mode != 0x0000000a);
EXIT_NOT_IMPLEMENTED(rt.attrib.num_samples != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib.num_fragments != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.width != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.height != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.num_mip_levels != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.depth != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.tile_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dimension != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.cmask_pipe_aligned != false);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dcc_pipe_aligned != false);
if (ps5)
{
EXIT_NOT_IMPLEMENTED(rt.attrib2.width == 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.height == 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.num_mip_levels != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.depth != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.tile_mode != 0x0000001b);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dimension != 0x00000001);
EXIT_NOT_IMPLEMENTED(rt.attrib3.cmask_pipe_aligned != true);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dcc_pipe_aligned != true);
} else
{
EXIT_NOT_IMPLEMENTED(rt.attrib2.width != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.height != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib2.num_mip_levels != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.depth != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.tile_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dimension != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.attrib3.cmask_pipe_aligned != false);
EXIT_NOT_IMPLEMENTED(rt.attrib3.dcc_pipe_aligned != false);
}
// EXIT_NOT_IMPLEMENTED(rt.dcc_max_uncompressed_block_size != 0x00000002);
// EXIT_NOT_IMPLEMENTED(rt.dcc.max_compressed_block_size != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.dcc.min_compressed_block_size != 0x00000000);
@@ -619,8 +637,11 @@ static void rt_check(const HW::RenderTarget& rt)
EXIT_NOT_IMPLEMENTED(rt.clear_word0.word0 != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.clear_word1.word1 != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.dcc_addr.addr != 0x0000000000000000);
// EXIT_NOT_IMPLEMENTED(rt.width != 0x00000780);
// EXIT_NOT_IMPLEMENTED(rt.height != 0x00000438);
if (ps5)
{
EXIT_NOT_IMPLEMENTED(rt.size.width != 0);
EXIT_NOT_IMPLEMENTED(rt.size.height != 0);
}
}
}
@@ -729,25 +750,46 @@ static void z_check(const HW::DepthRenderTarget& z)
if (z.z_info.format != 0 || z.stencil_info.format != 0)
{
EXIT_NOT_IMPLEMENTED(z.depth_info.addr5_swizzle_mask != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.depth_info.array_mode != 0x00000004);
EXIT_NOT_IMPLEMENTED(z.depth_info.pipe_config != (Config::IsNeo() ? 0x00000012 : 0x0c));
EXIT_NOT_IMPLEMENTED(z.depth_info.bank_width != 0x00000000);
// EXIT_NOT_IMPLEMENTED(z.depth_info.bank_height != (Config::IsNeo() ? 0x00000001 : 2));
// EXIT_NOT_IMPLEMENTED(z.depth_info.macro_tile_aspect != (Config::IsNeo() ? 0x00000000 : 2));
// EXIT_NOT_IMPLEMENTED(z.depth_info.num_banks != (Config::IsNeo() ? 0x00000002 : 3));
bool ps5 = Config::IsNextGen();
if (ps5)
{
EXIT_NOT_IMPLEMENTED(z.depth_info.addr5_swizzle_mask != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.array_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.pipe_config != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.bank_width != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.bank_height != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.macro_tile_aspect != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_info.num_banks != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.linear != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.full_cache != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.htile_uses_preload_win != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.preload != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_width != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_height != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.dst_outside_zero_to_one != 0x00000000);
} else
{
EXIT_NOT_IMPLEMENTED(z.depth_info.addr5_swizzle_mask != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.depth_info.array_mode != 0x00000004);
EXIT_NOT_IMPLEMENTED(z.depth_info.pipe_config != (Config::IsNeo() ? 0x00000012 : 0x0c));
EXIT_NOT_IMPLEMENTED(z.depth_info.bank_width != 0x00000000);
// EXIT_NOT_IMPLEMENTED(z.depth_info.bank_height != (Config::IsNeo() ? 0x00000001 : 2));
// EXIT_NOT_IMPLEMENTED(z.depth_info.macro_tile_aspect != (Config::IsNeo() ? 0x00000000 : 2));
// EXIT_NOT_IMPLEMENTED(z.depth_info.num_banks != (Config::IsNeo() ? 0x00000002 : 3));
EXIT_NOT_IMPLEMENTED(z.htile_surface.linear != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.full_cache != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.htile_uses_preload_win != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.preload != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_width != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_height != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.dst_outside_zero_to_one != 0x00000000);
}
EXIT_NOT_IMPLEMENTED(z.depth_view.slice_start != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_view.slice_max != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_view.current_mip_level != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.depth_view.depth_write_disable != false);
EXIT_NOT_IMPLEMENTED(z.depth_view.stencil_write_disable != false);
EXIT_NOT_IMPLEMENTED(z.htile_surface.linear != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.full_cache != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.htile_uses_preload_win != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.preload != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_width != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_height != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.htile_surface.dst_outside_zero_to_one != 0x00000000);
EXIT_NOT_IMPLEMENTED(z.z_read_base_addr != z.z_write_base_addr);
EXIT_NOT_IMPLEMENTED(z.stencil_read_base_addr != z.stencil_write_base_addr);
EXIT_NOT_IMPLEMENTED(z.z_write_base_addr == 0);
@@ -758,8 +800,13 @@ static void z_check(const HW::DepthRenderTarget& z)
// EXIT_NOT_IMPLEMENTED(z.htile_data_base_addr == 0);
// EXIT_NOT_IMPLEMENTED(z.width != 0x00000780);
// EXIT_NOT_IMPLEMENTED(z.height != 0x00000438);
EXIT_NOT_IMPLEMENTED(z.size.x_max != 0);
EXIT_NOT_IMPLEMENTED(z.size.y_max != 0);
if (ps5)
{
EXIT_NOT_IMPLEMENTED(z.width != 0);
EXIT_NOT_IMPLEMENTED(z.height != 0);
EXIT_NOT_IMPLEMENTED(z.size.x_max == 0);
EXIT_NOT_IMPLEMENTED(z.size.y_max == 0);
}
}
}
@@ -1044,14 +1091,23 @@ static void vp_print(const char* func, const HW::ScreenViewport& vp, const HW::S
static void vp_check(const HW::ScreenViewport& vp, const HW::ScanModeControl& smc)
{
bool ps5 = Config::IsNextGen();
EXIT_NOT_IMPLEMENTED(smc.msaa_enable);
// EXIT_NOT_IMPLEMENTED(smc.vport_scissor_enable);
EXIT_NOT_IMPLEMENTED(smc.line_stipple_enable);
bool generic_scissor =
(vp.generic_scissor_left != 0 || vp.generic_scissor_top != 0 || vp.generic_scissor_right != 0 || vp.generic_scissor_bottom != 0);
bool screen_scissor =
(vp.screen_scissor_left != 0 || vp.screen_scissor_top != 0 || vp.screen_scissor_right != 0 || vp.screen_scissor_bottom != 0);
bool viewport_scissor = (vp.viewports[0].viewport_scissor_left != 0 || vp.viewports[0].viewport_scissor_top != 0 ||
vp.viewports[0].viewport_scissor_right != 0 || vp.viewports[0].viewport_scissor_bottom != 0);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].zmin != 0.000000);
EXIT_NOT_IMPLEMENTED(viewport_scissor && (generic_scissor || screen_scissor));
EXIT_NOT_IMPLEMENTED(viewport_scissor && (!smc.vport_scissor_enable));
EXIT_NOT_IMPLEMENTED(vp.viewports[0].zmin > 0.000000);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].zmax != 1.000000);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].xscale != 960.000000);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].xoffset != 960.000000);
@@ -1068,14 +1124,21 @@ static void vp_check(const HW::ScreenViewport& vp, const HW::ScanModeControl& sm
// EXIT_NOT_IMPLEMENTED(vp.hw_offset_y != 32);
// EXIT_NOT_IMPLEMENTED(fabsf(vp.guard_band_horz_clip - 33.133327f) > 0.001f);
// EXIT_NOT_IMPLEMENTED(fabsf(vp.guard_band_vert_clip - 59.629623f) > 0.001f);
EXIT_NOT_IMPLEMENTED(vp.guard_band_horz_discard != 1.000000);
EXIT_NOT_IMPLEMENTED(vp.guard_band_vert_discard != 1.000000);
if (ps5)
{
EXIT_NOT_IMPLEMENTED(vp.guard_band_horz_discard != 0.0f);
EXIT_NOT_IMPLEMENTED(vp.guard_band_vert_discard != 0.0f);
} else
{
EXIT_NOT_IMPLEMENTED(vp.guard_band_horz_discard != 1.000000);
EXIT_NOT_IMPLEMENTED(vp.guard_band_vert_discard != 1.000000);
}
EXIT_NOT_IMPLEMENTED(vp.generic_scissor_window_offset_enable != false);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_left != 0);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_top != 0);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_right != 0);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_bottom != 0);
EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_window_offset_enable != false);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_left != 0);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_top != 0);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_right != 0);
