diff --git a/src/gpu-compute/gpu_command_processor.cc b/src/gpu-compute/gpu_command_processor.cc
index dbb909f624..02b1bb174a 100644
--- a/src/gpu-compute/gpu_command_processor.cc
+++ b/src/gpu-compute/gpu_command_processor.cc
@@ -723,7 +723,6 @@ GPUCommandProcessor::sanityCheckAKC(AMDKernelCode *akc)
     warn_if(akc->kernarg_preload_spec_length ||
             akc->kernarg_preload_spec_offset,
             "Kernarg preload not implemented\n");
-    warn_if(akc->accum_offset, "ACC offset not implemented\n");
     warn_if(akc->tg_split, "TG split not implemented\n");
 }
 
diff --git a/src/gpu-compute/hsa_queue_entry.hh b/src/gpu-compute/hsa_queue_entry.hh
index a464e4882d..f015b091fc 100644
--- a/src/gpu-compute/hsa_queue_entry.hh
+++ b/src/gpu-compute/hsa_queue_entry.hh
@@ -122,6 +122,11 @@ class HSAQueueEntry
         }
 
         parseKernelCode(akc);
+
+        // Offset of a first AccVGPR in the unified register file.
+        // Granularity 4. Value 0-63. 0 - accum-offset = 4,
+        // 1 - accum-offset = 8, ..., 63 - accum-offset = 256.
+        _accumOffset = (akc->accum_offset + 1) * 4;
     }
 
     const GfxVersion&
@@ -394,6 +399,12 @@ class HSAQueueEntry
         assert(_outstandingWbs >= 0);
     }
 
+    unsigned
+    accumOffset() const
+    {
+        return _accumOffset;
+    }
+
   private:
     void
     parseKernelCode(AMDKernelCode *akc)
@@ -489,6 +500,8 @@ class HSAQueueEntry
 
     std::bitset<NumVectorInitFields> initialVgprState;
     std::bitset<NumScalarInitFields> initialSgprState;
+
+    unsigned _accumOffset;
 };
 
 } // namespace gem5
diff --git a/src/gpu-compute/wavefront.cc b/src/gpu-compute/wavefront.cc
index cb8b6220e7..98d882b20e 100644
--- a/src/gpu-compute/wavefront.cc
+++ b/src/gpu-compute/wavefront.cc
@@ -430,6 +430,9 @@ Wavefront::initRegState(HSAQueueEntry *task, int wgSizeInWorkItems)
         }
     }
 
+    // Save the offset to the first accumulation VGPR number from HSA task.
+    accumOffset = task->accumOffset();
+
     regInitIdx = 0;
 
     // VGPRs are initialized to the work item IDs for a given thread. There
diff --git a/src/gpu-compute/wavefront.hh b/src/gpu-compute/wavefront.hh
index 43ac3e9ffc..82035f7d47 100644
--- a/src/gpu-compute/wavefront.hh
+++ b/src/gpu-compute/wavefront.hh
@@ -131,6 +131,8 @@ class Wavefront : public SimObject
     uint32_t maxVgprs;
     // number of SGPRs required by WF
     uint32_t maxSgprs;
+    // first accumulation vgpr number
+    uint32_t accumOffset;
     void freeResources();
     GPUDynInstPtr nextInstr();
     void setStatus(status_e newStatus);