Tweaking.

src/poly/logging.cc

@@ -62,14 +62,15 @@ void format_log_line(char* buffer, size_t buffer_count, const char* file_path,
   }
 }
 
+thread_local char log_buffer[2048];
+
 void log_line(const char* file_path, const uint32_t line_number,
               const char level_char, const char* fmt, ...) {
   // SCOPE_profile_cpu_i("emu", "log_line");
 
-  char buffer[2048];
   va_list args;
   va_start(args, fmt);
-  format_log_line(buffer, poly::countof(buffer), file_path, line_number,
+  format_log_line(log_buffer, poly::countof(log_buffer), file_path, line_number,
                   level_char, fmt, args);
   va_end(args);
 
@@ -77,9 +78,9 @@ void log_line(const char* file_path, const uint32_t line_number,
     log_lock.lock();
   }
 #if 0  // defined(OutputDebugString)
-  OutputDebugStringA(buffer);
+  OutputDebugStringA(log_buffer);
 #else
-  fprintf(stdout, "%s", buffer);
+  fprintf(stdout, "%s", log_buffer);
   fflush(stdout);
 #endif  // OutputDebugString
   if (!FLAGS_fast_stdout) {

src/xenia/gpu/gl4/command_processor.cc

@@ -2167,7 +2167,7 @@ bool CommandProcessor::IssueCopy(DrawCommand* draw_command) {
   auto copy_dest_format =
       static_cast<ColorFormat>((copy_dest_info >> 7) & 0x3F);
   uint32_t copy_dest_number = (copy_dest_info >> 13) & 0x7;
-  assert_true(copy_dest_number == 0);
+  // assert_true(copy_dest_number == 0); // ?
   uint32_t copy_dest_bias = (copy_dest_info >> 16) & 0x3F;
   assert_true(copy_dest_bias == 0);
   uint32_t copy_dest_swap = (copy_dest_info >> 25) & 0x1;
@@ -2233,6 +2233,14 @@ bool CommandProcessor::IssueCopy(DrawCommand* draw_command) {
       read_format = copy_dest_swap ? GL_BGRA : GL_RGBA;
       read_type = GL_UNSIGNED_BYTE;
       break;
+    case ColorFormat::k_16_16_16_16_FLOAT:
+      read_format = GL_RGBA;
+      read_type = GL_HALF_FLOAT;
+      break;
+    case ColorFormat::k_32_FLOAT:
+      read_format = GL_R32F;
+      read_type = GL_FLOAT;
+      break;
     default:
       assert_unhandled_case(copy_dest_format);
       return false;
@@ -2251,7 +2259,8 @@ bool CommandProcessor::IssueCopy(DrawCommand* draw_command) {
       glPixelStorei(GL_PACK_SWAP_BYTES, GL_TRUE);
       break;
     default:
-      assert_unhandled_case(copy_dest_endian);
+      //assert_unhandled_case(copy_dest_endian);
+      glPixelStorei(GL_PACK_SWAP_BYTES, GL_TRUE);
       return false;
   }
 
@@ -2315,13 +2324,13 @@ bool CommandProcessor::IssueCopy(DrawCommand* draw_command) {
   if (depth_clear_enabled) {
     // Clear the current depth buffer.
     // TODO(benvanik): verify format.
-    union {
+    GLfloat depth = {(copy_depth_clear & 0xFFFFFF00) / float(0xFFFFFF00)};
-      uint32_t uint_value;
-      GLfloat float_value;
-    } depth = {copy_depth_clear & 0xFFFFFF00};
     GLint stencil = copy_depth_clear & 0xFF;
-    glClearNamedFramebufferfi(source_framebuffer->framebuffer, GL_DEPTH_STENCIL,
+    // HACK: this should work, but throws INVALID_ENUM on nvidia drivers.
-                              depth.float_value, stencil);
+    //glClearNamedFramebufferfi(source_framebuffer->framebuffer, GL_DEPTH_STENCIL,
+    //                          depth, stencil);
+    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, source_framebuffer->framebuffer);
+    glClearBufferfi(GL_DEPTH_STENCIL, 0, depth, stencil);
   }
 
   return true;
@@ -2361,6 +2370,29 @@ GLuint CommandProcessor::GetColorRenderTarget(uint32_t pitch,
     case ColorRenderTargetFormat::k_8_8_8_8_GAMMA:
       internal_format = GL_RGBA8;
       break;
+    case ColorRenderTargetFormat::k_2_10_10_10:
+    case ColorRenderTargetFormat::k_2_10_10_10_unknown:
+      internal_format = GL_RGB10_A2UI;
+      break;
+    case ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
+    case ColorRenderTargetFormat::k_2_10_10_10_FLOAT_unknown:
+      internal_format = GL_RGB10_A2;
+      break;
+    case ColorRenderTargetFormat::k_16_16:
+      internal_format = GL_RG16;
+      break;
+    case ColorRenderTargetFormat::k_16_16_FLOAT:
+      internal_format = GL_RG16F;
+      break;
+    case ColorRenderTargetFormat::k_16_16_16_16:
+      internal_format = GL_RGBA16;
+      break;
+    case ColorRenderTargetFormat::k_16_16_16_16_FLOAT:
+      internal_format = GL_RGBA16F;
+      break;
+    case ColorRenderTargetFormat::k_32_FLOAT:
+      internal_format = GL_R32F;
+      break;
     default:
       assert_unhandled_case(format);
       return 0;
@@ -2399,7 +2431,9 @@ GLuint CommandProcessor::GetDepthRenderTarget(uint32_t pitch,
       internal_format = GL_DEPTH24_STENCIL8;
       break;
     case DepthRenderTargetFormat::kD24FS8:
-    // TODO(benvanik): not supported in GL?
+      // TODO(benvanik): not supported in GL?
+      internal_format = GL_DEPTH24_STENCIL8;
+      break;
     default:
       assert_unhandled_case(format);
       return 0;

