346 Source/Core/VideoCommon/FramebufferShaderGen.cpp
View file
@@ -1,26 +1,35 @@
// Copyright 2019 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.

#include "VideoCommon/FramebufferShaderGen.h"

#include <sstream>
#include <string_view>

#include "VideoCommon/FramebufferManager.h"
#include "VideoCommon/TextureDecoder.h"
#include "VideoCommon/VertexShaderGen.h"

namespace FramebufferShaderGen
{
static APIType GetAPIType()
namespace
{
APIType GetAPIType()
{
return g_ActiveConfig.backend_info.api_type;
}

static void EmitUniformBufferDeclaration(std::stringstream& ss)
void EmitUniformBufferDeclaration(std::ostringstream& ss)
{
if (GetAPIType() == APIType::D3D)
ss << "cbuffer PSBlock : register(b0)\n";
else
ss << "UBO_BINDING(std140, 1) uniform PSBlock\n";
}

static void EmitSamplerDeclarations(std::stringstream& ss, u32 start = 0, u32 end = 1,
bool multisampled = false)
void EmitSamplerDeclarations(std::ostringstream& ss, u32 start = 0, u32 end = 1,
bool multisampled = false)
{
switch (GetAPIType())
{
@@ -51,17 +60,17 @@ static void EmitSamplerDeclarations(std::stringstream& ss, u32 start = 0, u32 en
}
}

static void EmitSampleTexture(std::stringstream& ss, u32 n, const char* coords)
void EmitSampleTexture(std::ostringstream& ss, u32 n, std::string_view coords)
{
switch (GetAPIType())
{
case APIType::D3D:
ss << "tex" << n << ".Sample(samp" << n << ", " << coords << ")";
ss << "tex" << n << ".Sample(samp" << n << ", " << coords << ')';
break;

case APIType::OpenGL:
case APIType::Vulkan:
ss << "texture(samp" << n << ", " << coords << ")";
ss << "texture(samp" << n << ", " << coords << ')';
break;

default:
@@ -71,12 +80,12 @@ static void EmitSampleTexture(std::stringstream& ss, u32 n, const char* coords)

// Emits a texel fetch/load instruction. Assumes that "coords" is a 4-element vector, with z
// containing the layer, and w containing the mipmap level.
static void EmitTextureLoad(std::stringstream& ss, u32 n, const char* coords)
void EmitTextureLoad(std::ostringstream& ss, u32 n, std::string_view coords)
{
switch (GetAPIType())
{
case APIType::D3D:
ss << "tex" << n << ".Load(" << coords << ")";
ss << "tex" << n << ".Load(" << coords << ')';
break;

case APIType::OpenGL:
@@ -89,10 +98,9 @@ static void EmitTextureLoad(std::stringstream& ss, u32 n, const char* coords)
}
}

static void EmitVertexMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
u32 num_color_inputs, bool position_input,
u32 num_tex_outputs, u32 num_color_outputs,
const char* extra_inputs = "")
void EmitVertexMainDeclaration(std::ostringstream& ss, u32 num_tex_inputs, u32 num_color_inputs,
bool position_input, u32 num_tex_outputs, u32 num_color_outputs,
std::string_view extra_inputs = {})
{
switch (GetAPIType())
{
@@ -118,11 +126,15 @@ static void EmitVertexMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
case APIType::Vulkan:
{
for (u32 i = 0; i < num_tex_inputs; i++)
{
ss << "ATTRIBUTE_LOCATION(" << (SHADER_TEXTURE0_ATTRIB + i) << ") in float3 rawtex" << i
<< ";\n";
}
for (u32 i = 0; i < num_color_inputs; i++)
{
ss << "ATTRIBUTE_LOCATION(" << (SHADER_COLOR0_ATTRIB + i) << ") in float4 rawcolor" << i
<< ";\n";
}
if (position_input)
ss << "ATTRIBUTE_LOCATION(" << SHADER_POSITION_ATTRIB << ") in float4 rawpos;\n";

