YouHaveTrouble/Kyty
1
0
Fork 0
mirror of https://github.com/InoriRus/Kyty.git synced 2026-08-28 13:16:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

PS5 graphics (#42)

Browse Source
This commit is contained in:
InoriRus
2022-07-31 21:09:53 +03:00
parent 96b0202858
commit e110bd0f63
11 changed files with 2232 additions and 535 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/emulator/src/Graphics/Graphics.cpp
+44 -23
View File
@@ -1556,13 +1556,34 @@ int KYTY_SYSV_ABI GraphicsCreateShader(Shader** dst, void* header, const volatil
h->code = code;
dbg_dump_shader(h);
EXIT_NOT_IMPLEMENTED(h->file_header != 0x34333231);
EXIT_NOT_IMPLEMENTED(h->version != 0x00000018);
auto base = reinterpret_cast<uint64_t>(code);
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)
{
h->sh_registers[0].offset = Pm4::SPI_SHADER_PGM_LO_ES;
h->sh_registers[0].value = ((base >> 8u) & 0xffffffffu);
h->sh_registers[1].offset = Pm4::SPI_SHADER_PGM_HI_ES;
h->sh_registers[1].value = ((base >> 40u) & 0x000000ffu);
} else if (h->type == 1 && h->num_sh_registers >= 2 && h->sh_registers[0].offset == Pm4::SPI_SHADER_PGM_LO_PS &&
h->sh_registers[1].offset == Pm4::SPI_SHADER_PGM_HI_PS)
{
h->sh_registers[0].offset = Pm4::SPI_SHADER_PGM_LO_PS;
h->sh_registers[0].value = ((base >> 8u) & 0xffffffffu);
h->sh_registers[1].offset = Pm4::SPI_SHADER_PGM_HI_PS;
h->sh_registers[1].value = ((base >> 40u) & 0x000000ffu);
} else
{
EXIT("invalid shader\n");
}
*dst = h;
dbg_dump_shader(h);
return OK;
}
@@ -1774,6 +1795,10 @@ uint32_t* KYTY_SYSV_ABI GraphicsCbSetShRegisterRangeDirect(CommandBuffer* buf, u
buf->DbgDump();
auto* marker = buf->AllocateDW(2);
marker[0] = KYTY_PM4(2, Pm4::IT_NOP, Pm4::R_ZERO);
marker[1] = 0x6875000d;
auto* cmd = buf->AllocateDW(num_values + 2);
EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -1815,23 +1840,22 @@ uint32_t* KYTY_SYSV_ABI GraphicsCbReleaseMem(CommandBuffer* buf, uint8_t action,
EXIT_NOT_IMPLEMENTED(data_sel != 2);
EXIT_NOT_IMPLEMENTED(gds_offset != 0);
EXIT_NOT_IMPLEMENTED(gds_size != 1);
EXIT_NOT_IMPLEMENTED(interrupt != 0);
EXIT_NOT_IMPLEMENTED(interrupt_ctx_id != 0);
buf->DbgDump();
auto* cmd = buf->AllocateDW(10);
auto* cmd = buf->AllocateDW(7);
EXIT_NOT_IMPLEMENTED(cmd == nullptr);
cmd[0] = KYTY_PM4(10, Pm4::IT_NOP, Pm4::R_RELEASE_MEM);
cmd[1] = 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[3] = cache_policy;
cmd[4] = static_cast<uint32_t>(reinterpret_cast<uint64_t>(address) & 0xffffffffu);
cmd[5] = static_cast<uint32_t>((reinterpret_cast<uint64_t>(address) >> 32u) & 0xffffffffu);
cmd[6] = static_cast<uint32_t>(data & 0xffffffffu);
cmd[7] = static_cast<uint32_t>((data >> 32u) & 0xffffffffu);
cmd[8] = interrupt;
cmd[9] = interrupt_ctx_id;
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);
cmd[6] = static_cast<uint32_t>((data >> 32u) & 0xffffffffu);
return cmd;
}
@@ -2086,26 +2110,23 @@ uint32_t* KYTY_SYSV_ABI GraphicsDcbWriteData(CommandBuffer* buf, uint8_t dst, ui
printf("\t write_confirm = 0x%02" PRIx8 "\n", write_confirm);
EXIT_NOT_IMPLEMENTED(buf == nullptr);
EXIT_NOT_IMPLEMENTED((8 + num_dwords - 2u) > 0x3fffu);
EXIT_NOT_IMPLEMENTED((4 + num_dwords - 2u) > 0x3fffu);
EXIT_NOT_IMPLEMENTED(data == nullptr);
EXIT_NOT_IMPLEMENTED(address_or_offset == 0);
buf->DbgDump();
auto* cmd = buf->AllocateDW(8 + num_dwords);
auto* cmd = buf->AllocateDW(4 + num_dwords);
EXIT_NOT_IMPLEMENTED(cmd == nullptr);
cmd[0] = KYTY_PM4(8 + num_dwords, Pm4::IT_NOP, Pm4::R_WRITE_DATA);
cmd[1] = dst;
cmd[2] = cache_policy;
cmd[3] = address_or_offset & 0xffffffffu;
cmd[4] = (address_or_offset >> 32u) & 0xffffffffu;
cmd[5] = num_dwords;
cmd[6] = increment;
cmd[7] = write_confirm;
cmd[0] = KYTY_PM4(4 + num_dwords, Pm4::IT_NOP, Pm4::R_WRITE_DATA);
cmd[1] = dst | (static_cast<uint32_t>(cache_policy) << 8u) | (static_cast<uint32_t>(increment) << 16u) |
(static_cast<uint32_t>(write_confirm) << 24u);
cmd[2] = address_or_offset & 0xffffffffu;
cmd[3] = (address_or_offset >> 32u) & 0xffffffffu;
memcpy(cmd + 8, data, static_cast<size_t>(num_dwords) * 4);
memcpy(cmd + 4, data, static_cast<size_t>(num_dwords) * 4);
return cmd;
}
source/emulator/src/Graphics/GraphicsRender.cpp
+122 -23
View File
@@ -466,10 +466,27 @@ static void uc_print(const char* func, const HW::UserConfig& uc)
{
printf("%s\n", func);
printf("\t GetPrimType() = 0x%08" PRIx32 "\n", uc.GetPrimType());
const auto& ge_cntl = uc.GetGeControl();
const auto& user_en = uc.GetGeUserVgprEn();
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 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");
}
static void uc_check(const HW::UserConfig& /*uc*/) {}
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(user_en.vgpr1 != false);
