ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #10290 from OatmealDome/m1-earlyz-bug 

DriverDetails: Add broken discard with early-Z bug on Apple Silicon GPUs
This commit is contained in:
JMC47 2022-04-24 13:30:04 -04:00 committed by GitHub
parent 61edcf7e4f 259a5fc7c0
commit c42392c565
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 144 additions and 65 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

82
Source/Core/VideoBackends/Vulkan/VKPipeline.cpp
View File

@ -132,53 +132,65 @@ static VkPipelineDepthStencilStateCreateInfo GetVulkanDepthStencilState(const De
}; };
} }
static VkPipelineColorBlendAttachmentState GetVulkanAttachmentBlendState(const BlendingState& state) static VkPipelineColorBlendAttachmentState
GetVulkanAttachmentBlendState(const BlendingState& state, AbstractPipelineUsage usage)
{ {
VkPipelineColorBlendAttachmentState vk_state = {}; VkPipelineColorBlendAttachmentState vk_state = {};
vk_state.blendEnable = static_cast<VkBool32>(state.blendenable);
vk_state.colorBlendOp = state.subtract ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
vk_state.alphaBlendOp = state.subtractAlpha ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
bool use_dual_source = bool use_dual_source =
state.usedualsrc && g_ActiveConfig.backend_info.bSupportsDualSourceBlend && state.usedualsrc && g_ActiveConfig.backend_info.bSupportsDualSourceBlend &&
(!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || state.dstalpha); (!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || state.dstalpha);
bool use_shader_blend = !use_dual_source && state.usedualsrc && state.dstalpha &&
g_ActiveConfig.backend_info.bSupportsFramebufferFetch;
if (use_dual_source) if (use_shader_blend || (usage == AbstractPipelineUsage::GX &&
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z)))
{ {
static constexpr std::array<VkBlendFactor, 8> src_factors = { vk_state.blendEnable = VK_FALSE;
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
static constexpr std::array<VkBlendFactor, 8> dst_factors = {
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
} }
else else
{ {
static constexpr std::array<VkBlendFactor, 8> src_factors = { vk_state.blendEnable = static_cast<VkBool32>(state.blendenable);
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR, vk_state.colorBlendOp = state.subtract ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC_ALPHA, vk_state.alphaBlendOp = state.subtractAlpha ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
static constexpr std::array<VkBlendFactor, 8> dst_factors = { if (use_dual_source)
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR, {
VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC_ALPHA, static constexpr std::array<VkBlendFactor, 8> src_factors = {
VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA, {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}}; VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
static constexpr std::array<VkBlendFactor, 8> dst_factors = {
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())]; vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())]; vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())]; vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())]; vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
}
else
{
static constexpr std::array<VkBlendFactor, 8> src_factors = {
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
static constexpr std::array<VkBlendFactor, 8> dst_factors = {
{VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
}
} }
if (state.colorupdate) if (state.colorupdate)
@ -338,7 +350,7 @@ std::unique_ptr<VKPipeline> VKPipeline::Create(const AbstractPipelineConfig& con
VkPipelineDepthStencilStateCreateInfo depth_stencil_state = VkPipelineDepthStencilStateCreateInfo depth_stencil_state =
GetVulkanDepthStencilState(config.depth_state); GetVulkanDepthStencilState(config.depth_state);
VkPipelineColorBlendAttachmentState blend_attachment_state = VkPipelineColorBlendAttachmentState blend_attachment_state =
GetVulkanAttachmentBlendState(config.blending_state); GetVulkanAttachmentBlendState(config.blending_state, config.usage);
VkPipelineColorBlendStateCreateInfo blend_state = VkPipelineColorBlendStateCreateInfo blend_state =
GetVulkanColorBlendState(config.blending_state, &blend_attachment_state, 1); GetVulkanColorBlendState(config.blending_state, &blend_attachment_state, 1);