@@ -143,7 +155,7 @@ static void EmitVertexMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
ss << "VARYING_LOCATION(" << (num_tex_inputs + i) << ") out float4 v_col" << i << ";\n";
}
ss << "#define opos gl_Position\n";
ss << extra_inputs << "\n";
ss << extra_inputs << '\n';
ss << "void main()\n";
}
break;
@@ -152,9 +164,9 @@ static void EmitVertexMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
}
}

static void EmitPixelMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
u32 num_color_inputs, const char* output_type = "float4",
const char* extra_vars = "", bool emit_frag_coord = false)
void EmitPixelMainDeclaration(std::ostringstream& ss, u32 num_tex_inputs, u32 num_color_inputs,
std::string_view output_type = "float4",
std::string_view extra_vars = {}, bool emit_frag_coord = false)
{
switch (GetAPIType())
{
@@ -203,16 +215,17 @@ static void EmitPixelMainDeclaration(std::stringstream& ss, u32 num_tex_inputs,
break;
}
}
} // Anonymous namespace

std::string GenerateScreenQuadVertexShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitVertexMainDeclaration(ss, 0, 0, false, 1, 0,
GetAPIType() == APIType::D3D ? "in uint id : SV_VertexID, " :
"#define id gl_VertexID\n");
ss << "{\n";
ss << " v_tex0 = float3(float((id << 1) & 2), float(id & 2), 0.0f);\n";
ss << " opos = float4(v_tex0.xy * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);\n";
ss << "{\n"
" v_tex0 = float3(float((id << 1) & 2), float(id & 2), 0.0f);\n"
" opos = float4(v_tex0.xy * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);\n";

// NDC space is flipped in Vulkan. We also flip in GL so that (0,0) is in the lower-left.
if (GetAPIType() == APIType::Vulkan || GetAPIType() == APIType::OpenGL)
@@ -225,50 +238,52 @@ std::string GenerateScreenQuadVertexShader()

std::string GeneratePassthroughGeometryShader(u32 num_tex, u32 num_colors)
{
std::stringstream ss;
std::ostringstream ss;
if (GetAPIType() == APIType::D3D)
{
ss << "struct VS_OUTPUT\n";
ss << "{\n";
ss << "struct VS_OUTPUT\n"
"{\n";
for (u32 i = 0; i < num_tex; i++)
ss << " float3 tex" << i << " : TEXCOORD" << i << ";\n";
for (u32 i = 0; i < num_colors; i++)
ss << " float4 color" << i << " : COLOR" << i << ";\n";
ss << " float4 position : SV_Position;\n";
ss << "};\n";
ss << "struct GS_OUTPUT\n";
ss << "{";
ss << " float4 position : SV_Position;\n"
"};\n";

ss << "struct GS_OUTPUT\n"
"{";
for (u32 i = 0; i < num_tex; i++)
ss << " float3 tex" << i << " : TEXCOORD" << i << ";\n";
for (u32 i = 0; i < num_colors; i++)
ss << " float4 color" << i << " : COLOR" << i << ";\n";
ss << " float4 position : SV_Position;\n";
ss << " uint slice : SV_RenderTargetArrayIndex;\n";
ss << "};\n\n";
ss << "[maxvertexcount(6)]\n";
ss << "void main(triangle VS_OUTPUT vso[3], inout TriangleStream<GS_OUTPUT> output)\n";
ss << "{\n";
ss << " for (uint slice = 0; slice < 2u; slice++)\n";
ss << " {\n";
ss << " for (int i = 0; i < 3; i++)\n";
ss << " {\n";
ss << " GS_OUTPUT gso;\n";
ss << " gso.position = vso[i].position;\n";
ss << " float4 position : SV_Position;\n"
" uint slice : SV_RenderTargetArrayIndex;\n"
"};\n\n";