EXIT_NOT_IMPLEMENTED(user_en.vgpr2 != false);
EXIT_NOT_IMPLEMENTED(user_en.vgpr3 != false);
}
static void sh_print(const char* func, const HW::Shader& /*uc*/)
{
@@ -490,8 +507,13 @@ static Core::StringList rt_print(const char* func, const HW::RenderTarget& rt)
dst.Add(String::FromPrintf("\t slice.slice_div64_minus1 = 0x%08" PRIx32 "\n", rt.slice.slice_div64_minus1));
dst.Add(String::FromPrintf("\t view.base_array_slice_index = 0x%08" PRIx32 "\n", rt.view.base_array_slice_index));
dst.Add(String::FromPrintf("\t view.last_array_slice_index = 0x%08" PRIx32 "\n", rt.view.last_array_slice_index));
dst.Add(String::FromPrintf("\t view.current_mip_level = 0x%08" PRIx32 "\n", rt.view.current_mip_level));
dst.Add(String::FromPrintf("\t info.fmask_compression_enable = %s\n", rt.info.fmask_compression_enable ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.fmask_compression_mode = 0x%08" PRIx32 "\n", rt.info.fmask_compression_mode));
// dst.Add(String::FromPrintf("\t info.fmask_compression_mode = 0x%08" PRIx32 "\n", rt.info.fmask_compression_mode));
dst.Add(String::FromPrintf("\t info.fmask_data_compression_disable = %s\n", rt.info.fmask_data_compression_disable ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.fmask_one_frag_mode = %s\n", rt.info.fmask_one_frag_mode ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.cmask_fast_clear_enable = %s\n", rt.info.cmask_fast_clear_enable ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.dcc_compression_enable = %s\n", rt.info.dcc_compression_enable ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.neo_mode = %s\n", rt.info.neo_mode ? "true" : "false"));
@@ -500,17 +522,31 @@ static Core::StringList rt_print(const char* func, const HW::RenderTarget& rt)
dst.Add(String::FromPrintf("\t info.format = 0x%08" PRIx32 "\n", rt.info.format));
dst.Add(String::FromPrintf("\t info.channel_type = 0x%08" PRIx32 "\n", rt.info.channel_type));
dst.Add(String::FromPrintf("\t info.channel_order = 0x%08" PRIx32 "\n", rt.info.channel_order));
dst.Add(String::FromPrintf("\t info.blend_bypa = %s\n", rt.info.blend_bypass ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.blend_clamp = %s\n", rt.info.blend_clamp ? "true" : "false"));
dst.Add(String::FromPrintf("\t info.round_mode = %s\n", rt.info.round_mode ? "true" : "false"));
dst.Add(String::FromPrintf("\t attrib.force_dest_alpha_to_one = %s\n", rt.attrib.force_dest_alpha_to_one ? "true" : "false"));
dst.Add(String::FromPrintf("\t attrib.tile_mode = 0x%08" PRIx32 "\n", rt.attrib.tile_mode));
dst.Add(String::FromPrintf("\t attrib.fmask_tile_mode = 0x%08" PRIx32 "\n", rt.attrib.fmask_tile_mode));
dst.Add(String::FromPrintf("\t attrib.num_samples = 0x%08" PRIx32 "\n", rt.attrib.num_samples));
dst.Add(String::FromPrintf("\t attrib.num_fragments = 0x%08" PRIx32 "\n", rt.attrib.num_fragments));
dst.Add(String::FromPrintf("\t attrib2.width = 0x%08" PRIx32 "\n", rt.attrib2.width));
dst.Add(String::FromPrintf("\t attrib2.height = 0x%08" PRIx32 "\n", rt.attrib2.height));
dst.Add(String::FromPrintf("\t attrib2.num_mip_levels = 0x%08" PRIx32 "\n", rt.attrib2.num_mip_levels));
dst.Add(String::FromPrintf("\t attrib3.depth = 0x%08" PRIx32 "\n", rt.attrib3.depth));
dst.Add(String::FromPrintf("\t attrib3.tile_mode = 0x%08" PRIx32 "\n", rt.attrib3.tile_mode));
dst.Add(String::FromPrintf("\t attrib3.dimension = 0x%08" PRIx32 "\n", rt.attrib3.dimension));
dst.Add(String::FromPrintf("\t attrib3.cmask_pipe_aligned = %s\n", rt.attrib3.cmask_pipe_aligned ? "true" : "false"));
dst.Add(String::FromPrintf("\t attrib3.dcc_pipe_aligned = %s\n", rt.attrib3.dcc_pipe_aligned ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.max_uncompressed_block_size = 0x%08" PRIx32 "\n", rt.dcc.max_uncompressed_block_size));
dst.Add(String::FromPrintf("\t dcc.max_compressed_block_size = 0x%08" PRIx32 "\n", rt.dcc.max_compressed_block_size));
dst.Add(String::FromPrintf("\t dcc.min_compressed_block_size = 0x%08" PRIx32 "\n", rt.dcc.min_compressed_block_size));
dst.Add(String::FromPrintf("\t dcc.color_transform = 0x%08" PRIx32 "\n", rt.dcc.color_transform));
dst.Add(String::FromPrintf("\t dcc.enable_overwrite_combiner = %s\n", rt.dcc.enable_overwrite_combiner ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.force_independent_blocks = %s\n", rt.dcc.force_independent_blocks ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.overwrite_combiner_disable = %s\n", rt.dcc.overwrite_combiner_disable ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.independent_64b_blocks = %s\n", rt.dcc.independent_64b_blocks ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.independent_128b_blocks = %s\n", rt.dcc.independent_128b_blocks ? "true" : "false"));