12
Source/Core/VideoBackends/Vulkan/VulkanContext.cpp
View File

@ -371,6 +371,13 @@ void VulkanContext::PopulateBackendInfoFeatures(VideoConfig* config, VkPhysicalD
// with depth clamping. Fall back to inverted depth range for these. // with depth clamping. Fall back to inverted depth range for these.
if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_REVERSED_DEPTH_RANGE)) if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_REVERSED_DEPTH_RANGE))
config->backend_info.bSupportsReversedDepthRange = false; config->backend_info.bSupportsReversedDepthRange = false;
// Calling discard when early depth test is enabled can break on some Apple Silicon GPU drivers.
if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z))
{
// We will use shader blending, so disable hardware dual source blending.
config->backend_info.bSupportsDualSourceBlend = false;
}
} }
void VulkanContext::PopulateBackendInfoMultisampleModes( void VulkanContext::PopulateBackendInfoMultisampleModes(
@ -922,6 +929,11 @@ void VulkanContext::InitDriverDetails()
vendor = DriverDetails::VENDOR_IMGTEC; vendor = DriverDetails::VENDOR_IMGTEC;
driver = DriverDetails::DRIVER_IMGTEC; driver = DriverDetails::DRIVER_IMGTEC;
} }
else if (device_name.find("Apple") != std::string::npos)
{
vendor = DriverDetails::VENDOR_APPLE;
driver = DriverDetails::DRIVER_PORTABILITY;
}
else else
{ {
WARN_LOG_FMT(VIDEO, "Unknown Vulkan driver vendor, please report it to us."); WARN_LOG_FMT(VIDEO, "Unknown Vulkan driver vendor, please report it to us.");

2
Source/Core/VideoCommon/DriverDetails.cpp
View File

@ -140,6 +140,8 @@ constexpr BugInfo m_known_bugs[] = {
-1.0, -1.0, true}, -1.0, -1.0, true},
{API_VULKAN, OS_ALL, VENDOR_QUALCOMM, DRIVER_QUALCOMM, Family::UNKNOWN, BUG_PRIMITIVE_RESTART, {API_VULKAN, OS_ALL, VENDOR_QUALCOMM, DRIVER_QUALCOMM, Family::UNKNOWN, BUG_PRIMITIVE_RESTART,
-1.0, -1.0, true}, -1.0, -1.0, true},
{API_VULKAN, OS_OSX, VENDOR_APPLE, DRIVER_PORTABILITY, Family::UNKNOWN,
BUG_BROKEN_DISCARD_WITH_EARLY_Z, -1.0, -1.0, true},
}; };
static std::map<Bug, BugInfo> m_bugs; static std::map<Bug, BugInfo> m_bugs;

11
Source/Core/VideoCommon/DriverDetails.h
View File

@ -43,6 +43,7 @@ enum Vendor
VENDOR_TEGRA, VENDOR_TEGRA,
VENDOR_VIVANTE, VENDOR_VIVANTE,
VENDOR_MESA, VENDOR_MESA,
VENDOR_APPLE,
VENDOR_UNKNOWN VENDOR_UNKNOWN
}; };
@ -318,7 +319,15 @@ enum Bug
// BUG: Multi-threaded shader pre-compilation sometimes crashes // BUG: Multi-threaded shader pre-compilation sometimes crashes
// Used primarily in Videoconfig.cpp's GetNumAutoShaderPreCompilerThreads() // Used primarily in Videoconfig.cpp's GetNumAutoShaderPreCompilerThreads()
// refer to https://github.com/dolphin-emu/dolphin/pull/9414 for initial validation coverage // refer to https://github.com/dolphin-emu/dolphin/pull/9414 for initial validation coverage
BUG_BROKEN_MULTITHREADED_SHADER_PRECOMPILATION BUG_BROKEN_MULTITHREADED_SHADER_PRECOMPILATION,
// BUG: Some driver and Apple Silicon GPU combinations have problems with fragment discard when
// early depth test is enabled. Discarded fragments may appear corrupted (Super Mario Sunshine,
// Sonic Adventure 2: Battle, Phantasy Star Online Epsiodes 1 & 2, etc).
// Affected devices: Apple Silicon GPUs of Apple family 4 and newer.
// Started version: -1
// Ended version: -1
BUG_BROKEN_DISCARD_WITH_EARLY_Z,
}; };
// Initializes our internal vendor, device family, and driver version // Initializes our internal vendor, device family, and driver version