// EXIT_NOT_IMPLEMENTED(vp.viewports[0].viewport_scissor_bottom != 0);
EXIT_NOT_IMPLEMENTED(viewport_scissor && vp.viewports[0].viewport_scissor_window_offset_enable != true);
}
static void hw_check(const HW::Context& hw)
@@ -1785,33 +1848,53 @@ uint64_t SamplerCache::GetSamplerId(const ShaderSamplerResource& r)
return m_samplers_size;
}
static void get_input_format(const ShaderBufferResource& res, VkFormat* format, uint32_t* size)
static void get_input_format(const ShaderBufferResource& res, VkFormat* format, uint32_t* size, bool ps5)
{
EXIT_IF(format == nullptr);
EXIT_IF(size == nullptr);
auto nfmt = res.Nfmt();
auto dfmt = res.Dfmt();
if (nfmt == 7 && dfmt == 14)
if (ps5)
{
*format = VK_FORMAT_R32G32B32A32_SFLOAT;
*size = 4;
} else if (nfmt == 7 && dfmt == 13)
{
*format = VK_FORMAT_R32G32B32_SFLOAT;
*size = 3;
} else if (nfmt == 7 && dfmt == 11)
{
*format = VK_FORMAT_R32G32_SFLOAT;
*size = 2;
} else if (nfmt == 0 && dfmt == 10)
{
*format = VK_FORMAT_R8G8B8A8_UNORM;
*size = 4;
auto fmt = res.Format();
if (fmt == 74)
{
*format = VK_FORMAT_R32G32B32_SFLOAT;
*size = 3;
} else if (fmt == 64)
{
*format = VK_FORMAT_R32G32_SFLOAT;
*size = 2;
} else
{
EXIT("unknown format: fmt = %u\n", fmt);
*format = VK_FORMAT_UNDEFINED;
*size = 4;
}
} else
{
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
*format = VK_FORMAT_UNDEFINED;
*size = 4;
auto nfmt = res.Nfmt();
auto dfmt = res.Dfmt();
if (nfmt == 7 && dfmt == 14)
{
*format = VK_FORMAT_R32G32B32A32_SFLOAT;
*size = 4;
} else if (nfmt == 7 && dfmt == 13)
{
*format = VK_FORMAT_R32G32B32_SFLOAT;
*size = 3;
} else if (nfmt == 7 && dfmt == 11)
{
*format = VK_FORMAT_R32G32_SFLOAT;
*size = 2;
} else if (nfmt == 0 && dfmt == 10)
{
*format = VK_FORMAT_R8G8B8A8_UNORM;
*size = 4;
} else
{
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
*format = VK_FORMAT_UNDEFINED;
*size = 4;
}
}
}
@@ -1952,6 +2035,8 @@ static VulkanPipeline* CreatePipelineInternal(VkRenderPass render_pass, const Sh
VkVertexInputAttributeDescription input_attr[ShaderVertexInputInfo::RES_MAX];
VkVertexInputBindingDescription input_desc[ShaderVertexInputInfo::RES_MAX];
bool gen5 = Config::IsNextGen();
for (int bi = 0; bi < vs_input_info->buffers_num; bi++)
{
const auto& b = vs_input_info->buffers[bi];
@@ -1967,10 +2052,14 @@ static VulkanPipeline* CreatePipelineInternal(VkRenderPass render_pass, const Sh
input_attr[index].offset = b.attr_offsets[ai];
uint32_t attr_size = 4;
get_input_format(vs_input_info->resources[index], &input_attr[index].format, &attr_size);
get_input_format(vs_input_info->resources[index], &input_attr[index].format, &attr_size, gen5);
auto registers_num = vs_input_info->resources_dst[index].registers_num;
if (gen5)
{
EXIT_NOT_IMPLEMENTED(vs_input_info->resources[index].OutOfBounds() != 0);
}
EXIT_NOT_IMPLEMENTED(vs_input_info->resources[index].AddTid());
EXIT_NOT_IMPLEMENTED(vs_input_info->resources[index].SwizzleEnabled());
@@ -2484,17 +2573,23 @@ VulkanPipeline* PipelineCache::CreatePipeline(VulkanFramebuffer* framebuffer, Re
bool generic_scissor =
(vp.generic_scissor_left != 0 || vp.generic_scissor_top != 0 || vp.generic_scissor_right != 0 || vp.generic_scissor_bottom != 0);
bool viewport_scissor = (vp.viewports[0].viewport_scissor_left != 0 || vp.viewports[0].viewport_scissor_top != 0 ||
vp.viewports[0].viewport_scissor_right != 0 || vp.viewports[0].viewport_scissor_bottom != 0);
p.static_params->viewport_scale[0] = vp.viewports[0].xscale;
p.static_params->viewport_scale[1] = vp.viewports[0].yscale;
p.static_params->viewport_scale[2] = vp.viewports[0].zscale;
p.static_params->viewport_offset[0] = vp.viewports[0].xoffset;
p.static_params->viewport_offset[1] = vp.viewports[0].yoffset;
p.static_params->viewport_offset[2] = vp.viewports[0].zoffset;
p.static_params->scissor_ltrb[0] = (generic_scissor ? vp.generic_scissor_left : vp.screen_scissor_left);
p.static_params->scissor_ltrb[1] = (generic_scissor ? vp.generic_scissor_top : vp.screen_scissor_top);
p.static_params->scissor_ltrb[2] = (generic_scissor ? vp.generic_scissor_right : vp.screen_scissor_right);
p.static_params->scissor_ltrb[3] = (generic_scissor ? vp.generic_scissor_bottom : vp.screen_scissor_bottom);
p.static_params->viewport_scale[0] = vp.viewports[0].xscale;
p.static_params->viewport_scale[1] = vp.viewports[0].yscale;
p.static_params->viewport_scale[2] = vp.viewports[0].zscale;
p.static_params->viewport_offset[0] = vp.viewports[0].xoffset;
p.static_params->viewport_offset[1] = vp.viewports[0].yoffset;
p.static_params->viewport_offset[2] = vp.viewports[0].zoffset;
p.static_params->scissor_ltrb[0] =
(viewport_scissor ? vp.viewports[0].viewport_scissor_left : (generic_scissor ? vp.generic_scissor_left : vp.screen_scissor_left));
p.static_params->scissor_ltrb[1] =
(viewport_scissor ? vp.viewports[0].viewport_scissor_top : (generic_scissor ? vp.generic_scissor_top : vp.screen_scissor_top));
p.static_params->scissor_ltrb[2] = (viewport_scissor ? vp.viewports[0].viewport_scissor_right
: (generic_scissor ? vp.generic_scissor_right : vp.screen_scissor_right));
p.static_params->scissor_ltrb[3] = (viewport_scissor ? vp.viewports[0].viewport_scissor_bottom
: (generic_scissor ? vp.generic_scissor_bottom : vp.screen_scissor_bottom));
p.static_params->topology = topology;
p.static_params->with_depth = (depth->format != VK_FORMAT_UNDEFINED && depth->vulkan_buffer != nullptr);
p.static_params->depth_test_enable = depth->depth_test_enable;
@@ -3713,42 +3808,22 @@ static void FindRenderDepthInfo(uint64_t submit_id, CommandBuffer* /*buffer*/, c
EXIT_IF(r == nullptr);
bool neo = Config::IsNeo();
const auto& z = hw.GetDepthRenderTarget();
const auto& rc = hw.GetRenderControl();
const auto& dc = hw.GetDepthControl();
const auto& sc = hw.GetStencilControl();
const auto& sm = hw.GetStencilMask();
const auto& cc = hw.GetColorControl();
const auto& z = hw.GetDepthRenderTarget();
const auto& rc = hw.GetRenderControl();
const auto& dc = hw.GetDepthControl();
const auto& sc = hw.GetStencilControl();
const auto& sm = hw.GetStencilMask();
const auto& cc = hw.GetColorControl();
uint32_t size = 0;
uint32_t pitch = (z.pitch_div8_minus1 + 1) * 8;
TileSizeAlign stencil_size {};
TileSizeAlign htile_size {};
TileSizeAlign depth_size {};
if (z.z_info.format == 3)
{
size = (z.slice_div64_minus1 + 1) * 64 * 4;
} else if (z.z_info.format == 1)
{
size = (z.slice_div64_minus1 + 1) * 64 * 2;
}
bool htile = z.z_info.tile_surface_enable;
bool neo = Config::IsNeo();
bool ps5 = Config::IsNextGen();
bool htile = z.z_info.tile_surface_enable;
bool decompress = htile && (rc.depth_compress_disable || rc.stencil_compress_disable);
if (!TileGetDepthSize(z.width, z.height, pitch, z.z_info.format, z.stencil_info.format, htile, neo, &stencil_size, &htile_size,
&depth_size))
{
depth_size.size = size;
depth_size.align = neo ? 65536 : 32768;
} else
{
EXIT_NOT_IMPLEMENTED(depth_size.size != size);
}
if (dc.z_enable || dc.z_write_enable || dc.stencil_enable || decompress)
{
switch (z.z_info.format * 2 + z.stencil_info.format)
@@ -3763,6 +3838,38 @@ static void FindRenderDepthInfo(uint64_t submit_id, CommandBuffer* /*buffer*/, c
}
}
if (r->format != VK_FORMAT_UNDEFINED)
{
if (ps5)
{
bool size_found = TileGetDepthSize(z.size.x_max + 1, z.size.y_max + 1, 0, z.z_info.format, z.stencil_info.format, htile, true,
true, &stencil_size, &htile_size, &depth_size);
EXIT_NOT_IMPLEMENTED(!size_found);
} else
{
uint32_t size = 0;
uint32_t pitch = (z.pitch_div8_minus1 + 1) * 8;
if (z.z_info.format == 3)
{
size = (z.slice_div64_minus1 + 1) * 64 * 4;
} else if (z.z_info.format == 1)
{
size = (z.slice_div64_minus1 + 1) * 64 * 2;
}
if (!TileGetDepthSize(z.width, z.height, pitch, z.z_info.format, z.stencil_info.format, htile, neo, false, &stencil_size,
&htile_size, &depth_size))
{
depth_size.size = size;
depth_size.align = neo ? 65536 : 32768;
} else
{
EXIT_NOT_IMPLEMENTED(depth_size.size != size);
}
}
}
auto stencil_addr_mask = static_cast<uint64_t>(stencil_size.align) - 1;
auto depth_addr_mask = static_cast<uint64_t>(depth_size.align) - 1;
auto htile_addr_mask = static_cast<uint64_t>(htile_size.align) - 1;
@@ -3778,8 +3885,8 @@ static void FindRenderDepthInfo(uint64_t submit_id, CommandBuffer* /*buffer*/, c
r->htile_buffer_size = htile_size.size;
r->htile_buffer_vaddr = z.htile_data_base_addr & ~htile_addr_mask;
r->htile_tile_swizzle = z.htile_data_base_addr & htile_addr_mask;
r->width = z.width;
r->height = z.height;
r->width = (ps5 ? z.size.x_max + 1 : z.width);
r->height = (ps5 ? z.size.y_max + 1 : z.height);
r->depth_clear_enable = rc.depth_clear_enable;
r->depth_clear_value = hw.GetDepthClearValue();
r->depth_test_enable = dc.z_enable;
@@ -3888,38 +3995,69 @@ static void FindRenderColorInfo(uint64_t submit_id, CommandBuffer* buffer, const
return;
}
auto width = rt.size.width;
auto height = rt.size.height;
auto pitch = (rt.pitch.pitch_div8_minus1 + 1) * 8;
auto size = (rt.slice.slice_div64_minus1 + 1) * 64 * 4;
bool tile = false;
bool write_back = false;
bool ps5 = Config::IsNextGen();
uint32_t width = 0;
uint32_t height = 0;
uint32_t pitch = 0;
uint32_t size = 0;
bool tile = false;
bool write_back = false;
if (ps5)
{
switch (rt.attrib3.tile_mode)
{
case 0x1b:
tile = true;
write_back = false;
break;
default: EXIT("unknown tile mode: %u\n", rt.attrib3.tile_mode);
}
width = rt.attrib2.width + 1;