src/xenia/gpu/gl4/gl4_shader.cc

@@ -127,11 +127,7 @@ bool GL4Shader::PreparePixelShader(
       "layout(location = 0) in VertexData vtx;\n"
       "layout(location = 0) out vec4 oC[4];\n"
       "void processFragment();\n"
-      "void main() {\n"
+      "void main() {\n" +
-      "  for (int i = 0; i < oC.length(); ++i) {\n"
-      "    oC[i] = vec4(1.0, 0.0, 0.0, 1.0);\n"
-      "  }\n" +
-      (program_cntl.ps_export_depth ? "  gl_FragDepth = 0.0;\n" : "") +
       "  processFragment();\n"
       "}\n";
 

src/xenia/gpu/gl4/gl4_shader_translator.cc

@@ -69,6 +69,7 @@ std::string GL4ShaderTranslator::TranslateVertexShader(
   Reset(vertex_shader);
 
   // Normal shaders only, for now.
+  // TODO(benvanik): transform feedback/memexport.
   assert_true(program_cntl.vs_export_mode == 0);
 
   // Add vertex shader input.
@@ -199,7 +200,8 @@ void GL4ShaderTranslator::AppendDestRegName(uint32_t num, uint32_t dst_exp) {
             Append("gl_Position");
             break;
           case 63:
-            Append("gl_PointSize");
+            // Write to t, as we need to splice just x out of it.
+            Append("t");
             break;
           default:
             // Varying.
@@ -242,6 +244,8 @@ void GL4ShaderTranslator::AppendDestRegPost(uint32_t num, uint32_t mask,
   if (num == 61) {
     // gl_FragDepth handling to just get x from the temp result.
     Append("  gl_FragDepth = t.x;\n");
+  } else if (num == 63) {
+    Append("  gl_PointSize = t.x;\n");
   } else if (mask != 0xF) {
     // Masking.
     Append("  ");

src/xenia/gpu/gl4/texture_cache.cc

@@ -281,7 +281,7 @@ bool TextureCache::UploadTexture2D(GLuint texture, void* host_base,
   switch (texture_info.format) {
     case TextureFormat::k_8:
       internal_format = GL_R8;
-      format = GL_R;
+      format = GL_RED;
       type = GL_UNSIGNED_BYTE;
       break;
     case TextureFormat::k_1_5_5_5:
@@ -341,7 +341,7 @@ bool TextureCache::UploadTexture2D(GLuint texture, void* host_base,
       break;
     case TextureFormat::k_32_FLOAT:
       internal_format = GL_R32F;
-      format = GL_R;
+      format = GL_RED;
       type = GL_FLOAT;
       break;
     case TextureFormat::k_32_32_FLOAT:

src/xenia/gpu/shader.cc

@@ -153,8 +153,6 @@ void Shader::GatherExec(const instr_cf_exec_t* cf) {
 }
 
 void Shader::GatherVertexFetch(const instr_fetch_vtx_t* vtx) {
-  assert_true(shader_type_ == ShaderType::kVertex);
-
   // dst_reg/dst_swiz
   // src_reg/src_swiz
   // format = a2xx_sq_surfaceformat
@@ -166,6 +164,10 @@ void Shader::GatherVertexFetch(const instr_fetch_vtx_t* vtx) {
   // num_format_all ? integer : fraction
   // exp_adjust_all - [-32,31] - (2^exp_adjust_all)*fetch - 0 = default
 
+  if (!vtx->must_be_one) {
+    return;
+  }
+
   // Sometimes games have fetches that just produce constants. We can
   // ignore those.
   uint32_t dst_swiz = vtx->dst_swiz;

src/xenia/gpu/xenos.h

@@ -81,7 +81,15 @@ enum class MsaaSamples : uint32_t {
 enum class ColorRenderTargetFormat : uint32_t {
   k_8_8_8_8 = 0,        // D3DFMT_A8R8G8B8 (or ABGR?)
   k_8_8_8_8_GAMMA = 1,  // D3DFMT_A8R8G8B8 with gamma correction
-                        // ...
+  k_2_10_10_10 = 2,
+  k_2_10_10_10_FLOAT = 3,
+  k_16_16 = 4,
+  k_16_16_16_16 = 5,
+  k_16_16_FLOAT = 6,
+  k_16_16_16_16_FLOAT = 7,
+  k_2_10_10_10_unknown = 10,
+  k_2_10_10_10_FLOAT_unknown = 12,
+  k_32_FLOAT = 14,
 };
 
 enum class DepthRenderTargetFormat : uint32_t {