SPIRV: Fix performance issue when handling large arrays. (#4064)

* SPIRV: Fix performance issue when handling large arrays.

* Add test for packing.

* Fix clang.

1 files changed, 32 insertions, 0 deletions

diff --git a/tests/spirv/large-struct.slang b/tests/spirv/large-struct.slang
new file mode 100644
index 000000000..7738a5fcf
--- /dev/null
+++ b/tests/spirv/large-struct.slang
@@ -0,0 +1,32 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -emit-spirv-directly -profile glsl_460
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX(filecheck-buffer=BUF):-vk -compute -output-using-type
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX(filecheck-buffer=BUF):-d3d12 -compute -output-using-type
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX(filecheck-buffer=BUF):-cpu -compute -output-using-type
+
+// Check that when generating spirv directly, we use a loop
+// to copy large arrays in input data out into a local variable, instead of emitting
+// unrolled code that reads each element of the array individually.
+
+struct WorkData
+{
+    float A[2*2];
+    float B[1024];
+    
+    float Foo(uint i) { return A[i] * B[i]; }
+};
+
+//TEST_INPUT:set input = new WorkData{[1.0, 2.0, 3.0, 4.0], [10.0, 20.0, 30.0, 40.0]}
+ConstantBuffer<WorkData> input;
+
+//TEST_INPUT:set resultBuffer = out ubuffer(data=[0 0 0 0], stride=4)
+RWStructuredBuffer<float> resultBuffer;
+
+// CHECK: OpLoopMerge
+
+[numthreads(2, 1, 1)]
+void computeMain(uint3 tid: SV_DispatchThreadID)
+{
+    // BUF: 10.0
+    // BUF: 40.0
+    resultBuffer[tid.x] = input.Foo(tid.x);
+}
\ No newline at end of file