height = rt.attrib2.height + 1;
pitch = width;
Graphics::TileSizeAlign size32 {};
Graphics::TileGetVideoOutSize(width, height, pitch, tile, true, &size32);
size = size32.size;
EXIT_NOT_IMPLEMENTED(size == 0);
} else
{
switch (rt.attrib.tile_mode)
{
case 0x8:
tile = false;
write_back = false;
break;
case 0x1f:
tile = false;
write_back = true;
break;
case 0xa:
case 0xd:
case 0xe:
tile = true;
write_back = false;
break;
default: EXIT("unknown tile mode: %u\n", rt.attrib.tile_mode);
}
width = rt.size.width;
height = rt.size.height;
pitch = (rt.pitch.pitch_div8_minus1 + 1) * 8;
size = (rt.slice.slice_div64_minus1 + 1) * 64 * 4;
}
auto video_image = VideoOut::VideoOutGetImage(rt.base.addr);
bool render_to_texture = (video_image.image == nullptr);
switch (rt.attrib.tile_mode)
{
case 0x8:
tile = false;
write_back = false;
break;
case 0x1f:
tile = false;
write_back = true;
break;
case 0xa:
case 0xd:
case 0xe:
tile = true;
write_back = false;
break;
default: EXIT("unknown tile mode: %u\n", rt.attrib.tile_mode);
}
if (render_to_texture)
{
RenderTextureFormat rt_format = RenderTextureFormat::Unknown;
@@ -3955,8 +4093,8 @@ static void FindRenderColorInfo(uint64_t submit_id, CommandBuffer* buffer, const
r->buffer_size = size;
} else
{
EXIT_NOT_IMPLEMENTED(
!(rt.info.format == 0xa && (rt.info.channel_type == 0x6 || rt.info.channel_type == 0x0) && rt.info.channel_order == 0x1));
EXIT_NOT_IMPLEMENTED(!(rt.info.format == 0xa && (rt.info.channel_type == 0x6 || rt.info.channel_type == 0x0) &&
(rt.info.channel_order == 0x0 || rt.info.channel_order == 0x1)));
// Display buffer
EXIT_NOT_IMPLEMENTED(video_image.buffer_size != size);
@@ -3994,6 +4132,7 @@ static void InvalidateMemoryObject(const RenderDepthInfo& r)
}
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
static void PrepareStorageBuffers(uint64_t submit_id, CommandBuffer* buffer, const ShaderStorageResources& storage_buffers,
VulkanBuffer** buffers, uint32_t** sgprs)
{
@@ -4003,26 +4142,36 @@ static void PrepareStorageBuffers(uint64_t submit_id, CommandBuffer* buffer, con
EXIT_IF(sgprs == nullptr);
EXIT_IF(*sgprs == nullptr);
bool gen5 = Config::IsNextGen();
for (int i = 0; i < storage_buffers.buffers_num; i++)
{
auto r = storage_buffers.buffers[i];
EXIT_NOT_IMPLEMENTED(r.AddTid());
EXIT_NOT_IMPLEMENTED(r.SwizzleEnabled());
EXIT_NOT_IMPLEMENTED(!((r.Stride() == 4 && r.DstSelXYZW() == DstSel(4, 0, 0, 0) && r.Dfmt() == 4 && r.Nfmt() == 4) ||
(r.Stride() == 4 && r.DstSelXYZW() == DstSel(4, 0, 0, 1) && r.Dfmt() == 4 && r.Nfmt() == 7) ||
(r.Stride() == 8 && r.DstSelXYZW() == DstSel(4, 5, 0, 0) && r.Dfmt() == 11 && r.Nfmt() == 4) ||
(r.Stride() == 16 && r.DstSelXYZW() == DstSel(4, 5, 6, 7) && r.Dfmt() == 14 && r.Nfmt() == 7)));
EXIT_NOT_IMPLEMENTED(!(r.MemoryType() == 0x00 || r.MemoryType() == 0x10 || r.MemoryType() == 0x6d));
auto addr = r.Base();
if (gen5)
{
EXIT_NOT_IMPLEMENTED(r.OutOfBounds() != 0);
EXIT_NOT_IMPLEMENTED(!((r.Stride() == 16 && r.DstSelXYZW() == DstSel(4, 5, 6, 7) && r.Format() == 77)));
} else
{
EXIT_NOT_IMPLEMENTED(!((r.Stride() == 4 && r.DstSelXYZW() == DstSel(4, 0, 0, 0) && r.Dfmt() == 4 && r.Nfmt() == 4) ||
(r.Stride() == 4 && r.DstSelXYZW() == DstSel(4, 0, 0, 1) && r.Dfmt() == 4 && r.Nfmt() == 7) ||
(r.Stride() == 8 && r.DstSelXYZW() == DstSel(4, 5, 0, 0) && r.Dfmt() == 11 && r.Nfmt() == 4) ||
(r.Stride() == 16 && r.DstSelXYZW() == DstSel(4, 5, 6, 7) && r.Dfmt() == 14 && r.Nfmt() == 7)));
EXIT_NOT_IMPLEMENTED(!(r.MemoryType() == 0x00 || r.MemoryType() == 0x10 || r.MemoryType() == 0x6d));
}
auto addr = (gen5 ? r.Base48() : r.Base44());
auto stride = r.Stride();
auto num_records = r.NumRecords();
auto size = stride * num_records;
EXIT_NOT_IMPLEMENTED(size == 0);
EXIT_NOT_IMPLEMENTED((size & 0x3u) != 0);
bool read_only = (r.MemoryType() == 0x10);
bool read_only = (gen5 ? false : (r.MemoryType() == 0x10));
EXIT_NOT_IMPLEMENTED(read_only && !(storage_buffers.usages[i] == ShaderStorageUsage::ReadOnly ||
storage_buffers.usages[i] == ShaderStorageUsage::Constant));
@@ -4038,9 +4187,15 @@ static void PrepareStorageBuffers(uint64_t submit_id, CommandBuffer* buffer, con
buffers[i] = buf;
r.UpdateAddress(i);
if (gen5)
{
r.UpdateAddress48(i);
} else
{
r.UpdateAddress44(i);
}
EXIT_NOT_IMPLEMENTED((r.Base() >> 32u) != 0);
EXIT_NOT_IMPLEMENTED(((gen5 ? r.Base48() : r.Base44()) >> 32u) != 0);
(*sgprs)[0] = r.fields[0];
(*sgprs)[1] = r.fields[1];
@@ -4065,48 +4220,83 @@ static void PrepareTextures(uint64_t submit_id, CommandBuffer* buffer, const Sha
int index_sampled = 0;
int index_storage = 0;
bool gen5 = Config::IsNextGen();
for (int i = 0; i < textures.textures_num; i++)
{
auto r = textures.desc[i].texture;
EXIT_NOT_IMPLEMENTED(r.Base() == 0);
if (gen5)
{
EXIT_NOT_IMPLEMENTED(!(r.TileMode() == 0));
EXIT_NOT_IMPLEMENTED(r.Format() != 56);
EXIT_NOT_IMPLEMENTED(r.PerfMod5() != 7 && r.PerfMod5() != 0);
EXIT_NOT_IMPLEMENTED(r.BCSwizzle() != 0);
EXIT_NOT_IMPLEMENTED(r.BaseArray5() != 0);
EXIT_NOT_IMPLEMENTED(r.ArrayPitch() != 0);
EXIT_NOT_IMPLEMENTED(r.MaxMip() != 0);
EXIT_NOT_IMPLEMENTED(r.MinLodWarn5() != 0);
EXIT_NOT_IMPLEMENTED(r.MipStatsCntId() != 0);
EXIT_NOT_IMPLEMENTED(r.MipStatsCntEn() != false);
EXIT_NOT_IMPLEMENTED(r.CornerSample() != false);
EXIT_NOT_IMPLEMENTED(r.PrtDefColor() != false);
EXIT_NOT_IMPLEMENTED(r.MsaaDepth() != false);
EXIT_NOT_IMPLEMENTED(r.MaxUncompBlkSize() != 0);
EXIT_NOT_IMPLEMENTED(r.MaxCompBlkSize() != 0);
EXIT_NOT_IMPLEMENTED(r.MetaPipeAligned() != false);
EXIT_NOT_IMPLEMENTED(r.WriteCompress() != false);
EXIT_NOT_IMPLEMENTED(r.MetaCompress() != false);
EXIT_NOT_IMPLEMENTED(r.DccAlphaPos() != false);
EXIT_NOT_IMPLEMENTED(r.DccColorTransf() != false);
EXIT_NOT_IMPLEMENTED(r.MetaAddr() != 0);
} else
{
EXIT_NOT_IMPLEMENTED(r.Dfmt() != 1 && r.Dfmt() != 10 && r.Dfmt() != 37 && r.Dfmt() != 4 && r.Dfmt() != 35 && r.Dfmt() != 3 &&
r.Dfmt() != 36);
EXIT_NOT_IMPLEMENTED(r.Nfmt() != 9 && r.Nfmt() != 0 && r.Nfmt() != 7);
EXIT_NOT_IMPLEMENTED(r.PerfMod() != 7 && r.PerfMod() != 0);
EXIT_NOT_IMPLEMENTED(r.Interlaced() != false);
EXIT_NOT_IMPLEMENTED(!(r.TileMode() == 8 || r.TileMode() == 13 || r.TileMode() == 14 || r.TileMode() == 2 ||
r.TileMode() == 10 || r.TileMode() == 31));
EXIT_NOT_IMPLEMENTED(r.BaseArray() != 0);
EXIT_NOT_IMPLEMENTED(r.LastArray() != 0);
EXIT_NOT_IMPLEMENTED(r.MinLodWarn() != 0);
EXIT_NOT_IMPLEMENTED(r.CounterBankId() != 0);
EXIT_NOT_IMPLEMENTED(r.LodHdwCntEn() != false);
EXIT_NOT_IMPLEMENTED(r.MemoryType() != 0x10 && r.MemoryType() != 0x6d);
}
EXIT_NOT_IMPLEMENTED((gen5 ? r.Base40() : r.Base38()) == 0);
EXIT_NOT_IMPLEMENTED(r.MinLod() != 0);
EXIT_NOT_IMPLEMENTED(r.Dfmt() != 1 && r.Dfmt() != 10 && r.Dfmt() != 37 && r.Dfmt() != 4 && r.Dfmt() != 35 && r.Dfmt() != 3 &&
r.Dfmt() != 36);
EXIT_NOT_IMPLEMENTED(r.Nfmt() != 9 && r.Nfmt() != 0 && r.Nfmt() != 7);
EXIT_NOT_IMPLEMENTED(r.PerfMod() != 7 && r.PerfMod() != 0);
EXIT_NOT_IMPLEMENTED(r.Interlaced() != false);
EXIT_NOT_IMPLEMENTED(!(r.TilingIdx() == 8 || r.TilingIdx() == 13 || r.TilingIdx() == 14 || r.TilingIdx() == 2 ||
r.TilingIdx() == 10 || r.TilingIdx() == 31));
EXIT_NOT_IMPLEMENTED(r.Type() != 9);
EXIT_NOT_IMPLEMENTED(r.Depth() != 0);
EXIT_NOT_IMPLEMENTED(r.BaseArray() != 0);
EXIT_NOT_IMPLEMENTED(r.LastArray() != 0);
EXIT_NOT_IMPLEMENTED(r.MinLodWarn() != 0);
EXIT_NOT_IMPLEMENTED(r.CounterBankId() != 0);
EXIT_NOT_IMPLEMENTED(r.LodHdwCntEn() != false);
EXIT_NOT_IMPLEMENTED(r.MemoryType() != 0x10 && r.MemoryType() != 0x6d);
bool read_only = (r.MemoryType() == 0x10);
bool read_only = (gen5 ? false : (r.MemoryType() == 0x10));
EXIT_NOT_IMPLEMENTED(read_only && !(textures.desc[i].usage == ShaderTextureUsage::ReadOnly));
TileSizeAlign size {};
auto addr = r.Base();
auto addr = (gen5 ? r.Base40() : r.Base38());
bool neo = Config::IsNeo();
auto width = r.Width() + 1;
auto height = r.Height() + 1;
auto pitch = r.Pitch() + 1;
auto width = (gen5 ? r.Width5() : r.Width4()) + 1;
auto height = (gen5 ? r.Height5() : r.Height4()) + 1;
auto pitch = (gen5 ? width : r.Pitch() + 1);
auto base_level = r.BaseLevel();
auto levels = r.LastLevel() + 1;
auto tile = r.TilingIdx();
auto dfmt = r.Dfmt();
auto nfmt = r.Nfmt();
auto tile = r.TileMode();
auto dfmt = (gen5 ? 0 : r.Dfmt());
auto nfmt = (gen5 ? 0 : r.Nfmt());
auto fmt = (gen5 ? r.Format() : 0);
uint32_t swizzle = r.DstSelXYZW();
bool check_depth_texture = (tile == 2);
bool check_depth_texture = (!gen5 && tile == 2);
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, &size, nullptr, nullptr);
if (gen5)
{
TileGetTextureSize2(fmt, width, height, pitch, levels, tile, &size, nullptr, nullptr);
} else
{
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, &size, nullptr, nullptr);
}
EXIT_NOT_IMPLEMENTED(size.size == 0);
EXIT_NOT_IMPLEMENTED((addr & (static_cast<uint64_t>(size.align) - 1u)) != 0);
@@ -4149,14 +4339,14 @@ static void PrepareTextures(uint64_t submit_id, CommandBuffer* buffer, const Sha
{
EXIT_NOT_IMPLEMENTED(textures.desc[i].usage != ShaderTextureUsage::ReadWrite);
StorageTextureObject vulkan_texture_info(dfmt, nfmt, width, height, pitch, base_level, levels, tile, neo, swizzle);
StorageTextureObject vulkan_texture_info(dfmt, nfmt, fmt, width, height, pitch, base_level, levels, tile, neo, swizzle);
tex = static_cast<StorageTextureVulkanImage*>(
GpuMemoryCreateObject(submit_id, g_render_ctx->GetGraphicCtx(), buffer, addr, size.size, vulkan_texture_info));
} else
{
EXIT_NOT_IMPLEMENTED(textures.desc[i].usage != ShaderTextureUsage::ReadOnly);