51
Source/Core/VideoCommon/PixelShaderGen.cpp
View File

@ -321,7 +321,9 @@ PixelShaderUid GetPixelShaderUid()
BlendingState state = {}; BlendingState state = {};
state.Generate(bpmem); state.Generate(bpmem);
if (state.usedualsrc && state.dstalpha && g_ActiveConfig.backend_info.bSupportsFramebufferFetch && if (((state.usedualsrc && state.dstalpha) ||
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z)) &&
g_ActiveConfig.backend_info.bSupportsFramebufferFetch &&
!g_ActiveConfig.backend_info.bSupportsDualSourceBlend) !g_ActiveConfig.backend_info.bSupportsDualSourceBlend)
{ {
uid_data->blend_enable = state.blendenable; uid_data->blend_enable = state.blendenable;
@ -944,10 +946,15 @@ ShaderCode GeneratePixelShaderCode(APIType api_type, const ShaderHostConfig& hos
(!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || (!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) ||
uid_data->useDstAlpha); uid_data->useDstAlpha);
const bool use_shader_blend = const bool use_shader_blend =
!use_dual_source && uid_data->useDstAlpha && host_config.backend_shader_framebuffer_fetch; !use_dual_source &&
(uid_data->useDstAlpha ||
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z)) &&
host_config.backend_shader_framebuffer_fetch;
const bool use_shader_logic_op = !host_config.backend_logic_op && uid_data->logic_op_enable && const bool use_shader_logic_op = !host_config.backend_logic_op && uid_data->logic_op_enable &&
host_config.backend_shader_framebuffer_fetch; host_config.backend_shader_framebuffer_fetch;
const bool use_framebuffer_fetch = use_shader_blend || use_shader_logic_op; const bool use_framebuffer_fetch =
use_shader_blend || use_shader_logic_op ||
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z);
if (api_type == APIType::OpenGL || api_type == APIType::Vulkan) if (api_type == APIType::OpenGL || api_type == APIType::Vulkan)
{ {
@ -1869,9 +1876,23 @@ static void WriteAlphaTest(ShaderCode& out, const pixel_shader_uid_data* uid_dat
// ZCOMPLOC HACK: // ZCOMPLOC HACK:
if (!uid_data->alpha_test_use_zcomploc_hack) if (!uid_data->alpha_test_use_zcomploc_hack)
{ {
out.Write("\t\tdiscard;\n"); #ifdef __APPLE__
if (api_type == APIType::D3D) if (uid_data->forced_early_z &&
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z))
{
// Instead of using discard, fetch the framebuffer's color value and use it as the output
// for this fragment.
out.Write("\t\t{} = float4(initial_ocol0.xyz, 1.0);\n",
use_dual_source ? "real_ocol0" : "ocol0");
out.Write("\t\treturn;\n"); out.Write("\t\treturn;\n");
}
else
#endif
{
out.Write("\t\tdiscard;\n");
if (api_type == APIType::D3D)
out.Write("\t\treturn;\n");
}
} }
out.Write("\t}}\n"); out.Write("\t}}\n");
@ -2018,8 +2039,8 @@ static void WriteBlend(ShaderCode& out, const pixel_shader_uid_data* uid_data)