dst.Add(String::FromPrintf("\t data_write_on_dcc_clear_to_reg = %s\n", rt.dcc.data_write_on_dcc_clear_to_reg ? "true" : "false"));
dst.Add(String::FromPrintf("\t dcc.dcc_clear_key_enable = %s\n", rt.dcc.dcc_clear_key_enable ? "true" : "false"));
dst.Add(String::FromPrintf("\t cmask.addr = 0x%016" PRIx64 "\n", rt.cmask.addr));
dst.Add(String::FromPrintf("\t cmask_slice.slice_minus1 = 0x%08" PRIx32 "\n", rt.cmask_slice.slice_minus1));
dst.Add(String::FromPrintf("\t fmask.addr = 0x%016" PRIx64 "\n", rt.fmask.addr));
@@ -536,13 +572,21 @@ static void rt_check(const HW::RenderTarget& rt)
// EXIT_NOT_IMPLEMENTED(rt.slice_div64_minus1 != 0x000086ff);
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);
EXIT_NOT_IMPLEMENTED(rt.info.fmask_compression_enable != false);
EXIT_NOT_IMPLEMENTED(rt.info.fmask_compression_mode != 0x00000000);
// EXIT_NOT_IMPLEMENTED(rt.info.fmask_compression_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.info.fmask_data_compression_disable != false);
EXIT_NOT_IMPLEMENTED(rt.info.fmask_one_frag_mode != false);
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.info.cmask_tile_mode != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.info.cmask_tile_mode_neo != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.info.blend_bypass != false);
// EXIT_NOT_IMPLEMENTED(rt.info.blend_clamp != false);
EXIT_NOT_IMPLEMENTED(rt.info.round_mode != false);
// EXIT_NOT_IMPLEMENTED(rt.format != 0x0000000a);
// EXIT_NOT_IMPLEMENTED(rt.channel_type != 0x00000006);
// EXIT_NOT_IMPLEMENTED(rt.channel_order != 0x00000001);
@@ -551,12 +595,23 @@ 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);
// 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);
// EXIT_NOT_IMPLEMENTED(rt.dcc.color_transform != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.dcc.enable_overwrite_combiner != false);
EXIT_NOT_IMPLEMENTED(rt.dcc.overwrite_combiner_disable != false);
// EXIT_NOT_IMPLEMENTED(rt.dcc.force_independent_blocks != false);
// EXIT_NOT_IMPLEMENTED(rt.dcc.independent_128b_blocks != false);
// EXIT_NOT_IMPLEMENTED(rt.dcc.data_write_on_dcc_clear_to_reg != false);
EXIT_NOT_IMPLEMENTED(rt.dcc.dcc_clear_key_enable != false);
EXIT_NOT_IMPLEMENTED(rt.cmask.addr != 0x0000000000000000);
EXIT_NOT_IMPLEMENTED(rt.cmask_slice.slice_minus1 != 0x00000000);
EXIT_NOT_IMPLEMENTED(rt.fmask.addr != 0x0000000000000000);
@@ -579,11 +634,17 @@ static void z_print(const char* func, const HW::DepthRenderTarget& z)
printf("\t z_info.tile_surface_enable = %s\n", z.z_info.tile_surface_enable ? "true" : "false");
printf("\t z_info.expclear_enabled = %s\n", z.z_info.expclear_enabled ? "true" : "false");
printf("\t z_info.zrange_precision = 0x%08" PRIx32 "\n", z.z_info.zrange_precision);
printf("\t z_info.embedded_sample_locations = %s\n", z.z_info.embedded_sample_locations ? "true" : "false");
printf("\t z_info.partially_resident = %s\n", z.z_info.partially_resident ? "true" : "false");
printf("\t z_info.num_mip_levels = 0x%02" PRIx8 "\n", z.z_info.num_mip_levels);
printf("\t z_info.plane_compression = 0x%02" PRIx8 "\n", z.z_info.plane_compression);
printf("\t stencil_info.format = 0x%08" PRIx32 "\n", z.stencil_info.format);
printf("\t stencil_info.tile_stencil_disable = %s\n", z.stencil_info.tile_stencil_disable ? "true" : "false");
printf("\t stencil_info.expclear_enabled = %s\n", z.stencil_info.expclear_enabled ? "true" : "false");
printf("\t stencil_info.tile_mode_index = 0x%08" PRIx32 "\n", z.stencil_info.tile_mode_index);
printf("\t stencil_info.tile_split = 0x%08" PRIx32 "\n", z.stencil_info.tile_split);
printf("\t stencil_info.texture_compatible_stencil = %s\n", z.stencil_info.texture_compatible_stencil ? "true" : "false");
printf("\t stencil_info.partially_resident = %s\n", z.stencil_info.partially_resident ? "true" : "false");
printf("\t depth_info.addr5_swizzle_mask = 0x%08" PRIx32 "\n", z.depth_info.addr5_swizzle_mask);
printf("\t depth_info.array_mode = 0x%08" PRIx32 "\n", z.depth_info.array_mode);
printf("\t depth_info.pipe_config = 0x%08" PRIx32 "\n", z.depth_info.pipe_config);
@@ -593,6 +654,9 @@ static void z_print(const char* func, const HW::DepthRenderTarget& z)
printf("\t depth_info.num_banks = 0x%08" PRIx32 "\n", z.depth_info.num_banks);
printf("\t depth_view.slice_start = 0x%08" PRIx32 "\n", z.depth_view.slice_start);
printf("\t depth_view.slice_max = 0x%08" PRIx32 "\n", z.depth_view.slice_max);
printf("\t depth_view.current_mip_level = 0x%02" PRIx8 "\n", z.depth_view.current_mip_level);
printf("\t depth_view.depth_write_disable = %s\n", z.depth_view.depth_write_disable ? "true" : "false");
printf("\t depth_view.stencil_write_disable = %s\n", z.depth_view.stencil_write_disable ? "true" : "false");
