Support per field matrix layout (#3101)

* Support per field matrix layout

* Fix warnings.

* Fix.

* Fix tests.

* Fix spiv gen.

* Fix.

* More test fixes.

* Fix.

* Run only GPU tests on self-hosted servers.

* Remove -use-glsl-matrix-layout-modifier.

* Fix.

---------

Co-authored-by: Yong He <yhe@nvidia.com>

1 files changed, 13 insertions, 4 deletions

diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index 343c18916..4b95ca9c8 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -49,6 +49,7 @@
 #include "slang-ir-specialize-arrays.h"
 #include "slang-ir-specialize-buffer-load-arg.h"
 #include "slang-ir-specialize-resources.h"
+#include "slang-ir-specialize-matrix-layout.h"
 #include "slang-ir-ssa.h"
 #include "slang-ir-ssa-simplification.h"
 #include "slang-ir-strip-cached-dict.h"
@@ -59,6 +60,7 @@
 #include "slang-ir-liveness.h"
 #include "slang-ir-glsl-liveness.h"
 #include "slang-ir-legalize-uniform-buffer-load.h"
+#include "slang-ir-lower-buffer-element-type.h"
 #include "slang-ir-string-hash.h"
 #include "slang-ir-simplify-for-emit.h"
 #include "slang-ir-pytorch-cpp-binding.h"
@@ -353,6 +355,9 @@ Result linkAndOptimizeIR(
 
     simplifyIR(irModule, sink);
 
+    // Fill in default matrix layout into matrix types that left layout unspecified.
+    specializeMatrixLayout(codeGenContext->getTargetReq(), irModule);
+
     // It's important that this takes place before defunctionalization as we
     // want to be able to easily discover the cooperate and fallback funcitons
     // being passed to saturated_cooperation
@@ -588,10 +593,6 @@ Result linkAndOptimizeIR(
     }
     eliminateDeadCode(irModule);
 
-    // Rewrite functions that return arrays to return them via `out` parameter,
-    // since our target languages doesn't allow returning arrays.
-    legalizeArrayReturnType(irModule);
-
 #if 0
     dumpIRIfEnabled(codeGenContext, irModule, "AFTER RESOURCE SPECIALIZATION");
 #endif
@@ -857,6 +858,14 @@ Result linkAndOptimizeIR(
     // arrays that the emitters can deal with.
     legalizeMeshOutputTypes(irModule);
 
+    // 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.
+    lowerBufferElementTypeToStorageType(targetRequest, irModule);
+
+    // Rewrite functions that return arrays to return them via `out` parameter,
+    // since our target languages doesn't allow returning arrays.
+    legalizeArrayReturnType(irModule);
+
     if (isKhronosTarget(targetRequest) || target == CodeGenTarget::HLSL)
     {
         legalizeUniformBufferLoad(irModule);