"float3(1,1,1);", // ONE "float3(1,1,1);", // ONE
"initial_ocol0.rgb;", // DSTCLR "initial_ocol0.rgb;", // DSTCLR
"float3(1,1,1) - initial_ocol0.rgb;", // INVDSTCLR "float3(1,1,1) - initial_ocol0.rgb;", // INVDSTCLR
"ocol1.aaa;", // SRCALPHA "src_color.aaa;", // SRCALPHA
"float3(1,1,1) - ocol1.aaa;", // INVSRCALPHA "float3(1,1,1) - src_color.aaa;", // INVSRCALPHA
"initial_ocol0.aaa;", // DSTALPHA "initial_ocol0.aaa;", // DSTALPHA
"float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA "float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA
}; };
@ -2028,8 +2049,8 @@ static void WriteBlend(ShaderCode& out, const pixel_shader_uid_data* uid_data)
"1.0;", // ONE "1.0;", // ONE
"initial_ocol0.a;", // DSTCLR "initial_ocol0.a;", // DSTCLR
"1.0 - initial_ocol0.a;", // INVDSTCLR "1.0 - initial_ocol0.a;", // INVDSTCLR
"ocol1.a;", // SRCALPHA "src_color.a;", // SRCALPHA
"1.0 - ocol1.a;", // INVSRCALPHA "1.0 - src_color.a;", // INVSRCALPHA
"initial_ocol0.a;", // DSTALPHA "initial_ocol0.a;", // DSTALPHA
"1.0 - initial_ocol0.a;", // INVDSTALPHA "1.0 - initial_ocol0.a;", // INVDSTALPHA
}; };
@ -2038,8 +2059,8 @@ static void WriteBlend(ShaderCode& out, const pixel_shader_uid_data* uid_data)
"float3(1,1,1);", // ONE "float3(1,1,1);", // ONE
"ocol0.rgb;", // SRCCLR "ocol0.rgb;", // SRCCLR
"float3(1,1,1) - ocol0.rgb;", // INVSRCCLR "float3(1,1,1) - ocol0.rgb;", // INVSRCCLR
"ocol1.aaa;", // SRCALHA "src_color.aaa;", // SRCALHA
"float3(1,1,1) - ocol1.aaa;", // INVSRCALPHA "float3(1,1,1) - src_color.aaa;", // INVSRCALPHA
"initial_ocol0.aaa;", // DSTALPHA "initial_ocol0.aaa;", // DSTALPHA
"float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA "float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA
}; };
@ -2048,12 +2069,14 @@ static void WriteBlend(ShaderCode& out, const pixel_shader_uid_data* uid_data)
"1.0;", // ONE "1.0;", // ONE
"ocol0.a;", // SRCCLR "ocol0.a;", // SRCCLR
"1.0 - ocol0.a;", // INVSRCCLR "1.0 - ocol0.a;", // INVSRCCLR
"ocol1.a;", // SRCALPHA "src_color.a;", // SRCALPHA
"1.0 - ocol1.a;", // INVSRCALPHA "1.0 - src_color.a;", // INVSRCALPHA
"initial_ocol0.a;", // DSTALPHA "initial_ocol0.a;", // DSTALPHA
"1.0 - initial_ocol0.a;", // INVDSTALPHA "1.0 - initial_ocol0.a;", // INVDSTALPHA
}; };
out.Write("\tfloat4 blend_src;\n"); out.Write("\tfloat4 src_color = {};\n"
"\tfloat4 blend_src;",
uid_data->useDstAlpha ? "ocol1" : "ocol0");
out.Write("\tblend_src.rgb = {}\n", blend_src_factor[uid_data->blend_src_factor]); out.Write("\tblend_src.rgb = {}\n", blend_src_factor[uid_data->blend_src_factor]);
out.Write("\tblend_src.a = {}\n", blend_src_factor_alpha[uid_data->blend_src_factor_alpha]); out.Write("\tblend_src.a = {}\n", blend_src_factor_alpha[uid_data->blend_src_factor_alpha]);
out.Write("\tfloat4 blend_dst;\n"); out.Write("\tfloat4 blend_dst;\n");