printf("\t htile_surface.linear = 0x%08" PRIx32 "\n", z.htile_surface.linear);
printf("\t htile_surface.full_cache = 0x%08" PRIx32 "\n", z.htile_surface.full_cache);
printf("\t htile_surface.htile_uses_preload_win = 0x%08" PRIx32 "\n", z.htile_surface.htile_uses_preload_win);
@@ -610,6 +674,8 @@ static void z_print(const char* func, const HW::DepthRenderTarget& z)
printf("\t htile_data_base_addr = 0x%016" PRIx64 "\n", z.htile_data_base_addr);
printf("\t width = 0x%08" PRIx32 "\n", z.width);
printf("\t height = 0x%08" PRIx32 "\n", z.height);
printf("\t size.x_max = 0x%04" PRIx16 "\n", z.size.x_max);
printf("\t size.y_max = 0x%04" PRIx16 "\n", z.size.y_max);
}
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
@@ -623,6 +689,10 @@ static void z_check(const HW::DepthRenderTarget& z)
EXIT_NOT_IMPLEMENTED(z.z_info.tile_surface_enable != false);
EXIT_NOT_IMPLEMENTED(z.z_info.expclear_enabled != false);
EXIT_NOT_IMPLEMENTED(z.z_info.zrange_precision != 0);
EXIT_NOT_IMPLEMENTED(z.z_info.embedded_sample_locations != false);
EXIT_NOT_IMPLEMENTED(z.z_info.partially_resident != false);
EXIT_NOT_IMPLEMENTED(z.z_info.num_mip_levels != 0);
EXIT_NOT_IMPLEMENTED(z.z_info.plane_compression != 0);
} else
{
EXIT_NOT_IMPLEMENTED(z.z_info.format != 0x00000003);
@@ -631,22 +701,30 @@ static void z_check(const HW::DepthRenderTarget& z)
// EXIT_NOT_IMPLEMENTED(z.z_info.tile_surface_enable != true);
EXIT_NOT_IMPLEMENTED(z.z_info.expclear_enabled != false);
EXIT_NOT_IMPLEMENTED(z.z_info.zrange_precision != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.z_info.embedded_sample_locations != false);
EXIT_NOT_IMPLEMENTED(z.z_info.partially_resident != false);
EXIT_NOT_IMPLEMENTED(z.z_info.num_mip_levels != 0);
EXIT_NOT_IMPLEMENTED(z.z_info.plane_compression != 0);
}
if (z.stencil_info.format == 0)
{
EXIT_NOT_IMPLEMENTED(z.stencil_info.format != 0);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_stencil_disable != false);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.format != 0);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_stencil_disable != false);
EXIT_NOT_IMPLEMENTED(z.stencil_info.expclear_enabled != false);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_mode_index != 0);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_split != 0);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.texture_compatible_stencil != true);
EXIT_NOT_IMPLEMENTED(z.stencil_info.partially_resident != false);
} else
{
EXIT_NOT_IMPLEMENTED(z.stencil_info.format != 0x00000001);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.format != 0x00000001);
EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_stencil_disable != true);
EXIT_NOT_IMPLEMENTED(z.stencil_info.expclear_enabled != false);
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_mode_index != (Config::IsNeo() ? 0x00000002 : 0));
// EXIT_NOT_IMPLEMENTED(z.stencil_info.tile_split != (Config::IsNeo() ? 0x00000002 : 0));
// EXIT_NOT_IMPLEMENTED(z.stencil_info.texture_compatible_stencil != true);
EXIT_NOT_IMPLEMENTED(z.stencil_info.partially_resident != false);
}
if (z.z_info.format != 0 || z.stencil_info.format != 0)
@@ -660,6 +738,9 @@ static void z_check(const HW::DepthRenderTarget& z)
// EXIT_NOT_IMPLEMENTED(z.depth_info.num_banks != (Config::IsNeo() ? 0x00000002 : 3));
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);
@@ -677,6 +758,8 @@ 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);
}
}
@@ -820,6 +903,8 @@ static void d_print(const char* func, const HW::DepthControl& c, const HW::Stenc
printf("\t backface_enable = %s\n", c.backface_enable ? "true" : "false");
printf("\t stencilfunc = %" PRIu8 "\n", c.stencilfunc);
printf("\t stencilfunc_bf = %" PRIu8 "\n", c.stencilfunc_bf);
printf("\t color_writes_on_depth_fail_enable = %s\n", c.color_writes_on_depth_fail_enable ? "true" : "false");
printf("\t color_writes_on_depth_pass_disable = %s\n", c.color_writes_on_depth_pass_disable ? "true" : "false");
printf("\t stencil_fail = %" PRIu8 "\n", s.stencil_fail);
printf("\t stencil_zpass = %" PRIu8 "\n", s.stencil_zpass);
printf("\t stencil_zfail = %" PRIu8 "\n", s.stencil_zfail);
@@ -846,6 +931,8 @@ static void d_check(const HW::DepthControl& c, const HW::StencilControl& s, cons
EXIT_NOT_IMPLEMENTED(c.backface_enable != false);
// EXIT_NOT_IMPLEMENTED(c.stencilfunc != 0);
// EXIT_NOT_IMPLEMENTED(c.stencilfunc_bf != 0);
EXIT_NOT_IMPLEMENTED(c.color_writes_on_depth_fail_enable != false);
EXIT_NOT_IMPLEMENTED(c.color_writes_on_depth_pass_disable != false);
// EXIT_NOT_IMPLEMENTED(s.stencil_fail != 0);
// EXIT_NOT_IMPLEMENTED(s.stencil_zpass != 0);
// EXIT_NOT_IMPLEMENTED(s.stencil_zfail != 0);
