Allow CPP/CUDA/Metal to lower/legalize buffer-elements to support column_major/row_major. (#4653)

* Allow CPP/CUDA/Metal to legalize their buffer-elements.

Fixes: #4537

Changes:
1. Matrix inputs require legalization (pack/unpack) to ensure consistent row_major/column_major throughout entire shader, the following enabled legalization pass fixes this.
2. Added missing CUDA intrinsic so CUDA can run more tests.
3. Added a memory packing test since this still fails for cpp/cuda/metal (due to having no memory packing enforcement).

* change memory packing tests to run for targets without packing

---------

Co-authored-by: Yong He <yonghe@outlook.com>

3 files changed, 6 insertions, 10 deletions

diff --git a/source/slang/slang-compiler.h b/source/slang/slang-compiler.h
index 2409cedfb..7fc43d778 100755
--- a/source/slang/slang-compiler.h
+++ b/source/slang/slang-compiler.h
@@ -1769,6 +1769,8 @@ namespace Slang
     /// Are we generating code for a CUDA API (CUDA / OptiX)?
     bool isCUDATarget(TargetRequest* targetReq);
 
+    // Are we generating code for a CPU target
+    bool isCPUTarget(TargetRequest* targetReq);
 
         /// A request to generate output in some target format.
     class TargetRequest : public RefObject
diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index 679d8ce88..678b4137a 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -1267,15 +1267,7 @@ Result linkAndOptimizeIR(
     if (requiredLoweringPassSet.meshOutput)
         legalizeMeshOutputTypes(irModule);
 
-    if (options.shouldLegalizeExistentialAndResourceTypes)
-    {
-        if (!isMetalTarget(targetRequest))
-        {
-            // We need to lower any types used in a buffer resource (e.g. ContantBuffer or StructuredBuffer) into
-            // a simple storage type that has target independent layout based on the kind of buffer resource.
-            lowerBufferElementTypeToStorageType(targetProgram, irModule);
-        }
-    }
+    lowerBufferElementTypeToStorageType(targetProgram, irModule);
 
     // Rewrite functions that return arrays to return them via `out` parameter,
     // since our target languages doesn't allow returning arrays.
diff --git a/source/slang/slang-ir-lower-buffer-element-type.cpp b/source/slang/slang-ir-lower-buffer-element-type.cpp
index 981e29697..d042aae43 100644
--- a/source/slang/slang-ir-lower-buffer-element-type.cpp
+++ b/source/slang/slang-ir-lower-buffer-element-type.cpp
@@ -877,7 +877,9 @@ namespace Slang
     void lowerBufferElementTypeToStorageType(TargetProgram* target, IRModule* module, bool lowerBufferPointer)
     {
         SlangMatrixLayoutMode defaultMatrixMode = (SlangMatrixLayoutMode)target->getOptionSet().getMatrixLayoutMode();
-        if (defaultMatrixMode == SLANG_MATRIX_LAYOUT_MODE_UNKNOWN)
+        if ((isCPUTarget(target->getTargetReq()) || isCUDATarget(target->getTargetReq()) || isMetalTarget(target->getTargetReq())))
+            defaultMatrixMode = SLANG_MATRIX_LAYOUT_ROW_MAJOR;
+        else if (defaultMatrixMode == SLANG_MATRIX_LAYOUT_MODE_UNKNOWN)
             defaultMatrixMode = SLANG_MATRIX_LAYOUT_ROW_MAJOR;
         LoweredElementTypeContext context(target, lowerBufferPointer, defaultMatrixMode);
         context.processModule(module);