if (tile == 10)
if (!gen5 && tile == 10)
{
RenderTextureFormat rt_format = RenderTextureFormat::Unknown;
if (dfmt == 10 && nfmt == 0)
@@ -4169,7 +4359,7 @@ static void PrepareTextures(uint64_t submit_id, CommandBuffer* buffer, const Sha
submit_id, g_render_ctx->GetGraphicCtx(), buffer, addr, size.size, vulkan_buffer_info));
} else
{
TextureObject vulkan_texture_info(dfmt, nfmt, width, height, pitch, base_level, levels, tile, neo, swizzle);
TextureObject vulkan_texture_info(dfmt, nfmt, fmt, width, height, pitch, base_level, levels, tile, neo, swizzle);
tex = static_cast<TextureVulkanImage*>(
GpuMemoryCreateObject(submit_id, g_render_ctx->GetGraphicCtx(), buffer, addr, size.size, vulkan_texture_info));
}
@@ -4181,17 +4371,29 @@ static void PrepareTextures(uint64_t submit_id, CommandBuffer* buffer, const Sha
if (textures.desc[i].textures2d_without_sampler)
{
images_storage[index_storage] = tex;
r.UpdateAddress(index_storage);
if (gen5)
{
r.UpdateAddress40(index_storage);
} else
{
r.UpdateAddress38(index_storage);
}
index_storage++;
} else
{
images_sampled[index_sampled] = tex;
images_sampled_view[index_sampled] = (depth_texture ? VulkanImage::VIEW_DEPTH_TEXTURE : view_type);
r.UpdateAddress(index_sampled);
if (gen5)
{
r.UpdateAddress40(index_sampled);
} else
{
r.UpdateAddress38(index_sampled);
}
index_sampled++;
}
EXIT_NOT_IMPLEMENTED((r.Base() >> 32u) != 0);
EXIT_NOT_IMPLEMENTED(((gen5 ? r.Base40() : r.Base38()) >> 32u) != 0);
(*sgprs)[0] = r.fields[0];
(*sgprs)[1] = r.fields[1];
@@ -4215,6 +4417,8 @@ static void PrepareSamplers(const ShaderSamplerResources& samplers, uint64_t* sa
EXIT_IF(sgprs == nullptr);
EXIT_IF(*sgprs == nullptr);
bool gen5 = Config::IsNextGen();
for (int i = 0; i < samplers.samplers_num; i++)
{
auto r = samplers.samplers[i];
@@ -4226,8 +4430,12 @@ static void PrepareSamplers(const ShaderSamplerResources& samplers, uint64_t* sa
EXIT_NOT_IMPLEMENTED(r.DepthCompareFunc() != 0);
EXIT_NOT_IMPLEMENTED(r.ForceUnormCoords() != false);
EXIT_NOT_IMPLEMENTED(r.AnisoThreshold() != 0);
EXIT_NOT_IMPLEMENTED(r.McCoordTrunc() != false);
EXIT_NOT_IMPLEMENTED(!gen5 && r.McCoordTrunc() != false);
EXIT_NOT_IMPLEMENTED(r.ForceDegamma() != false);
EXIT_NOT_IMPLEMENTED(gen5 && r.SkipDegamma() != false);
EXIT_NOT_IMPLEMENTED(gen5 && r.PointPreclamp() != false);
EXIT_NOT_IMPLEMENTED(gen5 && r.AnisoOverride() != false);
EXIT_NOT_IMPLEMENTED(gen5 && r.BlendZeroPrt() != false);
EXIT_NOT_IMPLEMENTED(r.AnisoBias() != 0);
EXIT_NOT_IMPLEMENTED(r.TruncCoord() != false);
EXIT_NOT_IMPLEMENTED(r.DisableCubeWrap() != false);
@@ -4457,7 +4665,7 @@ void GraphicsRenderDrawIndex(uint64_t submit_id, CommandBuffer* buffer, HW::Cont
hw_print(*ctx);
hw_check(*ctx);
EXIT_NOT_IMPLEMENTED(ctx->GetShaderStages() != 0);
EXIT_NOT_IMPLEMENTED(ctx->GetShaderStages() != 0 && ctx->GetShaderStages() != 0x02002000);
VkPrimitiveTopology topology = VK_PRIMITIVE_TOPOLOGY_POINT_LIST;
@@ -4605,7 +4813,7 @@ void GraphicsRenderDrawIndexAuto(uint64_t submit_id, CommandBuffer* buffer, HW::
printf("\t flags = 0x%08" PRIx32 "\n", flags);
EXIT_NOT_IMPLEMENTED(flags != 0);
EXIT_NOT_IMPLEMENTED(ctx->GetShaderStages() != 0);
EXIT_NOT_IMPLEMENTED(ctx->GetShaderStages() != 0 && ctx->GetShaderStages() != 0x02002000);
RenderDepthInfo depth_info;
FindRenderDepthInfo(submit_id, buffer, *ctx, &depth_info);
source/emulator/src/Graphics/GraphicsRun.cpp
+48 -14
View File
@@ -5,6 +5,7 @@
#include "Kyty/Core/String.h"
#include "Kyty/Core/Threads.h"
#include "Emulator/Config.h"
#include "Emulator/Graphics/AsyncJob.h"
#include "Emulator/Graphics/GraphicContext.h"
#include "Emulator/Graphics/Graphics.h"
@@ -309,7 +310,7 @@ private:
CommandProcessor* m_compute_cp[8] = {};
ComputeRing* m_compute_ring[64] = {};
int m_done_num = 0;
std::atomic_int m_done_num = 0;
};
using hw_ctx_parser_func_t = uint32_t (*)(KYTY_HW_CTX_PARSER_ARGS);
@@ -452,7 +453,7 @@ bool Gpu::AreSubmitsAllowed()
int Gpu::GetFrameNum()
{
Core::LockGuard lock(m_mutex);
// Core::LockGuard lock(m_mutex);
return m_done_num;
}
@@ -1179,7 +1180,8 @@ void CommandProcessor::WriteAtEndOfPipe64(uint32_t cache_policy, uint32_t event_
cache_action == 0x38 && source64 && !with_interrupt)
{
GraphicsRenderWriteAtEndOfPipeWithWriteBack64(m_sumbit_id, m_buffer[m_current_buffer], static_cast<uint64_t*>(dst_gpu_addr), value);
} else if (eop_event_type == 0x04 && cache_action == 0x00 && event_index == 0x05 && source_counter && !with_interrupt)
} else if (((eop_event_type == 0x04 && event_index == 0x05) || (eop_event_type == 0x28 && event_index == 0x00)) &&
cache_action == 0x00 && source_counter && !with_interrupt)
{
GraphicsRenderWriteAtEndOfPipeClockCounter(m_sumbit_id, m_buffer[m_current_buffer], static_cast<uint64_t*>(dst_gpu_addr));
} else if ((eop_event_type == 0x04 && event_index == 0x05) && cache_action == 0x00 && source64 && with_interrupt)
@@ -1581,6 +1583,11 @@ KYTY_HW_CTX_PARSER(hw_ctx_set_color_info)
r.cmask_tile_mode_neo = KYTY_PM4_GET(buffer[0], CB_COLOR0_INFO, CMASK_ADDR_TYPE);
r.neo_mode = KYTY_PM4_GET(buffer[0], CB_COLOR0_INFO, ALT_TILE_MODE) != 0;
if (!r.neo_mode && Config::IsNextGen())
{
r.neo_mode = true;
}
cp->GetCtx()->SetColorInfo(param, r);
return 1;
@@ -1957,6 +1964,11 @@ KYTY_HW_CTX_PARSER(hw_ctx_set_render_target)
info.cmask_tile_mode_neo = KYTY_PM4_GET(buffer[4], CB_COLOR0_INFO, CMASK_ADDR_TYPE);
info.neo_mode = KYTY_PM4_GET(buffer[4], CB_COLOR0_INFO, ALT_TILE_MODE) != 0;
if (!info.neo_mode && Config::IsNextGen())
{
info.neo_mode = true;
}
// attrib.force_dest_alpha_to_one = (buffer[5] & 0x20000u) != 0;
// attrib.tile_mode = buffer[5] & 0x1fu;
// attrib.fmask_tile_mode = (buffer[5] >> 5u) & 0x1fu;
@@ -3223,21 +3235,39 @@ KYTY_CP_OP_PARSER(cp_op_release_mem)
if (custom)
{
uint32_t gcr_cntl = buffer[1];
uint32_t gcr_cntl = buffer[1] & 0xffffu;
uint32_t data_sel = (buffer[1] >> 16u) & 0xffu;
EXIT_NOT_IMPLEMENTED(gcr_cntl != 0x200);
EXIT_NOT_IMPLEMENTED((buffer[0] >> 8u) != 0x3);
EXIT_NOT_IMPLEMENTED(data_sel != 2 && data_sel != 3);
if (gcr_cntl == 0x200)
if (data_sel == 2)
{
cache_action = 0x38;
}
EXIT_NOT_IMPLEMENTED(gcr_cntl != 0x0200);
EXIT_NOT_IMPLEMENTED((buffer[0] >> 8u) != 0x3);
cache_policy = 0;
event_index = 0;
event_write_dest = 0;
event_write_source = 2;
interrupt_selector = 0;
if (gcr_cntl == 0x200)
{
cache_action = 0x38;
}
cache_policy = 0;
event_index = 0;
event_write_dest = 0;
event_write_source = 2;
interrupt_selector = 0;
} else if (data_sel == 3)
{
EXIT_NOT_IMPLEMENTED(gcr_cntl != 0x0000);
EXIT_NOT_IMPLEMENTED((buffer[0] >> 8u) != 0x0);
cache_action = 0x00;
cache_policy = 0;
event_index = 0;
event_write_dest = 0;
event_write_source = 4;
interrupt_selector = 0;
}
}
cp->WriteAtEndOfPipe64(cache_policy, event_write_dest, eop_event_type, cache_action, event_index, event_write_source, dst_gpu_addr,
@@ -3513,6 +3543,10 @@ static void graphics_init_jmp_tables_cx_indirect()
info.dcc_compression_enable = KYTY_PM4_GET(value, CB_COLOR0_INFO, DCC_ENABLE) != 0;
info.cmask_tile_mode_neo = KYTY_PM4_GET(value, CB_COLOR0_INFO, CMASK_ADDR_TYPE);
info.neo_mode = KYTY_PM4_GET(value, CB_COLOR0_INFO, ALT_TILE_MODE) != 0;
if (!info.neo_mode && Config::IsNextGen())
{
info.neo_mode = true;
}
cp->GetCtx()->SetColorInfo(slot, info);
};
}
source/emulator/src/Graphics/Objects/StorageTexture.cpp
+33 -18
View File
@@ -17,21 +17,27 @@
namespace Kyty::Libs::Graphics {
static VkFormat get_texture_format(uint32_t dfmt, uint32_t nfmt)
static VkFormat get_texture_format(uint32_t dfmt, uint32_t nfmt, uint32_t fmt)
{
if (nfmt == 9 && dfmt == 10)
if (fmt == 0)
{
return VK_FORMAT_R8G8B8A8_SRGB;
}
if (nfmt == 0 && dfmt == 10)
if (nfmt == 9 && dfmt == 10)
{
return VK_FORMAT_R8G8B8A8_SRGB;
}
if (nfmt == 0 && dfmt == 10)
{
return VK_FORMAT_R8G8B8A8_UNORM;
}
if (nfmt == 9 && dfmt == 37)
{
return VK_FORMAT_BC3_SRGB_BLOCK;
}
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
} else
{
return VK_FORMAT_R8G8B8A8_UNORM;
EXIT("unknown format: fmt = %u\n", fmt);
}
if (nfmt == 9 && dfmt == 37)
{
return VK_FORMAT_BC3_SRGB_BLOCK;
}
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
return VK_FORMAT_UNDEFINED;
}
@@ -130,8 +136,9 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
auto* vk_obj = static_cast<StorageTextureVulkanImage*>(obj);
auto tile = params[StorageTextureObject::PARAM_TILE];
auto dfmt = params[StorageTextureObject::PARAM_DFMT_NFMT] >> 32u;
auto nfmt = params[StorageTextureObject::PARAM_DFMT_NFMT] & 0xffffffffu;
auto fmt = (params[StorageTextureObject::PARAM_FORMAT] >> 16u) & 0xffffu;
auto dfmt = (params[StorageTextureObject::PARAM_FORMAT] >> 8u) & 0xffu;
auto nfmt = (params[StorageTextureObject::PARAM_FORMAT]) & 0xffu;
auto width = params[StorageTextureObject::PARAM_WIDTH_HEIGHT] >> 32u;
auto height = params[StorageTextureObject::PARAM_WIDTH_HEIGHT] & 0xffffffffu;
// auto base_level = params[StorageTextureObject::PARAM_LEVELS] >> 32u;
@@ -147,7 +154,13 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
TileSizeOffset level_sizes[16];
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, nullptr, level_sizes, nullptr);
if (fmt != 0)
{
TileGetTextureSize2(fmt, width, height, pitch, levels, tile, nullptr, level_sizes, nullptr);
} else
{
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, nullptr, level_sizes, nullptr);
}
// dbg_test_mipmaps(ctx, VK_FORMAT_BC3_SRGB_BLOCK, 512, 512);
@@ -187,6 +200,7 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
if (tile == 13)