@@ -3736,7 +3823,9 @@ static void FindRenderDepthInfo(uint64_t submit_id, CommandBuffer* /*buffer*/, c
if (r->format != VK_FORMAT_UNDEFINED)
{
DepthStencilBufferObject vulkan_buffer_info(r->format, r->width, r->height, htile, neo);
bool sampled = ((z.stencil_info.format == 0 && z.z_info.tile_mode_index != 0) || z.stencil_info.texture_compatible_stencil);
DepthStencilBufferObject vulkan_buffer_info(r->format, r->width, r->height, htile, neo, sampled);
EXIT_NOT_IMPLEMENTED(z.z_info.tile_mode_index != 0 && r->depth_tile_swizzle != 0);
EXIT_NOT_IMPLEMENTED(r->stencil_tile_swizzle != 0);
@@ -4321,6 +4410,26 @@ static void SetDynamicParams(VkCommandBuffer vk_buffer, VulkanPipeline* pipeline
}
}
static bool shader_is_disabled(HW::Shader* sh_ctx)
{
if (const auto& vs = sh_ctx->GetVs();
!vs.vs_embedded && ((vs.vs_regs.data_addr != 0 && ShaderIsDisabled(vs.vs_regs.data_addr)) ||
(vs.vs_regs.data_addr == 0 && vs.gs_regs.data_addr == 0 && vs.es_regs.data_addr != 0 &&
vs.gs_regs.chksum != 0 && ShaderIsDisabled2(vs.es_regs.data_addr, vs.gs_regs.chksum))))
{
return true;
}
if (const auto& ps = sh_ctx->GetPs();
!ps.ps_embedded && ((ps.ps_regs.chksum == 0 && ShaderIsDisabled(ps.ps_regs.data_addr)) ||
(ps.ps_regs.chksum != 0 && ShaderIsDisabled2(ps.ps_regs.data_addr, ps.ps_regs.chksum))))
{
return true;
}
return false;
}
void GraphicsRenderDrawIndex(uint64_t submit_id, CommandBuffer* buffer, HW::Context* ctx, HW::UserConfig* ucfg, HW::Shader* sh_ctx,
uint32_t index_type_and_size, uint32_t index_count, const void* index_addr, uint32_t flags, uint32_t type)
{
@@ -4334,12 +4443,7 @@ void GraphicsRenderDrawIndex(uint64_t submit_id, CommandBuffer* buffer, HW::Cont
Core::LockGuard lock(g_render_ctx->GetMutex());
if (const auto& vs = sh_ctx->GetVs(); !vs.vs_embedded && ShaderIsDisabled(vs.vs_regs.GetGpuAddress()))
{
return;
}
if (const auto& ps = sh_ctx->GetPs(); !ps.ps_embedded && ShaderIsDisabled(ps.ps_regs.data_addr))
if (shader_is_disabled(sh_ctx))
{
return;
}
@@ -4482,12 +4586,7 @@ void GraphicsRenderDrawIndexAuto(uint64_t submit_id, CommandBuffer* buffer, HW::
Core::LockGuard lock(g_render_ctx->GetMutex());
if (const auto& vs = sh_ctx->GetVs(); !vs.vs_embedded && ShaderIsDisabled(vs.vs_regs.GetGpuAddress()))
{
return;
}
if (const auto& ps = sh_ctx->GetPs(); !ps.ps_embedded && ShaderIsDisabled(ps.ps_regs.data_addr))
if (shader_is_disabled(sh_ctx))
{
return;
}
source/emulator/src/Graphics/GraphicsRun.cpp
+1212 -204
View File
File diff suppressed because it is too large Load Diff
source/emulator/src/Graphics/Objects/DepthStencilBuffer.cpp
+6 -10
View File
@@ -31,7 +31,8 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
auto pixel_format = static_cast<VkFormat>(params[DepthStencilBufferObject::PARAM_FORMAT]);
auto width = params[DepthStencilBufferObject::PARAM_WIDTH];
auto height = params[DepthStencilBufferObject::PARAM_HEIGHT];
bool htile = params[DepthStencilBufferObject::PARAM_HTILE] != 1;
bool htile = params[DepthStencilBufferObject::PARAM_HTILE] != 0;
bool sampled = params[DepthStencilBufferObject::PARAM_USAGE] == 1;
EXIT_NOT_IMPLEMENTED(pixel_format == VK_FORMAT_UNDEFINED);
EXIT_NOT_IMPLEMENTED(width == 0);
@@ -50,7 +51,7 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
view = nullptr;
}
vk_obj->compressed = htile;
vk_obj->compressed = !htile;
VkImageCreateInfo image_info {};
image_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
@@ -66,7 +67,7 @@ static void* create_func(GraphicContext* ctx, const uint64_t* params, const uint
image_info.tiling = VK_IMAGE_TILING_OPTIMAL;
image_info.initialLayout = vk_obj->layout;
image_info.usage = static_cast<VkImageUsageFlags>(VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) |
static_cast<VkImageUsageFlags>(VK_IMAGE_USAGE_SAMPLED_BIT);
(sampled ? static_cast<VkImageUsageFlags>(VK_IMAGE_USAGE_SAMPLED_BIT) : static_cast<VkImageUsageFlags>(0));
image_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
image_info.samples = VK_SAMPLE_COUNT_1_BIT;
@@ -160,15 +161,10 @@ static void delete_func(GraphicContext* ctx, void* obj, VulkanMemory* mem)
bool DepthStencilBufferObject::Equal(const uint64_t* other) const
{
return (params[PARAM_FORMAT] == other[PARAM_FORMAT] && params[PARAM_WIDTH] == other[PARAM_WIDTH] &&
params[PARAM_HEIGHT] == other[PARAM_HEIGHT] && params[PARAM_HTILE] == other[PARAM_HTILE]);
params[PARAM_HEIGHT] == other[PARAM_HEIGHT] && params[PARAM_HTILE] == other[PARAM_HTILE] &&
params[PARAM_NEO] == other[PARAM_NEO] && params[PARAM_USAGE] == other[PARAM_USAGE]);
}
// bool DepthStencilBufferObject::Reuse(const uint64_t* other) const
//{
// return (params[PARAM_FORMAT] == other[PARAM_FORMAT] && params[PARAM_WIDTH] <= other[PARAM_WIDTH] &&