ss << "[maxvertexcount(6)]\n"
"void main(triangle VS_OUTPUT vso[3], inout TriangleStream<GS_OUTPUT> output)\n"
"{\n"
" for (uint slice = 0; slice < 2u; slice++)\n"
" {\n"
" for (int i = 0; i < 3; i++)\n"
" {\n"
" GS_OUTPUT gso;\n"
" gso.position = vso[i].position;\n";
for (u32 i = 0; i < num_tex; i++)
ss << " gso.tex" << i << " = float3(vso[i].tex" << i << ".xy, float(slice));\n";
for (u32 i = 0; i < num_colors; i++)
ss << " gso.color" << i << " = vso[i].color" << i << ";\n";
ss << " gso.slice = slice;\n";
ss << " output.Append(gso);\n";
ss << " }\n";
ss << " output.RestartStrip();\n";
ss << " }\n";
ss << "}\n";
ss << " gso.slice = slice;\n"
" output.Append(gso);\n"
" }\n"
" output.RestartStrip();\n"
" }\n"
"}\n";
}
else if (GetAPIType() == APIType::OpenGL || GetAPIType() == APIType::Vulkan)
{
ss << "layout(triangles) in;\n";
ss << "layout(triangle_strip, max_vertices = 6) out;\n";
ss << "layout(triangles) in;\n"
"layout(triangle_strip, max_vertices = 6) out;\n";
if (num_tex > 0 || num_colors > 0)
{
ss << "VARYING_LOCATION(0) in VertexData {\n";
@@ -277,19 +292,20 @@ std::string GeneratePassthroughGeometryShader(u32 num_tex, u32 num_colors)
for (u32 i = 0; i < num_colors; i++)
ss << " float4 v_col" << i << ";\n";
ss << "} v_in[];\n";

ss << "VARYING_LOCATION(0) out VertexData {\n";
for (u32 i = 0; i < num_tex; i++)
ss << " float3 v_tex" << i << ";\n";
for (u32 i = 0; i < num_colors; i++)
ss << " float4 v_col" << i << ";\n";
ss << "} v_out;\n";
}
ss << "\n";
ss << "void main()\n";
ss << "{\n";
ss << " for (int j = 0; j < 2; j++)\n";
ss << " {\n";
ss << " gl_Layer = j;\n";
ss << "\n"
"void main()\n"
"{\n"
" for (int j = 0; j < 2; j++)\n"
" {\n"
" gl_Layer = j;\n";

// We have to explicitly unroll this loop otherwise the GL compiler gets cranky.
for (u32 v = 0; v < 3; v++)
@@ -302,30 +318,30 @@ std::string GeneratePassthroughGeometryShader(u32 num_tex, u32 num_colors)
ss << " v_out.v_col" << i << " = v_in[" << v << "].v_col" << i << ";\n";
ss << " EmitVertex();\n\n";
}
ss << " EndPrimitive();\n";
ss << " }\n";
ss << "}\n";
ss << " EndPrimitive();\n"
" }\n"
"}\n";
}

return ss.str();
}

std::string GenerateTextureCopyVertexShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitUniformBufferDeclaration(ss);
ss << "{";
ss << " float2 src_offset;\n";
ss << " float2 src_size;\n";
ss << "};\n\n";
ss << "{"
" float2 src_offset;\n"
" float2 src_size;\n"
"};\n\n";

EmitVertexMainDeclaration(ss, 0, 0, false, 1, 0,
GetAPIType() == APIType::D3D ? "in uint id : SV_VertexID, " :
"#define id gl_VertexID");
ss << "{\n";
ss << " v_tex0 = float3(float((id << 1) & 2), float(id & 2), 0.0f);\n";
ss << " opos = float4(v_tex0.xy * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);\n";
ss << " v_tex0 = float3(src_offset + (src_size * v_tex0.xy), 0.0f);\n";
ss << "{\n"
" v_tex0 = float3(float((id << 1) & 2), float(id & 2), 0.0f);\n"
" opos = float4(v_tex0.xy * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);\n"
" v_tex0 = float3(src_offset + (src_size * v_tex0.xy), 0.0f);\n";

// NDC space is flipped in Vulkan. We also flip in GL so that (0,0) is in the lower-left.
if (GetAPIType() == APIType::Vulkan || GetAPIType() == APIType::OpenGL)
@@ -338,35 +354,35 @@ std::string GenerateTextureCopyVertexShader()

std::string GenerateTextureCopyPixelShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 1, false);
EmitPixelMainDeclaration(ss, 1, 0);
ss << "{\n";
ss << " ocol0 = ";
ss << "{\n"
" ocol0 = ";
EmitSampleTexture(ss, 0, "v_tex0");
ss << ";\n";
ss << "}\n";
ss << ";\n"
"}\n";
return ss.str();
}

