VideoCommon: Add vertex shader point and line expansion

1 files changed, 216 insertions, 20 deletions

diff --git a/Source/Core/VideoCommon/VertexShaderGen.cpp b/Source/Core/VideoCommon/VertexShaderGen.cpp
index 2eb5a53a97..948dc3ad34 100644
--- a/Source/Core/VideoCommon/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/VertexShaderGen.cpp
@@ -6,6 +6,7 @@
 #include "Common/Assert.h"
 #include "Common/CommonTypes.h"
 #include "VideoCommon/BPMemory.h"
+#include "VideoCommon/ConstantManager.h"
 #include "VideoCommon/LightingShaderGen.h"
 #include "VideoCommon/NativeVertexFormat.h"
 #include "VideoCommon/VertexLoaderManager.h"
@@ -83,6 +84,8 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const ShaderHostConfig& ho
   const bool ssaa = host_config.ssaa;
   const bool vertex_rounding = host_config.vertex_rounding;
 
+  ShaderCode input_extract;
+
   out.Write("{}", s_lighting_struct);
 
   // uniforms
@@ -91,6 +94,21 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const ShaderHostConfig& ho
   out.Write("{}", s_shader_uniforms);
   out.Write("}};\n");
 
+  if (uid_data->vs_expand != VSExpand::None)
+  {
+    out.Write("UBO_BINDING(std140, 3) uniform GSBlock {{\n");
+    out.Write("{}", s_geometry_shader_uniforms);
+    out.Write("}};\n");
+
+    if (api_type == APIType::D3D)
+    {
+      // D3D doesn't include the base vertex in SV_VertexID
+      out.Write("UBO_BINDING(std140, 4) uniform DX_Constants {{\n"
+                "  uint base_vertex;\n"
+                "}};\n\n");
+    }
+  }
+
   out.Write("struct VS_OUTPUT {{\n");
   GenerateVSOutputMembers(out, api_type, uid_data->numTexGens, host_config, "",
                           ShaderStage::Vertex);
@@ -98,30 +116,113 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const ShaderHostConfig& ho
 
   WriteIsNanHeader(out, api_type);
 
-  out.Write("ATTRIBUTE_LOCATION({}) in float4 rawpos;\n", SHADER_POSITION_ATTRIB);
-  if ((uid_data->components & VB_HAS_POSMTXIDX) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in uint4 posmtx;\n", SHADER_POSMTX_ATTRIB);
-  if ((uid_data->components & VB_HAS_NORMAL) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in float3 rawnormal;\n", SHADER_NORMAL_ATTRIB);
-  if ((uid_data->components & VB_HAS_TANGENT) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in float3 rawtangent;\n", SHADER_TANGENT_ATTRIB);
-  if ((uid_data->components & VB_HAS_BINORMAL) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in float3 rawbinormal;\n", SHADER_BINORMAL_ATTRIB);
-
-  if ((uid_data->components & VB_HAS_COL0) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in float4 rawcolor0;\n", SHADER_COLOR0_ATTRIB);
-  if ((uid_data->components & VB_HAS_COL1) != 0)
-    out.Write("ATTRIBUTE_LOCATION({}) in float4 rawcolor1;\n", SHADER_COLOR1_ATTRIB);
-
-  for (u32 i = 0; i < 8; ++i)
+  if (uid_data->vs_expand == VSExpand::None)
   {
-    const u32 has_texmtx = (uid_data->components & (VB_HAS_TEXMTXIDX0 << i));
+    out.Write("ATTRIBUTE_LOCATION({}) in float4 rawpos;\n", SHADER_POSITION_ATTRIB);
+    if ((uid_data->components & VB_HAS_POSMTXIDX) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in uint4 posmtx;\n", SHADER_POSMTX_ATTRIB);
+    if ((uid_data->components & VB_HAS_NORMAL) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in float3 rawnormal;\n", SHADER_NORMAL_ATTRIB);
+    if ((uid_data->components & VB_HAS_TANGENT) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in float3 rawtangent;\n", SHADER_TANGENT_ATTRIB);
+    if ((uid_data->components & VB_HAS_BINORMAL) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in float3 rawbinormal;\n", SHADER_BINORMAL_ATTRIB);
+
+    if ((uid_data->components & VB_HAS_COL0) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in float4 rawcolor0;\n", SHADER_COLOR0_ATTRIB);
+    if ((uid_data->components & VB_HAS_COL1) != 0)
+      out.Write("ATTRIBUTE_LOCATION({}) in float4 rawcolor1;\n", SHADER_COLOR1_ATTRIB);
+
+    for (u32 i = 0; i < 8; ++i)
+    {
+      const u32 has_texmtx = (uid_data->components & (VB_HAS_TEXMTXIDX0 << i));
 