// params[PARAM_HEIGHT] <= other[PARAM_HEIGHT] && params[PARAM_HTILE] == other[PARAM_HTILE]);
//}
GpuObject::create_func_t DepthStencilBufferObject::GetCreateFunc() const
{
return create_func;
source/emulator/src/Graphics/Shader.cpp
+132 -32
View File
@@ -2088,20 +2088,59 @@ KYTY_SHADER_PARSER(shader_parse)
return ptr - src;
}
static void vs_print(const char* func, const HW::VsStageRegisters& vs, const HW::ShaderRegisters& sh)
static void vs_print(const char* func, const HW::VertexShaderInfo& vs, const HW::ShaderRegisters& sh)
{
printf("%s\n", func);
printf("\t GetGpuAddress() = 0x%016" PRIx64 "\n", vs.GetGpuAddress());
printf("\t GetStreamoutEnabled() = %s\n", vs.GetStreamoutEnabled() ? "true" : "false");
printf("\t GetSgprCount() = 0x%08" PRIx32 "\n", vs.GetSgprCount());
printf("\t GetInputComponentsCount() = 0x%08" PRIx32 "\n", vs.GetInputComponentsCount());
printf("\t GetUnknown1() = 0x%08" PRIx32 "\n", vs.GetUnknown1());
printf("\t GetUnknown2() = 0x%08" PRIx32 "\n", vs.GetUnknown2());
printf("\t vs.data_addr = 0x%016" PRIx64 "\n", vs.vs_regs.data_addr);
printf("\t es.data_addr = 0x%016" PRIx64 "\n", vs.es_regs.data_addr);
printf("\t gs.data_addr = 0x%016" PRIx64 "\n", vs.gs_regs.data_addr);
if (vs.vs_regs.data_addr != 0)
{
printf("\t vs.vgprs = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc1.vgprs);
printf("\t vs.sgprs = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc1.sgprs);
printf("\t vs.priority = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc1.priority);
printf("\t vs.float_mode = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc1.float_mode);
printf("\t vs.dx10_clamp = %s\n", vs.vs_regs.rsrc1.dx10_clamp ? "true" : "false");
printf("\t vs.ieee_mode = %s\n", vs.vs_regs.rsrc1.ieee_mode ? "true" : "false");
printf("\t vs.vgpr_component_count = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc1.vgpr_component_count);
printf("\t vs.cu_group_enable = %s\n", vs.vs_regs.rsrc1.cu_group_enable ? "true" : "false");
printf("\t vs.require_forward_progress = %s\n", vs.vs_regs.rsrc1.require_forward_progress ? "true" : "false");
printf("\t vs.fp16_overflow = %s\n", vs.vs_regs.rsrc1.fp16_overflow ? "true" : "false");
printf("\t vs.scratch_en = %s\n", vs.vs_regs.rsrc2.scratch_en ? "true" : "false");
printf("\t vs.user_sgpr = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc2.user_sgpr);
printf("\t vs.offchip_lds = %s\n", vs.vs_regs.rsrc2.offchip_lds ? "true" : "false");
printf("\t vs.streamout_enabled = %s\n", vs.vs_regs.rsrc2.streamout_enabled ? "true" : "false");
printf("\t vs.shared_vgprs = 0x%02" PRIx8 "\n", vs.vs_regs.rsrc2.shared_vgprs);
}
if (vs.gs_regs.data_addr != 0 || vs.es_regs.data_addr != 0)
{
printf("\t chksum = 0x%016" PRIx64 "\n", vs.gs_regs.chksum);
printf("\t gs.vgprs = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc1.vgprs);
printf("\t gs.sgprs = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc1.sgprs);
printf("\t gs.priority = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc1.priority);
printf("\t gs.float_mode = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc1.float_mode);
printf("\t gs.dx10_clamp = %s\n", vs.gs_regs.rsrc1.dx10_clamp ? "true" : "false");
printf("\t gs.ieee_mode = %s\n", vs.gs_regs.rsrc1.ieee_mode ? "true" : "false");
printf("\t gs.debug_mode = %s\n", vs.gs_regs.rsrc1.debug_mode ? "true" : "false");
printf("\t gs.lds_configuration = %s\n", vs.gs_regs.rsrc1.lds_configuration ? "true" : "false");
printf("\t gs.cu_group_enable = %s\n", vs.gs_regs.rsrc1.cu_group_enable ? "true" : "false");
printf("\t gs.require_forward_progress = %s\n", vs.gs_regs.rsrc1.require_forward_progress ? "true" : "false");
printf("\t gs.fp16_overflow = %s\n", vs.gs_regs.rsrc1.fp16_overflow ? "true" : "false");
printf("\t gs.gs_vgpr_component_count = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc1.gs_vgpr_component_count);
printf("\t gs.scratch_en = %s\n", vs.gs_regs.rsrc2.scratch_en ? "true" : "false");
printf("\t gs.user_sgpr = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc2.user_sgpr);
printf("\t gs.offchip_lds = %s\n", vs.gs_regs.rsrc2.offchip_lds ? "true" : "false");
printf("\t gs.shared_vgprs = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc2.shared_vgprs);
printf("\t gs.es_vgpr_component_count = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc2.es_vgpr_component_count);
printf("\t gs.lds_size = 0x%02" PRIx8 "\n", vs.gs_regs.rsrc2.lds_size);
}
printf("\t m_spiVsOutConfig = 0x%08" PRIx32 "\n", sh.m_spiVsOutConfig);
printf("\t m_spiShaderPosFormat = 0x%08" PRIx32 "\n", sh.m_spiShaderPosFormat);
printf("\t m_paClVsOutCntl = 0x%08" PRIx32 "\n", sh.m_paClVsOutCntl);
