From d5f37b53aecc9edfade6fff26313931d173535be Mon Sep 17 00:00:00 2001
From: foreman <dl.constructicon@amd.com>
Date: Tue, 10 Jan 2017 17:46:11 -0500
Subject: [PATCH] P4 to Git Change 1360479 by rerrabol@RocmOnLC on 2017/01/10
 17:32:55

	SWDEV-108854 - Support initialization of SGPR and VGPR usage from kernel header object. Changes to fix issue reported by SWDEV-108854

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#33 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#18 edit


[ROCm/clr commit: a3c0b6525f7f1cd8efa461449465c62e6f38adfa]
---
 .../rocclr/runtime/device/rocm/rocdevice.cpp   |  4 ++++
 .../rocclr/runtime/device/rocm/rockernel.cpp   | 18 ++++++++++++++++--
 2 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp b/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp
index a11471fb4d..0d5deb8788 100644
--- a/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp
+++ b/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp
@@ -285,6 +285,10 @@ bool NullDevice::init() {
     if (!initCompiler(offlineDevice_)){
         return false;
     }
+
+    // Return without initializing offline device list
+    return true;
+
 #if !defined(WITH_LIGHTNING_COMPILER)
     //If there is an HSA enabled device online then skip any offline device
     std::vector<Device*> devices;
diff --git a/projects/clr/rocclr/runtime/device/rocm/rockernel.cpp b/projects/clr/rocclr/runtime/device/rocm/rockernel.cpp
index d5fe95568c..a92dbb0807 100644
--- a/projects/clr/rocclr/runtime/device/rocm/rockernel.cpp
+++ b/projects/clr/rocclr/runtime/device/rocm/rockernel.cpp
@@ -865,9 +865,23 @@ bool Kernel::init()
     workGroupInfo_.privateMemSize_ = workitemPrivateSegmentByteSize_;
     workGroupInfo_.usedLDSSize_ = workgroupGroupSegmentByteSize_;
     workGroupInfo_.preferredSizeMultiple_ = wavefront_size;
-    workGroupInfo_.usedSGPRs_ = 0;
+
+    // Query kernel header object to initialize the number of
+    // SGPR's and VGPR's used by the kernel
+    const void* kernelHostPtr = nullptr;
+    if (Device::loaderQueryHostAddress(
+            reinterpret_cast<const void*>(kernelCodeHandle_), &kernelHostPtr
+            ) == HSA_STATUS_SUCCESS) {
+        auto akc = reinterpret_cast<const amd_kernel_code_t*>(kernelHostPtr);
+        workGroupInfo_.usedSGPRs_ = akc->wavefront_sgpr_count;
+        workGroupInfo_.usedVGPRs_ = akc->workitem_vgpr_count;
+    }
+    else {
+        workGroupInfo_.usedSGPRs_ = 0;
+        workGroupInfo_.usedVGPRs_ = 0;
+    }
+
     workGroupInfo_.usedStackSize_ = 0;
-    workGroupInfo_.usedVGPRs_ = 0;
     workGroupInfo_.wavefrontPerSIMD_ =
         program_->dev().info().maxWorkItemSizes_[0] / wavefront_size;
     workGroupInfo_.wavefrontSize_ = wavefront_size;