{
// EXIT_NOT_IMPLEMENTED(pitch != width);
EXIT_NOT_IMPLEMENTED(fmt != 0);
auto* temp_buf = new uint8_t[*size];
TileConvertTiledToLinear(temp_buf, reinterpret_cast<void*>(*vaddr), TileMode::TextureTiled, dfmt, nfmt, width, height, pitch,
levels, neo);
@@ -208,8 +222,9 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
EXIT_IF(ctx == nullptr);
EXIT_IF(params == nullptr);
auto dfmt = params[StorageTextureObject::PARAM_DFMT_NFMT] >> 32u;
auto nfmt = params[StorageTextureObject::PARAM_DFMT_NFMT] & 0xffffffffu;
auto fmt = (params[StorageTextureObject::PARAM_FORMAT] >> 16u) & 0xffffu;
auto dfmt = (params[StorageTextureObject::PARAM_FORMAT] >> 8u) & 0xffu;
auto nfmt = (params[StorageTextureObject::PARAM_FORMAT]) & 0xffu;
auto width = params[StorageTextureObject::PARAM_WIDTH_HEIGHT] >> 32u;
auto height = params[StorageTextureObject::PARAM_WIDTH_HEIGHT] & 0xffffffffu;
auto base_level = params[StorageTextureObject::PARAM_LEVELS] >> 32u;
@@ -227,7 +242,7 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
components.b = get_swizzle(GetDstSel(swizzle, 2));
components.a = get_swizzle(GetDstSel(swizzle, 3));
auto pixel_format = get_texture_format(dfmt, nfmt);
auto pixel_format = get_texture_format(dfmt, nfmt, fmt);
EXIT_NOT_IMPLEMENTED(pixel_format == VK_FORMAT_UNDEFINED);
EXIT_NOT_IMPLEMENTED(width == 0);
@@ -336,7 +351,7 @@ static void delete_func(GraphicContext* ctx, void* obj, VulkanMemory* mem)
bool StorageTextureObject::Equal(const uint64_t* other) const
{
return (params[PARAM_DFMT_NFMT] == other[PARAM_DFMT_NFMT] && params[PARAM_PITCH] == other[PARAM_PITCH] &&
return (params[PARAM_FORMAT] == other[PARAM_FORMAT] && params[PARAM_PITCH] == other[PARAM_PITCH] &&
params[PARAM_WIDTH_HEIGHT] == other[PARAM_WIDTH_HEIGHT] && params[PARAM_LEVELS] == other[PARAM_LEVELS] &&
params[PARAM_TILE] == other[PARAM_TILE] && params[PARAM_NEO] == other[PARAM_NEO] &&
params[PARAM_SWIZZLE] == other[PARAM_SWIZZLE]);
source/emulator/src/Graphics/Objects/Texture.cpp
+83 -52
View File
@@ -17,41 +17,51 @@
namespace Kyty::Libs::Graphics {
static VkFormat get_texture_format(uint32_t dfmt, uint32_t nfmt)
static VkFormat get_texture_format(uint32_t dfmt, uint32_t nfmt, uint32_t fmt)
{
if (nfmt == 9 && dfmt == 10)
if (fmt == 0)
{
return VK_FORMAT_R8G8B8A8_SRGB;
}
if (nfmt == 0 && dfmt == 10)
if (nfmt == 9 && dfmt == 10)
{
return VK_FORMAT_R8G8B8A8_SRGB;
}
if (nfmt == 0 && dfmt == 10)
{
return VK_FORMAT_R8G8B8A8_UNORM;
}
if (nfmt == 0 && dfmt == 1)
{
return VK_FORMAT_R8_UNORM;
}
if (nfmt == 0 && dfmt == 3)
{
return VK_FORMAT_R8G8_UNORM;
}
if (nfmt == 9 && dfmt == 37)
{
return VK_FORMAT_BC3_SRGB_BLOCK;
}
if (nfmt == 0 && dfmt == 37)
{
return VK_FORMAT_BC3_UNORM_BLOCK;
}
if (nfmt == 0 && dfmt == 36)
{
return VK_FORMAT_BC2_UNORM_BLOCK;
}
if (nfmt == 0 && dfmt == 35)
{
return VK_FORMAT_BC1_RGBA_UNORM_BLOCK;
}
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
} else
{
return VK_FORMAT_R8G8B8A8_UNORM;
if (fmt == 56)
{
return VK_FORMAT_R8G8B8A8_UNORM;
}
EXIT("unknown format: fmt = %u\n", fmt);
}
if (nfmt == 0 && dfmt == 1)
{
return VK_FORMAT_R8_UNORM;
}
if (nfmt == 0 && dfmt == 3)
{
return VK_FORMAT_R8G8_UNORM;
}
if (nfmt == 9 && dfmt == 37)
{
return VK_FORMAT_BC3_SRGB_BLOCK;
}
if (nfmt == 0 && dfmt == 37)
{
return VK_FORMAT_BC3_UNORM_BLOCK;
}
if (nfmt == 0 && dfmt == 36)
{
return VK_FORMAT_BC2_UNORM_BLOCK;
}
if (nfmt == 0 && dfmt == 35)
{
return VK_FORMAT_BC1_RGBA_UNORM_BLOCK;
}
EXIT("unknown format: nfmt = %u, dfmt = %u\n", nfmt, dfmt);
return VK_FORMAT_UNDEFINED;
}
@@ -150,8 +160,9 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
auto* vk_obj = static_cast<TextureVulkanImage*>(obj);
auto tile = params[TextureObject::PARAM_TILE];
auto dfmt = params[TextureObject::PARAM_DFMT_NFMT] >> 32u;
auto nfmt = params[TextureObject::PARAM_DFMT_NFMT] & 0xffffffffu;
auto fmt = (params[TextureObject::PARAM_FORMAT] >> 16u) & 0xffffu;
auto dfmt = (params[TextureObject::PARAM_FORMAT] >> 8u) & 0xffu;
auto nfmt = (params[TextureObject::PARAM_FORMAT]) & 0xffu;
auto width = params[TextureObject::PARAM_WIDTH_HEIGHT] >> 32u;
auto height = params[TextureObject::PARAM_WIDTH_HEIGHT] & 0xffffffffu;
auto levels = params[TextureObject::PARAM_LEVELS] & 0xffffffffu;
@@ -162,11 +173,19 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
EXIT_NOT_IMPLEMENTED(levels >= 16);
EXIT_NOT_IMPLEMENTED(tile != 8 && tile != 13);
TileSizeOffset level_sizes[16];
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, nullptr, level_sizes, nullptr);
if (fmt != 0)
{
EXIT_NOT_IMPLEMENTED(tile != 0);
TileGetTextureSize2(fmt, width, height, pitch, levels, tile, nullptr, level_sizes, nullptr);
} else
{
EXIT_NOT_IMPLEMENTED(tile != 8 && tile != 13);
TileGetTextureSize(dfmt, nfmt, width, height, pitch, levels, tile, neo, nullptr, level_sizes, nullptr);
}
// dbg_test_mipmaps(ctx, VK_FORMAT_BC3_SRGB_BLOCK, 512, 512);
@@ -201,17 +220,27 @@ static void update_func(GraphicContext* ctx, const uint64_t* params, void* obj,
}
}
if (tile == 13)
if (fmt == 0)
{
// EXIT_NOT_IMPLEMENTED(pitch != width);
auto* temp_buf = new uint8_t[*size];
TileConvertTiledToLinear(temp_buf, reinterpret_cast<void*>(*vaddr), TileMode::TextureTiled, dfmt, nfmt, width, height, pitch,
levels, neo);
UtilFillImage(ctx, vk_obj, temp_buf, *size, regions, static_cast<uint64_t>(vk_layout));
delete[] temp_buf;
} else if (tile == 8)
if (tile == 13)
{
// EXIT_NOT_IMPLEMENTED(pitch != width);
EXIT_NOT_IMPLEMENTED(fmt != 0);
auto* temp_buf = new uint8_t[*size];
TileConvertTiledToLinear(temp_buf, reinterpret_cast<void*>(*vaddr), TileMode::TextureTiled, dfmt, nfmt, width, height, pitch,
levels, neo);
UtilFillImage(ctx, vk_obj, temp_buf, *size, regions, static_cast<uint64_t>(vk_layout));
delete[] temp_buf;
} else if (tile == 8)
{
UtilFillImage(ctx, vk_obj, reinterpret_cast<void*>(*vaddr), *size, regions, static_cast<uint64_t>(vk_layout));
}
} else
{
UtilFillImage(ctx, vk_obj, reinterpret_cast<void*>(*vaddr), *size, regions, static_cast<uint64_t>(vk_layout));
if (tile == 0)
{
UtilFillImage(ctx, vk_obj, reinterpret_cast<void*>(*vaddr), *size, regions, static_cast<uint64_t>(vk_layout));
}
}
}
@@ -392,8 +421,9 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
EXIT_IF(ctx == nullptr);
EXIT_IF(params == nullptr);
auto dfmt = params[TextureObject::PARAM_DFMT_NFMT] >> 32u;
auto nfmt = params[TextureObject::PARAM_DFMT_NFMT] & 0xffffffffu;
auto fmt = (params[TextureObject::PARAM_FORMAT] >> 16u) & 0xffffu;
auto dfmt = (params[TextureObject::PARAM_FORMAT] >> 8u) & 0xffu;
auto nfmt = (params[TextureObject::PARAM_FORMAT]) & 0xffu;
auto width = params[TextureObject::PARAM_WIDTH_HEIGHT] >> 32u;
auto height = params[TextureObject::PARAM_WIDTH_HEIGHT] & 0xffffffffu;
auto base_level = params[TextureObject::PARAM_LEVELS] >> 32u;
@@ -409,7 +439,7 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
components.b = get_swizzle(GetDstSel(swizzle, 2));
components.a = get_swizzle(GetDstSel(swizzle, 3));
auto pixel_format = get_texture_format(dfmt, nfmt);
auto pixel_format = get_texture_format(dfmt, nfmt, fmt);
EXIT_NOT_IMPLEMENTED(pixel_format == VK_FORMAT_UNDEFINED);
EXIT_NOT_IMPLEMENTED(width == 0);
@@ -504,8 +534,9 @@ static void* create2_func(GraphicContext* ctx, CommandBuffer* buffer, const uint
EXIT_IF(ctx == nullptr);
EXIT_IF(params == nullptr);
auto dfmt = params[TextureObject::PARAM_DFMT_NFMT] >> 32u;
auto nfmt = params[TextureObject::PARAM_DFMT_NFMT] & 0xffffffffu;
auto fmt = (params[TextureObject::PARAM_FORMAT] >> 16u) & 0xffffu;
auto dfmt = (params[TextureObject::PARAM_FORMAT] >> 8u) & 0xffu;
auto nfmt = (params[TextureObject::PARAM_FORMAT]) & 0xffu;
auto width = params[TextureObject::PARAM_WIDTH_HEIGHT] >> 32u;
auto height = params[TextureObject::PARAM_WIDTH_HEIGHT] & 0xffffffffu;
auto base_level = params[TextureObject::PARAM_LEVELS] >> 32u;
@@ -521,7 +552,7 @@ static void* create2_func(GraphicContext* ctx, CommandBuffer* buffer, const uint
components.b = get_swizzle(GetDstSel(swizzle, 2));
components.a = get_swizzle(GetDstSel(swizzle, 3));
auto pixel_format = get_texture_format(dfmt, nfmt);
auto pixel_format = get_texture_format(dfmt, nfmt, fmt);
EXIT_NOT_IMPLEMENTED(pixel_format == VK_FORMAT_UNDEFINED);
EXIT_NOT_IMPLEMENTED(width == 0);
@@ -628,7 +659,7 @@ static void delete_func(GraphicContext* ctx, void* obj, VulkanMemory* mem)
bool TextureObject::Equal(const uint64_t* other) const
{
return (params[PARAM_DFMT_NFMT] == other[PARAM_DFMT_NFMT] && params[PARAM_PITCH] == other[PARAM_PITCH] &&
return (params[PARAM_FORMAT] == other[PARAM_FORMAT] && params[PARAM_PITCH] == other[PARAM_PITCH] &&
params[PARAM_WIDTH_HEIGHT] == other[PARAM_WIDTH_HEIGHT] && params[PARAM_LEVELS] == other[PARAM_LEVELS] &&
params[PARAM_TILE] == other[PARAM_TILE] && params[PARAM_NEO] == other[PARAM_NEO] &&
params[PARAM_SWIZZLE] == other[PARAM_SWIZZLE]);
source/emulator/src/Graphics/Shader.cpp
+986 -2010
View File
File diff suppressed because it is too large Load Diff
source/emulator/src/Graphics/ShaderParse.cpp
+3350
View File
File diff suppressed because it is too large Load Diff
source/emulator/src/Graphics/ShaderSpirv.cpp
+2193 -1366
View File
File diff suppressed because it is too large Load Diff
source/emulator/src/Graphics/Tables/TileTextureInfo_0_56.inc
+2373
View File
File diff suppressed because it is too large Load Diff
source/emulator/src/Graphics/Tile.cpp
+219 -12
View File
@@ -501,7 +501,7 @@ void TileConvertTiledToLinear(void* dst, const void* src, TileMode mode, uint32_