printf("\t m_spiShaderIdxFormat = 0x%08" PRIx32 "\n", sh.m_spiShaderIdxFormat);
printf("\t m_geNggSubgrpCntl = 0x%08" PRIx32 "\n", sh.m_geNggSubgrpCntl);
printf("\t m_vgtGsInstanceCnt = 0x%08" PRIx32 "\n", sh.m_vgtGsInstanceCnt);
@@ -2118,21 +2157,30 @@ static void ps_print(const char* func, const HW::PsStageRegisters& ps, const HW:
{
printf("%s\n", func);
// printf("\t GetGpuAddress() = 0x%016" PRIx64 "\n", ps.GetGpuAddress());
// printf("\t m_spiShaderPgmRsrc1Ps = 0x%08" PRIx32 "\n", ps.m_spiShaderPgmRsrc1Ps);
// printf("\t m_spiShaderPgmRsrc2Ps = 0x%08" PRIx32 "\n", ps.m_spiShaderPgmRsrc2Ps);
// printf("\t GetTargetOutputMode(0) = 0x%08" PRIx32 "\n", ps.GetTargetOutputMode(0));
printf("\t data_addr = 0x%016" PRIx64 "\n", ps.data_addr);
printf("\t chksum = 0x%016" PRIx64 "\n", ps.chksum);
printf("\t conservative_z_export_value = 0x%08" PRIx32 "\n", sh.db_shader_control.conservative_z_export_value);
printf("\t shader_z_behavior = 0x%08" PRIx32 "\n", sh.db_shader_control.shader_z_behavior);
printf("\t shader_kill_enable = %s\n", sh.db_shader_control.shader_kill_enable ? "true" : "false");
printf("\t shader_z_export_enable = %s\n", sh.db_shader_control.shader_z_export_enable ? "true" : "false");
printf("\t shader_execute_on_noop = %s\n", sh.db_shader_control.shader_execute_on_noop ? "true" : "false");
printf("\t vgprs = 0x%02" PRIx8 "\n", ps.vgprs);
printf("\t sgprs = 0x%02" PRIx8 "\n", ps.sgprs);
printf("\t scratch_en = 0x%02" PRIx8 "\n", ps.scratch_en);
printf("\t user_sgpr = 0x%02" PRIx8 "\n", ps.user_sgpr);
printf("\t wave_cnt_en = 0x%02" PRIx8 "\n", ps.wave_cnt_en);
printf("\t vgprs = 0x%02" PRIx8 "\n", ps.rsrc1.vgprs);
printf("\t sgprs = 0x%02" PRIx8 "\n", ps.rsrc1.sgprs);
printf("\t priority = 0x%02" PRIx8 "\n", ps.rsrc1.priority);
printf("\t float_mode = 0x%02" PRIx8 "\n", ps.rsrc1.float_mode);
printf("\t dx10_clamp = %s\n", ps.rsrc1.dx10_clamp ? "true" : "false");
printf("\t debug_mode = %s\n", ps.rsrc1.debug_mode ? "true" : "false");
printf("\t ieee_mode = %s\n", ps.rsrc1.ieee_mode ? "true" : "false");
printf("\t cu_group_disable = %s\n", ps.rsrc1.cu_group_disable ? "true" : "false");
printf("\t require_forward_progress = %s\n", ps.rsrc1.require_forward_progress ? "true" : "false");
printf("\t fp16_overflow = %s\n", ps.rsrc1.fp16_overflow ? "true" : "false");
printf("\t scratch_en = %s\n", ps.rsrc2.scratch_en ? "true" : "false");
printf("\t user_sgpr = 0x%02" PRIx8 "\n", ps.rsrc2.user_sgpr);
printf("\t wave_cnt_en = %s\n", ps.rsrc2.wave_cnt_en ? "true" : "false");
printf("\t extra_lds_size = 0x%02" PRIx8 "\n", ps.rsrc2.extra_lds_size);
printf("\t raster_ordered_shading = %s\n", ps.rsrc2.raster_ordered_shading ? "true" : "false");
printf("\t shared_vgprs = 0x%02" PRIx8 "\n", ps.rsrc2.shared_vgprs);
printf("\t shader_z_format = 0x%08" PRIx32 "\n", sh.shader_z_format);
printf("\t target_output_mode[0] = 0x%02" PRIx8 "\n", sh.target_output_mode[0]);
printf("\t ps_input_ena = 0x%08" PRIx32 "\n", sh.ps_input_ena);
@@ -2196,14 +2244,43 @@ static void bi_print(const char* func, const ShaderBinaryInfo& bi)
printf("\t crc32 = 0x%08" PRIx32 "\n", bi.crc32);
}
static void vs_check(const HW::VsStageRegisters& vs, const HW::ShaderRegisters& sh)
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
static void vs_check(const HW::VertexShaderInfo& vs, const HW::ShaderRegisters& sh)
{
EXIT_NOT_IMPLEMENTED(vs.GetStreamoutEnabled() != false);
// EXIT_NOT_IMPLEMENTED(vs.GetSgprCount() != 0x00000000);
// EXIT_NOT_IMPLEMENTED(vs.GetInputComponentsCount() != 0x00000003);
// EXIT_NOT_IMPLEMENTED(vs.GetUnknown1() != 0x002c0000);
// EXIT_NOT_IMPLEMENTED(vs.GetUnknown2() != 0x00000000);
// EXIT_NOT_IMPLEMENTED(vs.m_spiVsOutConfig != 0x00000000);
if (vs.vs_regs.data_addr != 0)
{
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.priority != 0);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.float_mode != 192);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.dx10_clamp != true);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.ieee_mode != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.cu_group_enable != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.require_forward_progress != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc1.fp16_overflow != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc2.scratch_en != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc2.offchip_lds != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc2.streamout_enabled != false);
EXIT_NOT_IMPLEMENTED(vs.vs_regs.rsrc2.shared_vgprs != 0);
}
if (vs.es_regs.data_addr != 0 || vs.gs_regs.data_addr != 0)
{