std::string GenerateColorPixelShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitPixelMainDeclaration(ss, 0, 1);
ss << "{\n";
ss << " ocol0 = v_col0;\n";
ss << "}\n";
ss << "{\n"
" ocol0 = v_col0;\n"
"}\n";
return ss.str();
}

std::string GenerateResolveDepthPixelShader(u32 samples)
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 1, true);
EmitPixelMainDeclaration(ss, 1, 0, "float",
GetAPIType() == APIType::D3D ? "in float4 ipos : SV_Position, " : "");
ss << "{\n";
ss << " int layer = int(v_tex0.z);\n";
ss << "{\n"
" int layer = int(v_tex0.z);\n";
if (GetAPIType() == APIType::D3D)
ss << " int3 coords = int3(int2(ipos.xy), layer);\n";
else
@@ -389,20 +405,20 @@ std::string GenerateResolveDepthPixelShader(u32 samples)

std::string GenerateClearVertexShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitUniformBufferDeclaration(ss);
ss << "{\n";
ss << " float4 clear_color;\n";
ss << " float clear_depth;\n";
ss << "};\n";
ss << "{\n"
" float4 clear_color;\n"
" float clear_depth;\n"
"};\n";

EmitVertexMainDeclaration(ss, 0, 0, false, 0, 1,
GetAPIType() == APIType::D3D ? "in uint id : SV_VertexID, " :
"#define id gl_VertexID\n");
ss << "{\n";
ss << " float2 coord = float2(float((id << 1) & 2), float(id & 2));\n";
ss << " opos = float4(coord * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), clear_depth, 1.0f);\n";
ss << " v_col0 = clear_color;\n";
ss << "{\n"
" float2 coord = float2(float((id << 1) & 2), float(id & 2));\n"
" opos = float4(coord * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), clear_depth, 1.0f);\n"
" v_col0 = clear_color;\n";

// NDC space is flipped in Vulkan
if (GetAPIType() == APIType::Vulkan)
@@ -415,11 +431,11 @@ std::string GenerateClearVertexShader()

std::string GenerateEFBPokeVertexShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitVertexMainDeclaration(ss, 0, 1, true, 0, 1);
ss << "{\n";
ss << " v_col0 = rawcolor0;\n";
ss << " opos = float4(rawpos.xyz, 1.0f);\n";
ss << "{\n"
" v_col0 = rawcolor0;\n"
" opos = float4(rawpos.xyz, 1.0f);\n";
if (g_ActiveConfig.backend_info.bSupportsLargePoints)
ss << " gl_PointSize = rawpos.w;\n";

@@ -433,7 +449,7 @@ std::string GenerateEFBPokeVertexShader()

std::string GenerateFormatConversionShader(EFBReinterpretType convtype, u32 samples)
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 1, samples > 1);
EmitPixelMainDeclaration(
ss, 1, 0, "float4",
@@ -442,8 +458,8 @@ std::string GenerateFormatConversionShader(EFBReinterpretType convtype, u32 samp
"in float4 ipos : SV_Position, in uint isample : SV_SampleIndex, " :
"in float4 ipos : SV_Position, ") :
"");
ss << "{\n";
ss << " int layer = int(v_tex0.z);\n";
ss << "{\n"
" int layer = int(v_tex0.z);\n";
if (GetAPIType() == APIType::D3D)
ss << " int3 coords = int3(int2(ipos.xy), layer);\n";
else
@@ -480,27 +496,27 @@ std::string GenerateFormatConversionShader(EFBReinterpretType convtype, u32 samp
switch (convtype)
{
case EFBReinterpretType::RGB8ToRGBA6:
ss << " int4 src8 = int4(round(val * 255.f));\n";
ss << " int4 dst6;\n";
ss << " dst6.r = src8.r >> 2;\n";
ss << " dst6.g = ((src8.r & 0x3) << 4) | (src8.g >> 4);\n";
ss << " dst6.b = ((src8.g & 0xF) << 2) | (src8.b >> 6);\n";
ss << " dst6.a = src8.b & 0x3F;\n";
ss << " ocol0 = float4(dst6) / 63.f;\n";
ss << " int4 src8 = int4(round(val * 255.f));\n"
" int4 dst6;\n"
" dst6.r = src8.r >> 2;\n"
" dst6.g = ((src8.r & 0x3) << 4) | (src8.g >> 4);\n"
" dst6.b = ((src8.g & 0xF) << 2) | (src8.b >> 6);\n"
" dst6.a = src8.b & 0x3F;\n"
" ocol0 = float4(dst6) / 63.f;\n";
break;