}
bool TileGetDepthSize(uint32_t width, uint32_t height, uint32_t pitch, uint32_t z_format, uint32_t stencil_format, bool htile, bool neo,
TileSizeAlign* stencil_size, TileSizeAlign* htile_size, TileSizeAlign* depth_size)
bool next_gen, TileSizeAlign* stencil_size, TileSizeAlign* htile_size, TileSizeAlign* depth_size)
{
struct DepthInfo
{
@@ -585,11 +585,53 @@ bool TileGetDepthSize(uint32_t width, uint32_t height, uint32_t pitch, uint32_t
{1280, 720, 1, 1, false, true, 1280, {983040, 32768}, {0, 0}, {1966080, 65536}},
};
static const DepthInfo infos_next_gen[] = {
{3840, 2160, 3, 0, true, true, 3840, {0, 0}, {655360, 32768}, {33423360, 65536}},
{3840, 2160, 3, 0, false, true, 3840, {0, 0}, {0, 0}, {33423360, 65536}},
{1920, 1080, 3, 0, true, true, 1920, {0, 0}, {196608, 32768}, {8847360, 65536}},
{1920, 1080, 3, 0, false, true, 1920, {0, 0}, {0, 0}, {8847360, 65536}},
{1280, 720, 3, 0, true, true, 1280, {0, 0}, {131072, 32768}, {3932160, 65536}},
{1280, 720, 3, 0, false, true, 1280, {0, 0}, {0, 0}, {3932160, 65536}},
{3840, 2160, 1, 0, true, true, 3840, {0, 0}, {655360, 32768}, {16711680, 65536}},
{3840, 2160, 1, 0, false, true, 3840, {0, 0}, {0, 0}, {16711680, 65536}},
{1920, 1080, 1, 0, true, true, 2048, {0, 0}, {196608, 32768}, {4718592, 65536}},
{1920, 1080, 1, 0, false, true, 2048, {0, 0}, {0, 0}, {4718592, 65536}},
{1280, 720, 1, 0, true, true, 1280, {0, 0}, {131072, 32768}, {1966080, 65536}},
{1280, 720, 1, 0, false, true, 1280, {0, 0}, {0, 0}, {1966080, 65536}},
{3840, 2160, 3, 1, true, true, 3840, {8847360, 65536}, {655360, 32768}, {33423360, 65536}},
{3840, 2160, 3, 1, false, true, 3840, {8847360, 65536}, {0, 0}, {33423360, 65536}},
{1920, 1080, 3, 1, true, true, 1920, {2621440, 65536}, {196608, 32768}, {8847360, 65536}},
{1920, 1080, 3, 1, false, true, 1920, {2621440, 65536}, {0, 0}, {8847360, 65536}},
{1280, 720, 3, 1, true, true, 1280, {983040, 65536}, {131072, 32768}, {3932160, 65536}},
{1280, 720, 3, 1, false, true, 1280, {983040, 65536}, {0, 0}, {3932160, 65536}},
{3840, 2160, 1, 1, true, true, 3840, {8847360, 65536}, {655360, 32768}, {16711680, 65536}},
{3840, 2160, 1, 1, false, true, 3840, {8847360, 65536}, {0, 0}, {16711680, 65536}},
{1920, 1080, 1, 1, true, true, 2048, {2621440, 65536}, {196608, 32768}, {4718592, 65536}},
{1920, 1080, 1, 1, false, true, 2048, {2621440, 65536}, {0, 0}, {4718592, 65536}},
{1280, 720, 1, 1, true, true, 1280, {983040, 65536}, {131072, 32768}, {1966080, 65536}},
{1280, 720, 1, 1, false, true, 1280, {983040, 65536}, {0, 0}, {1966080, 65536}},
};
EXIT_IF(depth_size == nullptr);
EXIT_IF(htile_size == nullptr);
EXIT_IF(stencil_size == nullptr);
if (neo)
if (next_gen)
{
EXIT_IF(pitch != 0);
EXIT_IF(!neo);
for (const auto& i: infos_next_gen)
{
if (i.width == width && i.height == height /*&& i.pitch == pitch*/ && i.tile == htile && i.z_format == z_format &&
i.stencil_format == stencil_format)
{
*depth_size = i.depth;
*htile_size = i.htile;
*stencil_size = i.stencil;
return true;
}
}
} else if (neo)
{
for (const auto& i: infos_neo)
{
@@ -719,6 +761,12 @@ struct TextureInfoSize
uint32_t mipmap_size = 0;
};
struct TextureInfoSize2
{
uint32_t mipmap_offset = 0;
uint32_t mipmap_size = 0;
};
struct TextureInfo
{
uint32_t dfmt = 0;
@@ -733,6 +781,21 @@ struct TextureInfo
TextureInfoPadded padded[16];
};
struct TextureInfo2
{
uint32_t fmt = 0;
uint32_t width = 0;
uint32_t height = 0;
uint32_t pitch = 0;
uint32_t levels = 0;
uint32_t tile = 0;
uint32_t total_size = 0;
uint32_t total_align = 0;
TextureInfoSize2 size[16];
TextureInfoPadded padded[16];
};
#include "Tables/TileTextureInfo_0_56.inc"
#include "Tables/TileTextureInfo_10_10_0.inc"
#include "Tables/TileTextureInfo_13_10_0.inc"
#include "Tables/TileTextureInfo_13_10_9.inc"
@@ -759,6 +822,8 @@ static const TextureInfo* g_info_map_8_1_0[16][16][16][2] = {};
static const TextureInfo* g_info_map_8_10_0[16][16][16][2] = {};
static const TextureInfo* g_info_map_8_10_9[16][16][16][2] = {};
static const TextureInfo2* g_info_map_0_56[16][16][16][16] = {};
template <class Map>
static void init_map(Map& map, const TextureInfo* info, int num)
{
@@ -786,7 +851,41 @@ static void init_map(Map& map, const TextureInfo* info, int num)
auto h = IntLog2(i.height);
auto p = IntLog2(i.pitch);
auto n = i.neo ? 1 : 0;
EXIT_IF(w > 16 || h > 16 || p > 16 || n > 2);
EXIT_IF(w >= 16 || h >= 16 || p >= 16 || n >= 2);
EXIT_IF(map[w][h][p][n] != nullptr);
map[w][h][p][n] = &i;
}
}
}
template <class Map>
static void init_map2(Map& map, const TextureInfo2* info, int num)
{
for (int w = 0; w < 16; w++)
{
for (int h = 0; h < 16; h++)
{
for (int p = 0; p < 16; p++)
{
for (int n = 0; n < 16; n++)
{
map[w][h][p][n] = nullptr;
}
}
}
}
for (int index = 0; index < num; index++)
{
const auto& i = info[index];
if (!(i.width == 0 && i.height == 0 && i.pitch == 0))
{
EXIT_IF(!(IsPowerOfTwo(i.width) && IsPowerOfTwo(i.height) && IsPowerOfTwo(i.pitch)));
auto w = IntLog2(i.width);
auto h = IntLog2(i.height);
auto p = IntLog2(i.pitch);
auto n = i.levels;
EXIT_IF(w >= 16 || h >= 16 || p >= 16 || n >= 16);
EXIT_IF(map[w][h][p][n] != nullptr);
map[w][h][p][n] = &i;
}
@@ -796,15 +895,20 @@ static void init_map(Map& map, const TextureInfo* info, int num)
template <class Map>
static const TextureInfo* check_map(Map& map, uint32_t width, uint32_t height, uint32_t pitch, bool neo)
{
// if (IsPowerOfTwo(width) && IsPowerOfTwo(height) && IsPowerOfTwo(pitch))
{
auto w = IntLog2(width);
auto h = IntLog2(height);
auto p = IntLog2(pitch);
auto n = neo ? 1 : 0;
return map[w][h][p][n];
}
// return nullptr;
auto w = IntLog2(width);
auto h = IntLog2(height);
auto p = IntLog2(pitch);
auto n = neo ? 1 : 0;
return map[w][h][p][n];
}
template <class Map>
static const TextureInfo2* check_map2(Map& map, uint32_t width, uint32_t height, uint32_t pitch, uint32_t levels)
{
auto w = IntLog2(width);
auto h = IntLog2(height);
auto p = IntLog2(pitch);
return map[w][h][p][levels];
}
static void init_maps()
@@ -821,6 +925,8 @@ static void init_maps()
init_map(g_info_map_8_1_0, infos_8_1_0_pow2, std::size(infos_8_1_0_pow2));
init_map(g_info_map_8_10_0, infos_8_10_0_pow2, std::size(infos_8_10_0_pow2));
init_map(g_info_map_8_10_9, infos_8_10_9_pow2, std::size(infos_8_10_9_pow2));
init_map2(g_info_map_0_56, infos_0_56_pow2, std::size(infos_0_56_pow2));
}
#define KYTY_TEXTURE_CHECK(n) \
@@ -841,6 +947,24 @@ static void init_maps()
*num = std::size(infos_##n); \
}
#define KYTY_TEXTURE_CHECK2(n) \
if (pow2) \
{ \
if (const auto* i = check_map2(g_info_map_##n, width, height, pitch, levels); i != nullptr) \
{ \
*info = i; \
*num = 1; \
} else \
{ \
*info = nullptr; \
*num = 0; \
} \
} else \
{ \
*info = infos_##n; \
*num = std::size(infos_##n); \
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
static void FindTextureInfo(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t tile, bool neo,
const TextureInfo** info, int* num, bool pow2)
@@ -905,6 +1029,23 @@ static void FindTextureInfo(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32
}
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
static void FindTextureInfo2(uint32_t fmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t tile, uint32_t levels,
const TextureInfo2** info, int* num, bool pow2)
{
EXIT_IF(info == nullptr);
EXIT_IF(num == nullptr);
if (tile == 0 && fmt == 56)
{
KYTY_TEXTURE_CHECK2(0_56);
} else
{
*info = nullptr;
*num = 0;
}
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
void TileGetTextureSize(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t levels, uint32_t tile,
bool neo, TileSizeAlign* total_size, TileSizeOffset* level_sizes, TilePaddedSize* padded_size)
@@ -1004,6 +1145,72 @@ void TileGetTextureSize(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t h
}
}
void TileGetTextureSize2(uint32_t format, uint32_t width, uint32_t height, uint32_t pitch, uint32_t levels, uint32_t tile,
TileSizeAlign* total_size, TileSizeOffset* level_sizes, TilePaddedSize* padded_size)
{
KYTY_PROFILER_FUNCTION();
const TextureInfo2* infos = nullptr;
int num = 0;
bool pow2 = (IsPowerOfTwo(width) && IsPowerOfTwo(height) && IsPowerOfTwo(pitch));
EXIT_IF(levels == 0 || levels > 16);
FindTextureInfo2(format, width, height, pitch, tile, levels, &infos, &num, pow2);
for (int index = 0; index < num; index++)
{
const auto& i = infos[index];
if (i.fmt == format && i.width == width && i.pitch == pitch && i.height == height && i.levels == levels && i.tile == tile)
{
if (total_size != nullptr)
{
total_size->size = i.total_size;
total_size->align = i.total_align;
}
if (level_sizes != nullptr)
{
if (levels == 1)
{
level_sizes[0].size = i.total_size;
level_sizes[0].offset = 0;
} else
{
for (uint32_t l = 0; l < levels; l++)
{
level_sizes[l].size = i.size[l].mipmap_size;
level_sizes[l].offset = i.size[l].mipmap_offset;
}
}
}
if (padded_size != nullptr)
{
for (uint32_t l = 0; l < levels; l++)
{
padded_size[l].width = i.padded[l].width;
padded_size[l].height = i.padded[l].height;
}
}
return;
}
}
if (total_size != nullptr && total_size->size == 0)
{
Core::StringList list;
list.Add(String::FromPrintf("format = %u", format));
list.Add(String::FromPrintf("width = %u", width));
list.Add(String::FromPrintf("height = %u", height));
list.Add(String::FromPrintf("pitch = %u", pitch));
list.Add(String::FromPrintf("levels = %u", levels));
list.Add(String::FromPrintf("tile = %u", tile));