-    if ((uid_data->components & (VB_HAS_UV0 << i)) != 0 || has_texmtx != 0)
+      if ((uid_data->components & (VB_HAS_UV0 << i)) != 0 || has_texmtx != 0)
+      {
+        out.Write("ATTRIBUTE_LOCATION({}) in float{} rawtex{};\n", SHADER_TEXTURE0_ATTRIB + i,
+                  has_texmtx != 0 ? 3 : 2, i);
+      }
+    }
+  }
+  else
+  {
+    // Can't use float3, etc because we want 4-byte alignment
+    out.Write(
+        "uint4 unpack_ubyte4(uint value) {{\n"
+        "  return uint4(value & 0xff, (value >> 8) & 0xff, (value >> 16) & 0xff, value >> 24);\n"
+        "}}\n\n"
+        "struct InputData {{\n");
+    if (uid_data->components & VB_HAS_POSMTXIDX)
     {
-      out.Write("ATTRIBUTE_LOCATION({}) in float{} rawtex{};\n", SHADER_TEXTURE0_ATTRIB + i,
-                has_texmtx != 0 ? 3 : 2, i);
+      out.Write("  uint posmtx;\n");
+      input_extract.Write("uint4 posmtx = unpack_ubyte4(i.posmtx);\n");
+    }
+    if (uid_data->position_has_3_elems)
+    {
+      out.Write("  float pos0;\n"
+                "  float pos1;\n"
+                "  float pos2;\n");
+      input_extract.Write("float4 rawpos = float4(i.pos0, i.pos1, i.pos2, 1.0f);\n");
+    }
+    else
+    {
+      out.Write("  float pos0;\n"
+                "  float pos1;\n");
+      input_extract.Write("float4 rawpos = float4(i.pos0, i.pos1, 0.0f, 1.0f);\n");
+    }
+    std::array<std::string_view, 3> names = {"normal", "binormal", "tangent"};
+    for (int i = 0; i < 3; i++)
+    {
+      if (uid_data->components & (VB_HAS_NORMAL << i))
+      {
+        out.Write("  float {0}0;\n"
+                  "  float {0}1;\n"
+                  "  float {0}2;\n",
+                  names[i]);
+        input_extract.Write("float3 raw{0} = float3(i.{0}0, i.{0}1, i.{0}2);\n", names[i]);
+      }
+    }
+    for (int i = 0; i < 2; i++)
+    {
+      if (uid_data->components & (VB_HAS_COL0 << i))
+      {
+        out.Write("  uint color{};\n", i);
+        input_extract.Write("float4 rawcolor{0} = float4(unpack_ubyte4(i.color{0})) / 255.0f;\n",
+                            i);
+      }
+    }
+    for (int i = 0; i < 8; i++)
+    {
+      if (uid_data->components & (VB_HAS_UV0 << i))
+      {
+        u32 ncomponents = (uid_data->texcoord_elem_count >> (2 * i)) & 3;
+        if (ncomponents < 2)
+        {
+          out.Write("  float tex{};\n", i);
+          input_extract.Write("float3 rawtex{0} = float3(i.tex{0}, 0.0f, 0.0f);\n", i);
+        }
+        else if (ncomponents == 2)
+        {
+          out.Write("  float tex{0}_0;\n"
+                    "  float tex{0}_1;\n",
+                    i);
+          input_extract.Write("float3 rawtex{0} = float3(i.tex{0}_0, i.tex{0}_1, 0.0f);\n", i);
+        }
+        else
+        {
+          out.Write("  float tex{0}_0;\n"
+                    "  float tex{0}_1;\n"
+                    "  float tex{0}_2;\n",
+                    i);
+          input_extract.Write("float3 rawtex{0} = float3(i.tex{0}_0, i.tex{0}_1, i.tex{0}_2);\n",
+                              i);
+        }
+      }
     }
+    out.Write("}};\n\n"
+              "SSBO_BINDING(1) readonly restrict buffer InputBuffer {{\n"
+              "  InputData input_buffer[];\n"
+              "}};\n\n");
   }
 
   if (host_config.backend_geometry_shaders)
@@ -161,6 +262,21 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const ShaderHostConfig& ho
 
   out.Write("void main()\n{{\n");
 
+  if (uid_data->vs_expand != VSExpand::None)
+  {
+    out.Write("bool is_bottom = (gl_VertexID & 2) != 0;\n"
+              "bool is_right = (gl_VertexID & 1) != 0;\n");
+    // D3D doesn't include the base vertex in SV_VertexID
+    // See comment in UberShaderVertex for details
+    if (api_type == APIType::D3D)
+      out.Write("uint vertex_id = (gl_VertexID >> 2) + base_vertex;\n");
+    else
+      out.Write("uint vertex_id = gl_VertexID >> 2;\n");
+    out.Write("InputData i = input_buffer[vertex_id];\n"
+              "{}",
+              input_extract.GetBuffer());
+  }
+
   out.Write("VS_OUTPUT o;\n");
 
   // xfmem.numColorChans controls the number of color channels available to TEV, but we still need
@@ -403,6 +519,86 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const ShaderHostConfig& ho
     out.Write("}}\n");
   }
 
