[D3D12] Update some comments about numeric formats

src/xenia/gpu/dxbc_shader_translator_memexport.cc

@@ -447,8 +447,6 @@ void DxbcShaderTranslator::ExportToMemory() {
     uint32_t format_param_temp = PushSystemTemp();
 
     // Denormalize, clamp and convert to integer.
-    // TODO(Triang3l): Really clamp? It's GPUSURFACENUMBER_UREPEAT, not clamp,
-    // but need to verify since not clamping doesn't look very safe.
     // A lot of the code is similar for both signed and unsigned. Start by
     // checking the signedness.
     shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_IF) |

src/xenia/gpu/dxbc_shader_translator_om.cc

@@ -729,10 +729,10 @@ void DxbcShaderTranslator::CompletePixelShader_DepthTo24Bit(
   ++stat_.instruction_count;
   ++stat_.float_instruction_count;
 
-  // Round to the nearest integer. This is the correct way of rounding, rounding
+  // Round to the nearest even integer. This seems to be the correct way:
-  // towards zero gives 0xFF instead of 0x100 in clear shaders in, for instance,
+  // rounding towards zero gives 0xFF instead of 0x100 in clear shaders in, for
-  // Halo 3.
+  // instance, Halo 3, but other clear shaders in it are also broken if 0.5 is
-  // https://docs.microsoft.com/en-us/windows/desktop/direct3d10/d3d10-graphics-programming-guide-resources-data-conversion
+  // added before ftou instead of round_ne.
   shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ROUND_NE) |
                          ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
   shader_code_.push_back(
@@ -3646,8 +3646,9 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV_PackColor(
   ++stat_.instruction_count;
   ++stat_.float_instruction_count;
 
-  // Convert to fixed-point, rounding to the nearest integer.
+  // Convert to fixed-point, rounding towards the nearest even integer.
-  // https://docs.microsoft.com/en-us/windows/desktop/direct3d10/d3d10-graphics-programming-guide-resources-data-conversion
+  // Rounding towards the nearest (adding +-0.5 before truncating) is giving
+  // incorrect results for depth, so better to use round_ne here too.
   uint32_t fixed_temp = PushSystemTemp();
   shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ROUND_NE) |
                          ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));

src/xenia/gpu/xenos.h

@@ -156,7 +156,8 @@ enum class IndexFormat : uint32_t {
   kInt32,
 };
 
-// GPUSURFACENUMBER from a game .pdb.
+// GPUSURFACENUMBER from a game .pdb. "Repeat" means repeating fraction, it's
+// what ATI calls normalized.
 enum class SurfaceNumFormat : uint32_t {
   kUnsignedRepeat = 0,
   kSignedRepeat = 1,