EXIT("unknown format:\n%s\n", list.Concat(U'\n').C_Str());
}
}
} // namespace Kyty::Libs::Graphics
#endif // KYTY_EMU_ENABLED
source/emulator/src/Graphics/VideoOut.cpp
+2 -2
View File
@@ -236,7 +236,7 @@ static void calc_buffer_size(const VideoOutBufferAttribute* attribute, const Vid
if (attribute2 != nullptr)
{
EXIT_NOT_IMPLEMENTED(attribute2->option != 0);
EXIT_NOT_IMPLEMENTED(attribute2->option != 0 && attribute2->option != 8);
EXIT_NOT_IMPLEMENTED(attribute2->aspect_ratio != 0);
EXIT_NOT_IMPLEMENTED(attribute2->pixel_format != 0x8000000000000000ULL && attribute2->pixel_format != 0x8000000022000000ULL);
} else
@@ -1015,7 +1015,7 @@ KYTY_SYSV_ABI int VideoOutRegisterBuffers2(int handle, int set_index, int buffer
EXIT_NOT_IMPLEMENTED(attribute->tiling_mode != 0);
EXIT_NOT_IMPLEMENTED(attribute->aspect_ratio != 0);
EXIT_NOT_IMPLEMENTED(attribute->pitch_in_pixel != 0);
EXIT_NOT_IMPLEMENTED(attribute->option != 0);
EXIT_NOT_IMPLEMENTED(attribute->option != 0 && attribute->option != 8);
EXIT_NOT_IMPLEMENTED(attribute->dcc_cb_register_clear_color != 0);
EXIT_NOT_IMPLEMENTED(attribute->dcc_control != 0);
source/emulator/src/Kernel/Memory.cpp
+51
View File
@@ -52,6 +52,9 @@ public:
bool Find(uint64_t vaddr, uint64_t* base_addr, size_t* len, int* prot, VirtualMemory::Mode* mode, Graphics::GpuMemoryMode* gpu_mode);
bool Find(uint64_t phys_addr, bool next, PhysicalMemory::AllocatedBlock* out);
[[nodiscard]] Core::Mutex& GetMutex() { return m_mutex; }
[[nodiscard]] const Vector<AllocatedBlock>& GetBlocks() const { return m_allocated; }
private:
Vector<AllocatedBlock> m_allocated;
Core::Mutex m_mutex;
@@ -81,6 +84,9 @@ public:
bool Unmap(uint64_t vaddr, uint64_t size, Graphics::GpuMemoryMode* gpu_mode);
bool Find(uint64_t vaddr, uint64_t* base_addr, size_t* len, int* prot, VirtualMemory::Mode* mode, Graphics::GpuMemoryMode* gpu_mode);
[[nodiscard]] Core::Mutex& GetMutex() { return m_mutex; }
[[nodiscard]] const Vector<AllocatedBlock>& GetBlocks() const { return m_allocated; }
private:
Vector<AllocatedBlock> m_allocated;
uint64_t m_allocated_total = 0;
@@ -89,6 +95,8 @@ private:
static PhysicalMemory* g_physical_memory = nullptr;
static FlexibleMemory* g_flexible_memory = nullptr;
static callback_func_t g_alloc_callback = nullptr;
static callback_func_t g_free_callback = nullptr;
KYTY_SUBSYSTEM_INIT(Memory)
{
@@ -107,6 +115,29 @@ static uint64_t get_aligned_pos(uint64_t pos, size_t align)
return (align != 0 ? (pos + (align - 1)) & ~(align - 1) : pos);
}
void RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func)
{
EXIT_IF(g_alloc_callback != nullptr || g_free_callback != nullptr);
EXIT_IF(alloc_func == nullptr || free_func == nullptr);
g_alloc_callback = alloc_func;
g_free_callback = free_func;
g_physical_memory->GetMutex().Lock();
for (const auto& b: g_physical_memory->GetBlocks())
{
g_alloc_callback(b.map_vaddr, b.map_size);
}
g_physical_memory->GetMutex().Unlock();
g_flexible_memory->GetMutex().Lock();
for (const auto& b: g_flexible_memory->GetBlocks())
{
g_alloc_callback(b.map_vaddr, b.map_size);
}
g_flexible_memory->GetMutex().Unlock();
}
bool PhysicalMemory::Alloc(uint64_t search_start, uint64_t search_end, size_t len, size_t alignment, uint64_t* phys_addr_out,
int memory_type)
{
@@ -445,6 +476,11 @@ int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t le
Graphics::GpuMemorySetAllocatedRange(out_addr, len);
}
if (g_alloc_callback != nullptr)
{
g_alloc_callback(out_addr, len);
}
return OK;
}
@@ -490,6 +526,11 @@ int KYTY_SYSV_ABI KernelMunmap(uint64_t vaddr, size_t len)
Graphics::GpuMemoryFree(Graphics::WindowGetGraphicContext(), vaddr, len, true);
}
if (g_free_callback != nullptr)
{
g_free_callback(vaddr, len);
}
return OK;
}
@@ -639,6 +680,11 @@ int KYTY_SYSV_ABI KernelReleaseDirectMemory(int64_t start, size_t len)
Graphics::GpuMemoryFree(Graphics::WindowGetGraphicContext(), vaddr, size, true);
}
if (g_free_callback != nullptr)
{
g_free_callback(vaddr, len);
}
return OK;
}
@@ -722,6 +768,11 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
Graphics::GpuMemorySetAllocatedRange(out_addr, len);
}
if (g_alloc_callback != nullptr)
{
g_alloc_callback(out_addr, len);
}
printf(FG_GREEN "\t [Ok]\n" FG_DEFAULT);
return OK;
source/emulator/src/Kyty.cpp
+26
View File
@@ -46,6 +46,15 @@ static void load_symbols(const String& id, Loader::RuntimeLinker* rt)
}
}
static void load_symbols_all(Loader::RuntimeLinker* rt)
{
KYTY_PROFILER_FUNCTION();
EXIT_IF(rt == nullptr);
Libs::InitAll(rt->Symbols());
}
static void print_system_info()
{
Loader::SystemInfo info = Loader::GetSystemInfo();
@@ -212,6 +221,22 @@ KYTY_SCRIPT_FUNC(kyty_load_symbols_func)
return 0;
}
KYTY_SCRIPT_FUNC(kyty_load_symbols_all_func)
{
auto count = Scripts::ArgGetVarCount();
if (count != 0)
{
EXIT("invalid args\n");
}
auto* rt = Core::Singleton<Loader::RuntimeLinker>::Instance();
load_symbols_all(rt);
return 0;
}
KYTY_SCRIPT_FUNC(kyty_load_param_sfo_func)
{
auto count = Scripts::ArgGetVarCount();
@@ -405,6 +430,7 @@ void kyty_reg()
Scripts::RegisterFunc("kyty_load_elf", LuaFunc::kyty_load_elf_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_save_main_elf", LuaFunc::kyty_save_main_elf_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_load_symbols", LuaFunc::kyty_load_symbols_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_load_symbols_all", LuaFunc::kyty_load_symbols_all_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_load_param_sfo", LuaFunc::kyty_load_param_sfo_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_dbg_dump", LuaFunc::kyty_dbg_dump_func, LuaFunc::kyty_help);
Scripts::RegisterFunc("kyty_execute", LuaFunc::kyty_execute_func, LuaFunc::kyty_help);
source/emulator/src/Libs/LibDbgAsan.cpp
+251
View File
@@ -0,0 +1,251 @@
#include "Kyty/Core/Common.h"
#include "Kyty/Core/DbgAssert.h"
#include "Kyty/Core/String.h"
#include "Kyty/Core/Threads.h"
#include "Kyty/Core/Vector.h"
#include "Kyty/Sys/SysDbg.h"
#include "Emulator/Common.h"
#include "Emulator/Kernel/Memory.h"
#include "Emulator/Libs/Libs.h"
#include "Emulator/Loader/SymbolDatabase.h"
#include "Emulator/Loader/VirtualMemory.h"
#include <algorithm>
#ifdef KYTY_EMU_ENABLED
namespace Kyty::Libs {
LIB_VERSION("DbgAddressSanitizer", 1, "DbgAddressSanitizer", 1, 1);
namespace DbgAddressSanitizer {
struct SourceLocation
{
const char* filename;
int line_no;
int column_no;
};
struct Global
{
uintptr_t beg;
uintptr_t size;
uintptr_t size_with_redzone;
const char* name;
const char* module_name;
uintptr_t has_dynamic_init;
SourceLocation* location;
uintptr_t odr_indicator;
};
struct Shadow
{
uintptr_t start = 0;
int count = 0;
};
struct AsanContext
{
Core::Mutex mutex;
Vector<Shadow> shadows;
};
static uint32_t g_fake_stack = 0;
static AsanContext* g_ctx = nullptr;
static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr)
{
EXIT_IF(g_ctx == nullptr);
Core::LockGuard lock(g_ctx->mutex);
static const uintptr_t page_size = 0x10000ull;
auto min_shadow = ((min_addr >> 3u) + 0x10000000000ull) & ~(page_size - 1);
auto max_shadow = ((max_addr >> 3u) + 0x10000000000ull) & ~(page_size - 1);
auto size = max_shadow - min_shadow + page_size;
printf("\t shadow = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(min_shadow));
printf("\t size = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(size));
for (uintptr_t page = min_shadow; page <= max_shadow; page += page_size)
{
if (auto index = g_ctx->shadows.Find(page, [](auto& s, auto p) { return s.start == p; }); g_ctx->shadows.IndexValid(index))
{
g_ctx->shadows[index].count++;
} else
{
if (!Loader::VirtualMemory::AllocFixed(page, page_size, Loader::VirtualMemory::Mode::ReadWrite))
{
EXIT("can't allocate shadow memory\n");
}
Shadow s;
s.start = page;
s.count = 1;
g_ctx->shadows.Add(s);
}
}
}
static void FreeShadow(uintptr_t min_addr, uintptr_t max_addr)
{
EXIT_IF(g_ctx == nullptr);
Core::LockGuard lock(g_ctx->mutex);
static const uintptr_t page_size = 0x10000ull;
auto min_shadow = ((min_addr >> 3u) + 0x10000000000ull) & ~(page_size - 1);
auto max_shadow = ((max_addr >> 3u) + 0x10000000000ull) & ~(page_size - 1);
auto size = max_shadow - min_shadow + page_size;
printf("\t shadow = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(min_shadow));
printf("\t size = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(size));
for (uintptr_t page = min_shadow; page <= max_shadow; page += page_size)
{
if (auto index = g_ctx->shadows.Find(page, [](auto& s, auto p) { return s.start == p; }); g_ctx->shadows.IndexValid(index))
{
g_ctx->shadows[index].count--;
if (g_ctx->shadows[index].count <= 0)
{
if (!Loader::VirtualMemory::Free(page))
{
EXIT("can't free shadow memory\n");
}
g_ctx->shadows.RemoveAt(index);
}
}
}
}
static void KernelAlloc(uintptr_t addr, size_t size)
{
AllocShadow(addr, addr + size - 1);
}
static void KernelFree(uintptr_t addr, size_t size)
{
FreeShadow(addr, addr + size - 1);
}
static void KYTY_SYSV_ABI asan_init()
{
PRINT_NAME();
EXIT_NOT_IMPLEMENTED(g_ctx != nullptr);
g_ctx = new AsanContext;
sys_dbg_stack_info_t s {};
sys_stack_usage(s);
AllocShadow(s.reserved_addr, reinterpret_cast<uintptr_t>(&s));
LibKernel::Memory::RegisterCallbacks(KernelAlloc, KernelFree);
}
static void KYTY_SYSV_ABI asan_version_mismatch_check_v8()
{
PRINT_NAME();
}
static void KYTY_SYSV_ABI asan_register_elf_globals(uintptr_t* flag, Global* start, Global* stop)
{
PRINT_NAME();
printf("\t flag = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(flag));
