From bd32c277bb1221c79dc59df5a34daa48da8645ab Mon Sep 17 00:00:00 2001
From: foreman <dl.constructicon@amd.com>
Date: Fri, 26 Aug 2016 01:47:40 -0400
Subject: [PATCH] P4 to Git Change 1307417 by lmoriche@lmoriche_opencl_dev on
 2016/08/26 01:38:31

	SWDEV-94610 - Setup the (hidden) arguments for global_offset [x,y,z].

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#10 edit
---
 rocclr/runtime/device/rocm/rocprogram.cpp |  4 ++--
 rocclr/runtime/device/rocm/rocvirtual.cpp | 22 ++++++++++++++++++++++
 2 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/rocclr/runtime/device/rocm/rocprogram.cpp b/rocclr/runtime/device/rocm/rocprogram.cpp
index 8955c1b21c..8155590efd 100644
--- a/rocclr/runtime/device/rocm/rocprogram.cpp
+++ b/rocclr/runtime/device/rocm/rocprogram.cpp
@@ -1024,7 +1024,7 @@ namespace roc {
             }
 
             // for OpenCL default hidden kernel arguments assuming there is no printf
-            size_t numHiddenKernelArgs = 0; // FIXME_lmoriche:3;
+            size_t numHiddenKernelArgs = 3; // FIXME_Wilkin
 
             // Fix the kernel name issue that causes string comparison does not work
             // due to an extra character at the end
@@ -1040,7 +1040,7 @@ namespace roc {
                 // TODO: remove the workaround
                 //   add 24 bytes for global offsets as workaround for LC reporting
                 //   excluded the hidden arguments
-                kernargSegmentByteSize /* FIXME_lmoriche:+24*/,
+                kernargSegmentByteSize + numHiddenKernelArgs * sizeof(size_t),
                 kernargSegmentAlignment,
                 numHiddenKernelArgs
             );
diff --git a/rocclr/runtime/device/rocm/rocvirtual.cpp b/rocclr/runtime/device/rocm/rocvirtual.cpp
index fd698954bb..2fea3d5f56 100644
--- a/rocclr/runtime/device/rocm/rocvirtual.cpp
+++ b/rocclr/runtime/device/rocm/rocvirtual.cpp
@@ -1498,6 +1498,7 @@ VirtualGPU::submitKernelInternal(
 
     address argPtr = argBuffer;
 
+#if !defined(WITH_LIGHTNING_COMPILER)
     // The HLC generates Kernenv arguments, first 3 are global offsets.
     const uint extraAargs = ((roc::Kernel*)devKernel)->extraArgumentsNum();
     for (uint j = 0; j < extraAargs; ++j) {
@@ -1510,6 +1511,7 @@ VirtualGPU::submitKernelInternal(
             addArg(&argPtr, &offset, sizeof(void*)); //Should be uint32_t for small model and uint64_t for large!
         }
     }
+#endif // defined(WITH_LIGHTNING_COMPILER)
 
     const amd::KernelSignature& signature = kernel.signature();
     const amd::KernelParameters& kernelParams = kernel.parameters();
@@ -1625,6 +1627,26 @@ VirtualGPU::submitKernelInternal(
         }
     }
 
+#if defined(WITH_LIGHTNING_COMPILER)
+    const uint extraAargs = ((roc::Kernel*)devKernel)->extraArgumentsNum();
+    for (uint j = 0; j < extraAargs; ++j) {
+        switch(j) {
+        case 3: { // Printf buffer
+            address bufferPtr = printfDbg()->dbgBuffer();
+            addArg(&argPtr, &bufferPtr, sizeof(void*));
+            break;
+        }
+        case 0: case 1: case 2: { // Global offsets
+            const size_t offset = j < sizes.dimensions() ? sizes.offset()[j] : 0;
+            addArg(&argPtr, &offset, sizeof(size_t));
+            break;
+        }
+        default:
+            assert(!"Unknown hidden argument index");
+        }
+    }
+#endif // defined(WITH_LIGHTNING_COMPILER)
+
     // Check there is no arguments' buffer overflow
     assert(argPtr <= argBuffer + gpuKernel.KernargSegmentByteSize());