51
Source/Core/VideoCommon/UberShaderPixel.cpp
View File

@ -57,7 +57,9 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
const bool use_shader_blend = !use_dual_source && host_config.backend_shader_framebuffer_fetch; const bool use_shader_blend = !use_dual_source && host_config.backend_shader_framebuffer_fetch;
const bool use_shader_logic_op = const bool use_shader_logic_op =
!host_config.backend_logic_op && host_config.backend_shader_framebuffer_fetch; !host_config.backend_logic_op && host_config.backend_shader_framebuffer_fetch;
const bool use_framebuffer_fetch = use_shader_blend || use_shader_logic_op; const bool use_framebuffer_fetch =
use_shader_blend || use_shader_logic_op ||
DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z);
const bool early_depth = uid_data->early_depth != 0; const bool early_depth = uid_data->early_depth != 0;
const bool per_pixel_depth = uid_data->per_pixel_depth != 0; const bool per_pixel_depth = uid_data->per_pixel_depth != 0;
const bool bounding_box = host_config.bounding_box; const bool bounding_box = host_config.bounding_box;
@ -1007,8 +1009,21 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
out.Write(" depth = float(zbuffer_zCoord) / 16777216.0;\n"); out.Write(" depth = float(zbuffer_zCoord) / 16777216.0;\n");
} }
out.Write(" // Alpha Test\n" out.Write(" // Alpha Test\n");
" if (bpmem_alphaTest != 0u) {{\n"
if (early_depth && DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z))
{
// Instead of using discard, fetch the framebuffer's color value and use it as the output
// for this fragment.
out.Write(" #define discard_fragment {{ {} = float4(initial_ocol0.xyz, 1.0); return; }}\n",
use_shader_blend ? "real_ocol0" : "ocol0");
}
else
{
out.Write(" #define discard_fragment discard\n");
}
out.Write(" if (bpmem_alphaTest != 0u) {{\n"
" bool comp0 = alphaCompare(TevResult.a, " I_ALPHA ".r, {});\n", " bool comp0 = alphaCompare(TevResult.a, " I_ALPHA ".r, {});\n",
BitfieldExtract<&AlphaTest::comp0>("bpmem_alphaTest")); BitfieldExtract<&AlphaTest::comp0>("bpmem_alphaTest"));
out.Write(" bool comp1 = alphaCompare(TevResult.a, " I_ALPHA ".g, {});\n", out.Write(" bool comp1 = alphaCompare(TevResult.a, " I_ALPHA ".g, {});\n",
@ -1019,13 +1034,13 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
" switch ({}) {{\n", " switch ({}) {{\n",
BitfieldExtract<&AlphaTest::logic>("bpmem_alphaTest")); BitfieldExtract<&AlphaTest::logic>("bpmem_alphaTest"));
out.Write(" case 0u: // AND\n" out.Write(" case 0u: // AND\n"
" if (comp0 && comp1) break; else discard; break;\n" " if (comp0 && comp1) break; else discard_fragment; break;\n"
" case 1u: // OR\n" " case 1u: // OR\n"
" if (comp0 || comp1) break; else discard; break;\n" " if (comp0 || comp1) break; else discard_fragment; break;\n"
" case 2u: // XOR\n" " case 2u: // XOR\n"
" if (comp0 != comp1) break; else discard; break;\n" " if (comp0 != comp1) break; else discard_fragment; break;\n"
" case 3u: // XNOR\n" " case 3u: // XNOR\n"
" if (comp0 == comp1) break; else discard; break;\n" " if (comp0 == comp1) break; else discard_fragment; break;\n"
" }}\n" " }}\n"
" }}\n" " }}\n"
"\n"); "\n");
@ -1194,8 +1209,8 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
"blend_src.rgb = float3(1,1,1);", // ONE "blend_src.rgb = float3(1,1,1);", // ONE
"blend_src.rgb = initial_ocol0.rgb;", // DSTCLR "blend_src.rgb = initial_ocol0.rgb;", // DSTCLR