case EFBReinterpretType::RGB8ToRGB565:
ss << " ocol0 = val;\n";
break;

case EFBReinterpretType::RGBA6ToRGB8:
ss << " int4 src6 = int4(round(val * 63.f));\n";
ss << " int4 dst8;\n";
ss << " dst8.r = (src6.r << 2) | (src6.g >> 4);\n";
ss << " dst8.g = ((src6.g & 0xF) << 4) | (src6.b >> 2);\n";
ss << " dst8.b = ((src6.b & 0x3) << 6) | src6.a;\n";
ss << " dst8.a = 255;\n";
ss << " ocol0 = float4(dst8) / 255.f;\n";
ss << " int4 src6 = int4(round(val * 63.f));\n"
" int4 dst8;\n"
" dst8.r = (src6.r << 2) | (src6.g >> 4);\n"
" dst8.g = ((src6.g & 0xF) << 4) | (src6.b >> 2);\n"
" dst8.b = ((src6.b & 0x3) << 6) | src6.a;\n"
" dst8.a = 255;\n"
" ocol0 = float4(dst8) / 255.f;\n";
break;

case EFBReinterpretType::RGBA6ToRGB565:
@@ -523,12 +539,12 @@ std::string GenerateFormatConversionShader(EFBReinterpretType convtype, u32 samp

std::string GenerateTextureReinterpretShader(TextureFormat from_format, TextureFormat to_format)
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 1, false);
EmitPixelMainDeclaration(ss, 1, 0, "float4", "", true);
ss << "{\n";
ss << " int layer = int(v_tex0.z);\n";
ss << " int4 coords = int4(int2(frag_coord.xy), layer, 0);\n";
ss << "{\n"
" int layer = int(v_tex0.z);\n"
" int4 coords = int4(int2(frag_coord.xy), layer, 0);\n";

// Convert to a 32-bit value encompassing all channels, filling the most significant bits with
// zeroes.
@@ -540,36 +556,36 @@ std::string GenerateTextureReinterpretShader(TextureFormat from_format, TextureF
{
ss << " float4 temp_value = ";
EmitTextureLoad(ss, 0, "coords");
ss << ";\n";
ss << " raw_value = uint(temp_value.r * 255.0);\n";
ss << ";\n"
" raw_value = uint(temp_value.r * 255.0);\n";
}
break;

case TextureFormat::IA8:
{
ss << " float4 temp_value = ";
EmitTextureLoad(ss, 0, "coords");
ss << ";\n";
ss << " raw_value = uint(temp_value.r * 255.0) | (uint(temp_value.a * 255.0) << 8);\n";
ss << ";\n"
" raw_value = uint(temp_value.r * 255.0) | (uint(temp_value.a * 255.0) << 8);\n";
}
break;

case TextureFormat::IA4:
{
ss << " float4 temp_value = ";
EmitTextureLoad(ss, 0, "coords");
ss << ";\n";
ss << " raw_value = uint(temp_value.r * 15.0) | (uint(temp_value.a * 15.0) << 4);\n";
ss << ";\n"
" raw_value = uint(temp_value.r * 15.0) | (uint(temp_value.a * 15.0) << 4);\n";
}
break;

case TextureFormat::RGB565:
{
ss << " float4 temp_value = ";
EmitTextureLoad(ss, 0, "coords");
ss << ";\n";
ss << " raw_value = uint(temp_value.b * 31.0) | (uint(temp_value.g * 63.0) << 5) |\n";
ss << " (uint(temp_value.r * 31.0) << 11);\n";
ss << ";\n"
" raw_value = uint(temp_value.b * 31.0) | (uint(temp_value.g * 63.0) << 5) |\n"
" (uint(temp_value.r * 31.0) << 11);\n";
}
break;

@@ -580,13 +596,13 @@ std::string GenerateTextureReinterpretShader(TextureFormat from_format, TextureF
ss << ";\n";