printf("\t start = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(start));
printf("\t stop = %016" PRIx64 "\n", reinterpret_cast<uint64_t>(stop));
if (flag == nullptr || start == nullptr || stop == nullptr)
{
return;
}
if (*flag != 0)
{
return;
}
auto num = stop - start;
if (num != 0)
{
uintptr_t min_addr = UINTPTR_MAX;
uintptr_t max_addr = 0;
for (ptrdiff_t i = 0; i < num; i++)
{
auto* g = start + i;
printf("asan global:\n");
printf("\t beg = %016" PRIx64 "\n", static_cast<uint64_t>(g->beg));
printf("\t size = %016" PRIx64 "\n", static_cast<uint64_t>(g->size));
printf("\t size_with_redzone = %016" PRIx64 "\n", static_cast<uint64_t>(g->size_with_redzone));
printf("\t name = %s\n", g->name);
printf("\t module_name = %s\n", g->module_name);
printf("\t has_dynamic_init = %" PRIu64 "\n", static_cast<uint64_t>(g->has_dynamic_init));
printf("\t odr_indicator = %016" PRIx64 "\n", static_cast<uint64_t>(g->odr_indicator));
if (g->location != nullptr)
{
printf("\t location = %s:%d\n", g->location->filename, g->location->line_no);
}
min_addr = std::min(g->beg, min_addr);
max_addr = std::max(g->beg + g->size_with_redzone - 1, max_addr);
}
AllocShadow(min_addr, max_addr);
}
*flag = 0;
}
static void KYTY_SYSV_ABI asan_before_dynamic_init(const char* module_name)
{
PRINT_NAME();
printf("\t name = %s\n", module_name);
}
static void KYTY_SYSV_ABI asan_after_dynamic_init()
{
PRINT_NAME();
}
static void* KYTY_SYSV_ABI asan_memcpy(void* dst, const void* src, size_t size)
{
return std::memcpy(dst, src, size);
}
static void* KYTY_SYSV_ABI asan_memset(void* dst, int ch, size_t size)
{
return std::memset(dst, ch, size);
}
} // namespace DbgAddressSanitizer
LIB_DEFINE(InitDbgAddressSanitizer_1)
{
LIB_OBJECT("OmecxEgrhCs", &DbgAddressSanitizer::g_fake_stack);
LIB_FUNC("xLoVJKHvRMU", DbgAddressSanitizer::asan_init);
LIB_FUNC("GY2I-Okc7q0", DbgAddressSanitizer::asan_version_mismatch_check_v8);
LIB_FUNC("W+2HQAGdPuU", DbgAddressSanitizer::asan_register_elf_globals);
LIB_FUNC("G9xSYlaNv+Y", DbgAddressSanitizer::asan_before_dynamic_init);
LIB_FUNC("z5Gy5T3tptE", DbgAddressSanitizer::asan_after_dynamic_init);
LIB_FUNC("g+Neom2EHO8", DbgAddressSanitizer::asan_memcpy);
LIB_FUNC("q-hZf47nHhQ", DbgAddressSanitizer::asan_memset);
}
} // namespace Kyty::Libs
#endif // KYTY_EMU_ENABLED
source/emulator/src/Libs/LibGraphicsDriver.cpp
+1
View File
@@ -85,6 +85,7 @@ LIB_DEFINE(InitGraphicsDriver_1)
LIB_FUNC("wr23dPKyWc0", Gen5::GraphicsCbReleaseMem);
LIB_FUNC("TRO721eVt4g", Gen5::GraphicsDcbResetQueue);
LIB_FUNC("MWiElSNE8j8", Gen5::GraphicsDcbWaitUntilSafeForRendering);
LIB_FUNC("pFLArOT53+w", Gen5::GraphicsDcbSetShRegisterDirect);
LIB_FUNC("ZvwO9euwYzc", Gen5::GraphicsDcbSetCxRegistersIndirect);
LIB_FUNC("-HOOCn0JY48", Gen5::GraphicsDcbSetShRegistersIndirect);
LIB_FUNC("hvUfkUIQcOE", Gen5::GraphicsDcbSetUcRegistersIndirect);
source/emulator/src/Libs/LibKernel.cpp
+21 -2
View File
@@ -243,11 +243,16 @@ static void KYTY_SYSV_ABI KernelRtldSetApplicationHeapAPI(void* api[])
[[maybe_unused]] auto* heap_posix_memalign = api[6];
}
static int KYTY_SYSV_ABI write(int d, const char* str, int64_t size)
static int64_t KYTY_SYSV_ABI write(int d, const char* str, int64_t size)
{
// PRINT_NAME();
EXIT_NOT_IMPLEMENTED(d < 0 || d > 2);
EXIT_NOT_IMPLEMENTED(d < 0);
if (d > 2)
{
return POSIX_N_CALL(FileSystem::KernelWrite(d, str, size));
}
int size_int = static_cast<int>(size);
@@ -256,6 +261,18 @@ static int KYTY_SYSV_ABI write(int d, const char* str, int64_t size)
return size_int;
}
static int KYTY_SYSV_ABI open(const char* path, int flags, int mode)
{
return POSIX_N_CALL(FileSystem::KernelOpen(path, flags, mode));
}
static int KYTY_SYSV_ABI close(int d)
{
EXIT_NOT_IMPLEMENTED(d <= 2);
return POSIX_CALL(FileSystem::KernelClose(d));
}
static int64_t KYTY_SYSV_ABI read(int d, void* buf, uint64_t nbytes)
{
// PRINT_NAME();
@@ -623,6 +640,7 @@ LIB_DEFINE(InitLibKernel_1)
LIB_OBJECT("djxxOmW6-aw", &LibKernel::g_progname);
LIB_FUNC("1jfXLRVzisc", LibKernel::KernelUsleep);
LIB_FUNC("6c3rCVE-fTU", LibKernel::open);
LIB_FUNC("6xVpy0Fdq+I", LibKernel::sigprocmask);
LIB_FUNC("6Z83sYWFlA8", LibKernel::exit);
LIB_FUNC("8OnWXlgQlvo", LibKernel::KernelRtldThreadAtexitDecrement);
@@ -636,6 +654,7 @@ LIB_DEFINE(InitLibKernel_1)
LIB_FUNC("FxVZqBAA7ks", LibKernel::write);
LIB_FUNC("kbw4UHHSYy0", LibKernel::pthread_cxa_finalize);
LIB_FUNC("lLMT9vJAck0", LibKernel::clock_gettime);
LIB_FUNC("NNtFaKJbPt0", LibKernel::close);
LIB_FUNC("OMDRKKAZ8I4", LibKernel::KernelDebugRaiseException);
LIB_FUNC("Ou3iL1abvng", LibKernel::stack_chk_fail);
LIB_FUNC("p5EcQeEeJAE", LibKernel::KernelRtldSetApplicationHeapAPI);
source/emulator/src/Libs/Libs.cpp
+24 -1
View File
@@ -10,6 +10,7 @@ LIB_DEFINE(InitLibcInternal_1);
LIB_DEFINE(InitAppContent_1);
LIB_DEFINE(InitAudio_1);
LIB_DEFINE(InitDbgAddressSanitizer_1);
LIB_DEFINE(InitDebug_1);
LIB_DEFINE(InitDialog_1);
LIB_DEFINE(InitDiscMap_1);
@@ -26,9 +27,10 @@ LIB_DEFINE(InitVideoOut_1);
bool Init(const String& id, Loader::SymbolDatabase* s)
{
LIB_CHECK(U"libAudio_1", InitAudio_1);
LIB_CHECK(U"libAppContent_1", InitAppContent_1);
LIB_CHECK(U"libAudio_1", InitAudio_1);
LIB_CHECK(U"libc_internal_1", LibcInternal::InitLibcInternal_1);
LIB_CHECK(U"libDbgAddressSanitizer_1", InitDbgAddressSanitizer_1);
LIB_CHECK(U"libDebug_1", InitDebug_1);
LIB_CHECK(U"libDialog_1", InitDialog_1);
LIB_CHECK(U"libDiscMap_1", InitDiscMap_1);
@@ -46,6 +48,27 @@ bool Init(const String& id, Loader::SymbolDatabase* s)
return false;
}
void InitAll(Loader::SymbolDatabase* s)
{
LIB_LOAD(InitAudio_1);
LIB_LOAD(InitAppContent_1);
LIB_LOAD(LibcInternal::InitLibcInternal_1);
LIB_LOAD(InitDbgAddressSanitizer_1);
LIB_LOAD(InitDebug_1);
LIB_LOAD(InitDialog_1);
LIB_LOAD(InitDiscMap_1);
LIB_LOAD(InitGraphicsDriver_1);
LIB_LOAD(InitLibKernel_1);
LIB_LOAD(InitNet_1);
LIB_LOAD(InitPad_1);
LIB_LOAD(InitPlayGo_1);
LIB_LOAD(InitSaveData_1);
LIB_LOAD(InitSysmodule_1);
LIB_LOAD(InitSystemService_1);
LIB_LOAD(InitUserService_1);
LIB_LOAD(InitVideoOut_1);
}
} // namespace Kyty::Libs
#endif // KYTY_EMU_ENABLED
source/emulator/src/Loader/RuntimeLinker.cpp
+13 -3
View File
@@ -9,6 +9,7 @@
#include "Kyty/Core/Threads.h"
#include "Kyty/Sys/SysDbg.h"
#include "Emulator/Config.h"
#include "Emulator/Graphics/Objects/GpuMemory.h"
#include "Emulator/Kernel/Pthread.h"
#include "Emulator/Loader/Elf.h"
@@ -211,13 +212,16 @@ static void kyty_exception_handler(const VirtualMemory::ExceptionHandler::Except
return;
}
Core::Singleton<Loader::RuntimeLinker>::Instance()->StackTrace(info->rbp);
if (info->rbp != 0)
{
Core::Singleton<Loader::RuntimeLinker>::Instance()->StackTrace(info->rbp);
}
EXIT("Access violation: %s [%016" PRIx64 "] %s\n", Core::EnumName(info->access_violation_type).C_Str(),
info->access_violation_vaddr, (info->access_violation_vaddr == g_invalid_memory ? "(Unpatched object)" : ""));
}
EXIT("Unknown exception!!!");
EXIT("Unknown exception!!! (%08" PRIx32 ")", info->exception_win_code);
}
static void encode_id_64(uint16_t in_id, String* out_id)
@@ -1105,6 +1109,7 @@ static uint64_t calc_base_size(const Elf64_Ehdr* ehdr, const Elf64_Phdr* phdr)
return base_size;
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
void RuntimeLinker::LoadProgramToMemory(Program* program)
{
KYTY_PROFILER_FUNCTION();
@@ -1122,6 +1127,11 @@ void RuntimeLinker::LoadProgramToMemory(Program* program)
EXIT_IF(phdr == nullptr || ehdr == nullptr);
if (is_next_gen && !is_shared)
{
Config::SetNextGen(true);
}
program->base_size = calc_base_size(ehdr, phdr);
program->base_size_aligned = (program->base_size & ~(static_cast<uint64_t>(0x1000) - 1)) + 0x1000;
@@ -1163,7 +1173,7 @@ void RuntimeLinker::LoadProgramToMemory(Program* program)
SYSTEM_RESERVED,
kyty_exception_handler);
if (Libs::Graphics::GpuMemoryWatcherEnabled())
// if (Libs::Graphics::GpuMemoryWatcherEnabled())
{
VirtualMemory::ExceptionHandler::InstallVectored(kyty_exception_handler);
}
source/emulator/src/Loader/VirtualMemory.cpp
+18 -1
View File
@@ -115,7 +115,8 @@ public:
info.access_violation_vaddr = exception_record->ExceptionInformation[1];
}
info.rbp = dispatcher_context->ContextRecord->Rbp;
info.rbp = dispatcher_context->ContextRecord->Rbp;
info.exception_win_code = exception_record->ExceptionCode;
auto* p = *static_cast<ExceptionHandlerPrivate**>(dispatcher_context->HandlerData);
p->func(&info);
@@ -205,6 +206,19 @@ static LONG WINAPI ExceptionFilter(PEXCEPTION_POINTERS exception)
info.exception_address = reinterpret_cast<uint64_t>(exception_record->ExceptionAddress);
// printf("exception_record->ExceptionCode = %u\n", static_cast<uint32_t>(exception_record->ExceptionCode));
if (exception_record->ExceptionCode == DBG_PRINTEXCEPTION_C || exception_record->ExceptionCode == DBG_PRINTEXCEPTION_WIDE_C)
{
return EXCEPTION_CONTINUE_EXECUTION;
}
if (exception_record->ExceptionCode == 0x406D1388)
{
// Set a thread name
return EXCEPTION_CONTINUE_EXECUTION;
}
if (exception_record->ExceptionCode == EXCEPTION_ACCESS_VIOLATION)
{
info.type = ExceptionHandler::ExceptionType::AccessViolation;
@@ -218,6 +232,9 @@ static LONG WINAPI ExceptionFilter(PEXCEPTION_POINTERS exception)
info.access_violation_vaddr = exception_record->ExceptionInformation[1];
}
info.rbp = exception->ContextRecord->Rbp;
info.exception_win_code = exception_record->ExceptionCode;
ExceptionHandlerPrivate::g_vec_func(&info);
return EXCEPTION_CONTINUE_EXECUTION;