"blend_src.rgb = float3(1,1,1) - initial_ocol0.rgb;", // INVDSTCLR "blend_src.rgb = float3(1,1,1) - initial_ocol0.rgb;", // INVDSTCLR
"blend_src.rgb = ocol1.aaa;", // SRCALPHA "blend_src.rgb = src_color.aaa;", // SRCALPHA
"blend_src.rgb = float3(1,1,1) - ocol1.aaa;", // INVSRCALPHA "blend_src.rgb = float3(1,1,1) - src_color.aaa;", // INVSRCALPHA
"blend_src.rgb = initial_ocol0.aaa;", // DSTALPHA "blend_src.rgb = initial_ocol0.aaa;", // DSTALPHA
"blend_src.rgb = float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA "blend_src.rgb = float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA
}; };
@ -1204,8 +1219,8 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
"blend_src.a = 1.0;", // ONE "blend_src.a = 1.0;", // ONE
"blend_src.a = initial_ocol0.a;", // DSTCLR "blend_src.a = initial_ocol0.a;", // DSTCLR
"blend_src.a = 1.0 - initial_ocol0.a;", // INVDSTCLR "blend_src.a = 1.0 - initial_ocol0.a;", // INVDSTCLR
"blend_src.a = ocol1.a;", // SRCALPHA "blend_src.a = src_color.a;", // SRCALPHA
"blend_src.a = 1.0 - ocol1.a;", // INVSRCALPHA "blend_src.a = 1.0 - src_color.a;", // INVSRCALPHA
"blend_src.a = initial_ocol0.a;", // DSTALPHA "blend_src.a = initial_ocol0.a;", // DSTALPHA
"blend_src.a = 1.0 - initial_ocol0.a;", // INVDSTALPHA "blend_src.a = 1.0 - initial_ocol0.a;", // INVDSTALPHA
}; };
@ -1214,8 +1229,8 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
"blend_dst.rgb = float3(1,1,1);", // ONE "blend_dst.rgb = float3(1,1,1);", // ONE
"blend_dst.rgb = ocol0.rgb;", // SRCCLR "blend_dst.rgb = ocol0.rgb;", // SRCCLR
"blend_dst.rgb = float3(1,1,1) - ocol0.rgb;", // INVSRCCLR "blend_dst.rgb = float3(1,1,1) - ocol0.rgb;", // INVSRCCLR
"blend_dst.rgb = ocol1.aaa;", // SRCALHA "blend_dst.rgb = src_color.aaa;", // SRCALHA
"blend_dst.rgb = float3(1,1,1) - ocol1.aaa;", // INVSRCALPHA "blend_dst.rgb = float3(1,1,1) - src_color.aaa;", // INVSRCALPHA
"blend_dst.rgb = initial_ocol0.aaa;", // DSTALPHA "blend_dst.rgb = initial_ocol0.aaa;", // DSTALPHA
"blend_dst.rgb = float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA "blend_dst.rgb = float3(1,1,1) - initial_ocol0.aaa;", // INVDSTALPHA
}; };
@ -1224,13 +1239,19 @@ ShaderCode GenPixelShader(APIType api_type, const ShaderHostConfig& host_config,
"blend_dst.a = 1.0;", // ONE "blend_dst.a = 1.0;", // ONE
"blend_dst.a = ocol0.a;", // SRCCLR "blend_dst.a = ocol0.a;", // SRCCLR
"blend_dst.a = 1.0 - ocol0.a;", // INVSRCCLR "blend_dst.a = 1.0 - ocol0.a;", // INVSRCCLR
"blend_dst.a = ocol1.a;", // SRCALPHA "blend_dst.a = src_color.a;", // SRCALPHA
"blend_dst.a = 1.0 - ocol1.a;", // INVSRCALPHA "blend_dst.a = 1.0 - src_color.a;", // INVSRCALPHA
"blend_dst.a = initial_ocol0.a;", // DSTALPHA "blend_dst.a = initial_ocol0.a;", // DSTALPHA
"blend_dst.a = 1.0 - initial_ocol0.a;", // INVDSTALPHA "blend_dst.a = 1.0 - initial_ocol0.a;", // INVDSTALPHA
}; };
out.Write(" if (blend_enable) {{\n" out.Write(" if (blend_enable) {{\n"
" float4 src_color;\n"
" if (bpmem_dstalpha != 0u) {{\n"
" src_color = ocol1;\n"
" }} else {{\n"
" src_color = ocol0;\n"
" }}"
" float4 blend_src;\n"); " float4 blend_src;\n");
WriteSwitch(out, api_type, "blend_src_factor", blendSrcFactor, 4, true); WriteSwitch(out, api_type, "blend_src_factor", blendSrcFactor, 4, true);
WriteSwitch(out, api_type, "blend_src_factor_alpha", blendSrcFactorAlpha, 4, true); WriteSwitch(out, api_type, "blend_src_factor_alpha", blendSrcFactorAlpha, 4, true);