// 0.8784 = 224 / 255 which is the maximum alpha value that can be represented in 3 bits
ss << " if (temp_value.a > 0.878f) {\n";
ss << " raw_value = (uint(temp_value.b * 31.0)) | (uint(temp_value.g * 31.0) << 5) |\n";
ss << " (uint(temp_value.r * 31.0) << 10) | 0x8000u;\n";
ss << " } else {\n";
ss << " raw_value = (uint(temp_value.b * 15.0)) | (uint(temp_value.g * 15.0) << 4) |\n";
ss << " (uint(temp_value.r * 15.0) << 8) | (uint(temp_value.a * 7.0) << 12);\n";
ss << " }\n";
ss << " if (temp_value.a > 0.878f) {\n"
" raw_value = (uint(temp_value.b * 31.0)) | (uint(temp_value.g * 31.0) << 5) |\n"
" (uint(temp_value.r * 31.0) << 10) | 0x8000u;\n"
" } else {\n"
" raw_value = (uint(temp_value.b * 15.0)) | (uint(temp_value.g * 15.0) << 4) |\n"
" (uint(temp_value.r * 15.0) << 8) | (uint(temp_value.a * 7.0) << 12);\n"
" }\n";
}
break;
}
@@ -597,45 +613,45 @@ std::string GenerateTextureReinterpretShader(TextureFormat from_format, TextureF
case TextureFormat::I8:
case TextureFormat::C8:
{
ss << " float orgba = float(raw_value & 0xFFu) / 255.0;\n";
ss << " ocol0 = float4(orgba, orgba, orgba, orgba);\n";
ss << " float orgba = float(raw_value & 0xFFu) / 255.0;\n"
" ocol0 = float4(orgba, orgba, orgba, orgba);\n";
}
break;

case TextureFormat::IA8:
{
ss << " float orgb = float(raw_value & 0xFFu) / 255.0;\n";
ss << " ocol0 = float4(orgb, orgb, orgb, float((raw_value >> 8) & 0xFFu) / 255.0);\n";
ss << " float orgb = float(raw_value & 0xFFu) / 255.0;\n"
" ocol0 = float4(orgb, orgb, orgb, float((raw_value >> 8) & 0xFFu) / 255.0);\n";
}
break;

case TextureFormat::IA4:
{
ss << " float orgb = float(raw_value & 0xFu) / 15.0;\n";
ss << " ocol0 = float4(orgb, orgb, orgb, float((raw_value >> 4) & 0xFu) / 15.0);\n";
ss << " float orgb = float(raw_value & 0xFu) / 15.0;\n"
" ocol0 = float4(orgb, orgb, orgb, float((raw_value >> 4) & 0xFu) / 15.0);\n";
}
break;

case TextureFormat::RGB565:
{
ss << " ocol0 = float4(float((raw_value >> 10) & 0x1Fu) / 31.0,\n";
ss << " float((raw_value >> 5) & 0x1Fu) / 31.0,\n";
ss << " float(raw_value & 0x1Fu) / 31.0, 1.0);\n";
ss << " ocol0 = float4(float((raw_value >> 10) & 0x1Fu) / 31.0,\n"
" float((raw_value >> 5) & 0x1Fu) / 31.0,\n"
" float(raw_value & 0x1Fu) / 31.0, 1.0);\n";
}
break;

case TextureFormat::RGB5A3:
{
ss << " if ((raw_value & 0x8000u) != 0u) {\n";
ss << " ocol0 = float4(float((raw_value >> 10) & 0x1Fu) / 31.0,\n";
ss << " float((raw_value >> 5) & 0x1Fu) / 31.0,\n";
ss << " float(raw_value & 0x1Fu) / 31.0, 1.0);\n";
ss << " } else {\n";
ss << " ocol0 = float4(float((raw_value >> 8) & 0x0Fu) / 15.0,\n";
ss << " float((raw_value >> 4) & 0x0Fu) / 15.0,\n";
ss << " float(raw_value & 0x0Fu) / 15.0,\n";
ss << " float((raw_value >> 12) & 0x07u) / 7.0);\n";
ss << " }\n";
ss << " if ((raw_value & 0x8000u) != 0u) {\n"
" ocol0 = float4(float((raw_value >> 10) & 0x1Fu) / 31.0,\n"
" float((raw_value >> 5) & 0x1Fu) / 31.0,\n"
" float(raw_value & 0x1Fu) / 31.0, 1.0);\n"
" } else {\n"
" ocol0 = float4(float((raw_value >> 8) & 0x0Fu) / 15.0,\n"
" float((raw_value >> 4) & 0x0Fu) / 15.0,\n"
" float(raw_value & 0x0Fu) / 15.0,\n"
" float((raw_value >> 12) & 0x07u) / 7.0);\n"
" }\n";
}
break;
}
@@ -646,37 +662,37 @@ std::string GenerateTextureReinterpretShader(TextureFormat from_format, TextureF

std::string GenerateEFBRestorePixelShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 2, false);
EmitPixelMainDeclaration(ss, 1, 0, "float4",
GetAPIType() == APIType::D3D ? "out float depth : SV_Depth, " : "");
ss << "{\n";
ss << " ocol0 = ";
ss << "{\n"
" ocol0 = ";
EmitSampleTexture(ss, 0, "v_tex0");
ss << ";\n";
ss << " " << (GetAPIType() == APIType::D3D ? "depth" : "gl_FragDepth") << " = ";
EmitSampleTexture(ss, 1, "v_tex0");
ss << ".r;\n";
ss << "}\n";
ss << ".r;\n"
"}\n";
return ss.str();
}

std::string GenerateImGuiVertexShader()
{
std::stringstream ss;
std::ostringstream ss;

// Uniform buffer contains the viewport size, and we transform in the vertex shader.
EmitUniformBufferDeclaration(ss);
ss << "{\n";
ss << "float2 u_rcp_viewport_size_mul2;\n";
ss << "};\n\n";
ss << "{\n"
"float2 u_rcp_viewport_size_mul2;\n"
"};\n\n";

EmitVertexMainDeclaration(ss, 1, 1, true, 1, 1);
ss << "{\n"
<< " v_tex0 = float3(rawtex0.xy, 0.0);\n"
<< " v_col0 = rawcolor0;\n"
<< " opos = float4(rawpos.x * u_rcp_viewport_size_mul2.x - 1.0,"
<< " 1.0 - rawpos.y * u_rcp_viewport_size_mul2.y, 0.0, 1.0);\n";
" v_tex0 = float3(rawtex0.xy, 0.0);\n"
" v_col0 = rawcolor0;\n"
" opos = float4(rawpos.x * u_rcp_viewport_size_mul2.x - 1.0,"
" 1.0 - rawpos.y * u_rcp_viewport_size_mul2.y, 0.0, 1.0);\n";

// NDC space is flipped in Vulkan.
if (GetAPIType() == APIType::Vulkan)
@@ -688,14 +704,14 @@ std::string GenerateImGuiVertexShader()

std::string GenerateImGuiPixelShader()
{
std::stringstream ss;
std::ostringstream ss;
EmitSamplerDeclarations(ss, 0, 1, false);
EmitPixelMainDeclaration(ss, 1, 1);
ss << "{\n";
ss << " ocol0 = ";
ss << "{\n"
" ocol0 = ";
EmitSampleTexture(ss, 0, "float3(v_tex0.xy, 0.0)");
ss << " * v_col0;\n";
ss << "}\n";
ss << " * v_col0;\n"
"}\n";

return ss.str();
}
20 Source/Core/VideoCommon/FramebufferShaderGen.h
View file
@@ -1,25 +1,17 @@
// Copyright 2019 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.

#pragma once

#include <string>
#include "VideoCommon/VideoCommon.h"
#include "Common/CommonTypes.h"

enum class EFBReinterpretType;
enum class TextureFormat;

namespace FramebufferShaderGen
{
struct Config
{
Config(APIType api_type_, u32 efb_samples_, u32 efb_layers_, bool ssaa_)
: api_type(api_type_), efb_samples(efb_samples_), efb_layers(efb_layers_), ssaa(ssaa_)
{
}

APIType api_type;
u32 efb_samples;
u32 efb_layers;
bool ssaa;
};

std::string GenerateScreenQuadVertexShader();
std::string GeneratePassthroughGeometryShader(u32 num_tex, u32 num_colors);
std::string GenerateTextureCopyVertexShader();