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.priority != 0);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.float_mode != 192);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.dx10_clamp != true);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.debug_mode != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.ieee_mode != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.cu_group_enable != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.require_forward_progress != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.lds_configuration != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.gs_vgpr_component_count != 0);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc1.fp16_overflow != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc2.scratch_en != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc2.offchip_lds != false);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc2.es_vgpr_component_count != 0);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc2.lds_size != 0);
EXIT_NOT_IMPLEMENTED(vs.gs_regs.rsrc2.shared_vgprs != 0);
}
EXIT_NOT_IMPLEMENTED(sh.m_spiShaderPosFormat != 0x00000004);
EXIT_NOT_IMPLEMENTED(sh.m_paClVsOutCntl != 0x00000000);
@@ -2231,9 +2308,21 @@ static void ps_check(const HW::PsStageRegisters& ps, const HW::ShaderRegisters&
// EXIT_NOT_IMPLEMENTED(ps.m_spiShaderPgmRsrc2Ps != 0x00000000);
// EXIT_NOT_IMPLEMENTED(ps.vgprs != 0x00 && ps.vgprs != 0x01);
// EXIT_NOT_IMPLEMENTED(ps.sgprs != 0x00 && ps.sgprs != 0x01);
EXIT_NOT_IMPLEMENTED(ps.scratch_en != 0);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.priority != 0);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.float_mode != 192);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.dx10_clamp != true);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.debug_mode != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.ieee_mode != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.cu_group_disable != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.require_forward_progress != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc1.fp16_overflow != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc2.scratch_en != false);
// EXIT_NOT_IMPLEMENTED(ps.user_sgpr != 0 && ps.user_sgpr != 4 && ps.user_sgpr != 12);
EXIT_NOT_IMPLEMENTED(ps.wave_cnt_en != 0);
EXIT_NOT_IMPLEMENTED(ps.rsrc2.wave_cnt_en != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc2.extra_lds_size != 0);
EXIT_NOT_IMPLEMENTED(ps.rsrc2.raster_ordered_shading != false);
EXIT_NOT_IMPLEMENTED(ps.rsrc2.shared_vgprs != 0);
EXIT_NOT_IMPLEMENTED(sh.shader_z_format != 0x00000000);
EXIT_NOT_IMPLEMENTED(sh.ps_input_ena != 0x00000002 && sh.ps_input_ena != 0x00000302);
EXIT_NOT_IMPLEMENTED(sh.ps_input_addr != 0x00000002 && sh.ps_input_addr != 0x00000302);
@@ -2756,7 +2845,7 @@ void ShaderGetInputInfoVS(const HW::VertexShaderInfo* regs, const HW::ShaderRegi
return;
}
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.GetGpuAddress());
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.data_addr);
auto usages = GetUsageSlots(src);
@@ -3400,12 +3489,14 @@ ShaderCode ShaderParseVS(const HW::VertexShaderInfo* regs, const HW::ShaderRegis
code.SetVsEmbeddedId(regs->vs_embedded_id);
} else
{
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.GetGpuAddress());
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.data_addr);
EXIT_NOT_IMPLEMENTED(src == nullptr);
vs_print("ShaderParseVS()", regs->vs_regs, *sh);
vs_check(regs->vs_regs, *sh);
vs_print("ShaderParseVS()", *regs, *sh);
vs_check(*regs, *sh);
EXIT_NOT_IMPLEMENTED(regs->vs_regs.rsrc2.user_sgpr > regs->vs_user_sgpr.count);
const auto* header = GetBinaryInfo(src);
@@ -3489,7 +3580,7 @@ ShaderCode ShaderParsePS(const HW::PixelShaderInfo* regs, const HW::ShaderRegist
ps_print("ShaderParsePS()", regs->ps_regs, *sh);
ps_check(regs->ps_regs, *sh);
EXIT_NOT_IMPLEMENTED(regs->ps_regs.user_sgpr > regs->ps_user_sgpr.count);
EXIT_NOT_IMPLEMENTED(regs->ps_regs.rsrc2.user_sgpr > regs->ps_user_sgpr.count);
const auto* header = GetBinaryInfo(src);
@@ -3853,7 +3944,7 @@ ShaderId ShaderGetIdVS(const HW::VertexShaderInfo* regs, const ShaderVertexInput
return ret;
}
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.GetGpuAddress());
const auto* src = reinterpret_cast<const uint32_t*>(regs->vs_regs.data_addr);
EXIT_NOT_IMPLEMENTED(src == nullptr);
@@ -3999,6 +4090,15 @@ bool ShaderIsDisabled(uint64_t addr)
return disabled;
}
bool ShaderIsDisabled2(uint64_t addr, uint64_t chksum)
{
bool disabled = (g_disabled_shaders != nullptr && g_disabled_shaders->Contains(chksum));
printf("Shader 0x%016" PRIx64 ": id = 0x%016" PRIx64 " - %s\n", addr, chksum, (disabled ? "disabled" : "enabled"));
return disabled;
}
void ShaderDisable(uint64_t id)
{
if (g_disabled_shaders == nullptr)