From a3633dea8d2e8ce07ea10a3f8998693562ad9344 Mon Sep 17 00:00:00 2001
From: Jake Turner <jake@evansturner.co.uk>
Date: Sun, 15 Dec 2024 15:16:05 +0000
Subject: [PATCH] Add D3D12 Shader Debug tests for SM6.4 instructions

dot4add_i8packed
dot4add_u8packed
dot2add
---
 .../demos/d3d12/d3d12_shader_debug_zoo.cpp    | 25 +++++++++++++++++++
 util/test/demos/d3d12/d3d12_test.cpp          |  8 +++---
 2 files changed, 30 insertions(+), 3 deletions(-)

diff --git a/util/test/demos/d3d12/d3d12_shader_debug_zoo.cpp b/util/test/demos/d3d12/d3d12_shader_debug_zoo.cpp
index d4d3d6764..ab988cb91 100644
--- a/util/test/demos/d3d12/d3d12_shader_debug_zoo.cpp
+++ b/util/test/demos/d3d12/d3d12_shader_debug_zoo.cpp
@@ -866,6 +866,31 @@ float4 main(v2f IN) : SV_Target0
     return inttexrwtest[uv];
   }
 #endif // #if SM_6_0 || SM_6_2 || SM_6_6
+#if SM_6_6
+  if(IN.tri == 94)
+  {
+    uint a = IN.tri - 94 + 0x01020304;
+    uint b = IN.tri - 94 + 0x05060708;
+    uint c = IN.tri - 94 + 0x090a0b0c;
+    uint res = dot4add_i8packed(a, b, c);
+    return float4(res & 0xFF, (res >> 8) & 0xFF, (res >> 16) & 0xFF, (res >> 24) & 0xFF);
+  }
+  if(IN.tri == 95)
+  {
+    uint a = IN.tri - 94 + 0x01020304;
+    uint b = IN.tri - 94 + 0x05060708;
+    uint c = IN.tri - 94 + 0x090a0b0c;
+    uint res = dot4add_u8packed(a, b, c);
+    return float4(res & 0xFF, (res >> 8) & 0xFF, (res >> 16) & 0xFF, (res >> 24) & 0xFF);
+  }
+  if(IN.tri == 96)
+  {
+    half2 a = half2(IN.tri - 96 + 0.1f, IN.tri - 96 + 0.2f);
+    half2 b = half2(IN.tri - 96 + 0.2f, IN.tri - 96 + 0.3f);
+    float c = IN.tri - 96 + 0.3f;
+    return dot2add(a, b, c);
+  }
+#endif // #if SM_6_6
 
   return float4(0.4f, 0.4f, 0.4f, 0.4f);
 }
diff --git a/util/test/demos/d3d12/d3d12_test.cpp b/util/test/demos/d3d12/d3d12_test.cpp
index a13e3f85d..1741cea0b 100644
--- a/util/test/demos/d3d12/d3d12_test.cpp
+++ b/util/test/demos/d3d12/d3d12_test.cpp
@@ -1335,8 +1335,8 @@ ID3DBlobPtr D3D12GraphicsTest::Compile(std::string src, std::string entry, std::
                                        uint32_t compileOptions)
 {
   ID3DBlobPtr blob = NULL;
-  bool skipoptimise = (compileOptions && CompileOptionFlags::SkipOptimise);
-  bool enable16BitTypes = (compileOptions && CompileOptionFlags::Enable16BitTypes);
+  bool skipoptimise = ((compileOptions & CompileOptionFlags::SkipOptimise) != 0);
+  bool enable16BitTypes = ((compileOptions & CompileOptionFlags::Enable16BitTypes) != 0);
 
   if(profile[3] >= '6')
   {
@@ -1390,10 +1390,12 @@ ID3DBlobPtr D3D12GraphicsTest::Compile(std::string src, std::string entry, std::
       argStorage.push_back(L"-O1");
     }
     argStorage.push_back(L"-Zi");
-    argStorage.push_back(L"-Qembed_debug");
     if(enable16BitTypes)
       argStorage.push_back(L"-enable-16bit-types");
 
+    // Must be the final option
+    argStorage.push_back(L"-Qembed_debug");
+
     for(size_t i = 0; i < argStorage.size(); i++)
       args[0].push_back(argStorage[i].c_str());