+  if (uid_data->vs_expand == VSExpand::Line)
+  {
+    out.Write("// Line expansion\n"
+              "uint other_id = vertex_id;\n"
+              "if (is_bottom) {{\n"
+              "  other_id -= 1;\n"
+              "}} else {{\n"
+              "  other_id += 1;\n"
+              "}}\n"
+              "InputData other = input_buffer[other_id];\n");
+    if (uid_data->position_has_3_elems)
+      out.Write("float4 other_pos = float4(other.pos0, other.pos1, other.pos2, 1.0f);\n");
+    else
+      out.Write("float4 other_pos = float4(other.pos0, other.pos1, 0.0f, 1.0f);\n");
+    if (uid_data->components & VB_HAS_POSMTXIDX)
+    {
+      out.Write("uint other_posidx = other.posmtx & 0xff;\n"
+                "float4 other_p0 = " I_TRANSFORMMATRICES "[other_posidx];\n"
+                "float4 other_p1 = " I_TRANSFORMMATRICES "[other_posidx + 1];\n"
+                "float4 other_p2 = " I_TRANSFORMMATRICES "[other_posidx + 2];\n"
+                "other_pos = float4(dot(other_p0, other_pos), dot(other_p1, other_pos), "
+                "dot(other_p2, other_pos), 1.0f);\n");
+    }
+    else
+    {
+      out.Write("other_pos = float4(dot(P0, other_pos), dot(P1, other_pos), dot(P2, other_pos), "
+                "1.0f);\n");
+    }
+    out.Write("other_pos = float4(dot(" I_PROJECTION "[0], other_pos), dot(" I_PROJECTION
+              "[1], other_pos), dot(" I_PROJECTION "[2], other_pos), dot(" I_PROJECTION
+              "[3], other_pos));\n"
+              "float expand_sign = is_right ? 1.0f : -1.0f;\n"
+              "float2 offset;\n"
+              "float2 to = abs(o.pos.xy / o.pos.w - other_pos.xy / other_pos.w);\n"
+              // FIXME: What does real hardware do when line is at a 45-degree angle?
+              // FIXME: Lines aren't drawn at the correct width. See Twilight Princess map.
+              "if (" I_LINEPTPARAMS ".y * to.y > " I_LINEPTPARAMS ".x * to.x) {{\n"
+              // Line is more tall. Extend geometry left and right.
+              // Lerp LineWidth/2 from [0..VpWidth] to [-1..1]
+              "  offset = float2(expand_sign * " I_LINEPTPARAMS ".z / " I_LINEPTPARAMS ".x, 0);\n"
+              "}} else {{\n"
+              // Line is more wide. Extend geometry up and down.
+              // Lerp LineWidth/2 from [0..VpHeight] to [1..-1]
+              "  offset = float2(0, expand_sign * " I_LINEPTPARAMS ".z / " I_LINEPTPARAMS ".y);\n"
+              "}}\n"
+              "\n"
+              "o.pos.xy += offset * o.pos.w;\n");
+    if (uid_data->numTexGens > 0)
+    {
+      out.Write("if ((" I_TEXOFFSET "[2] != 0) && is_right) {{\n"
+                "  float texOffset = 1.0 / float(" I_TEXOFFSET "[2]);\n");
+      for (u32 i = 0; i < uid_data->numTexGens; i++)
+      {
+        out.Write("  if (((" I_TEXOFFSET "[0] >> {}) & 0x1) != 0)\n", i);
+        out.Write("    o.tex{}.x += texOffset;\n", i);
+      }
+      out.Write("}}\n");
+    }
+  }
+  else if (uid_data->vs_expand == VSExpand::Point)
+  {
+    out.Write("// Point expansion\n"
+              "float2 expand_sign = float2(is_right ? 1.0f : -1.0f, is_bottom ? 1.0f : -1.0f);\n"
+              "float2 offset = expand_sign * " I_LINEPTPARAMS ".ww / " I_LINEPTPARAMS ".xy;\n"
+              "o.pos.xy += offset * o.pos.w;\n");
+    if (uid_data->numTexGens > 0)
+    {
+      out.Write("if (" I_TEXOFFSET "[3] != 0) {{\n"
+                "  float texOffsetMagnitude = 1.0f / float(" I_TEXOFFSET "[3]);\n"
+                "  float2 texOffset = float2(is_right ? texOffsetMagnitude : 0.0f, "
+                "is_bottom ? texOffsetMagnitude : 0.0f);");
+      for (u32 i = 0; i < uid_data->numTexGens; i++)
+      {
+        out.Write("  if (((" I_TEXOFFSET "[1] >> {}) & 0x1) != 0)\n", i);
+        out.Write("    o.tex{}.xy += texOffset;\n", i);
+      }
+      out.Write("}}\n");
+    }
+  }
+
   if (per_pixel_lighting)
   {
     // When per-pixel lighting is enabled, the vertex colors are passed through