arch-gcn3: ensure that atomics follow HSA conventions

Add asserts to make sure atomics are following the HSA conventions
that atomics should be word aligned (i.e., can't be byte aligned)
and should not be misaligned such that a given lane's access
spans multiple cache lines.

Change-Id: Ia48758b9ed96764864234dc607f337e30e287d1c
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29941
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>

diff --git a/src/arch/gcn3/gpu_mem_helpers.hh b/src/arch/gcn3/gpu_mem_helpers.hh
index 40ca56561bef89a6354a26cb9fea151920461a49..562158d408427481a03c30a37eb69f23ab5e0f6f 100644 (file)
--- a/src/arch/gcn3/gpu_mem_helpers.hh
+++ b/src/arch/gcn3/gpu_mem_helpers.hh
@@ -80,6 +80,12 @@ initMemReqHelper(GPUDynInstPtr gpuDynInst, MemCmd mem_req_type,
             misaligned_acc = split_addr > vaddr;
 
             if (is_atomic) {
+                // make sure request is word aligned
+                assert((vaddr & 0x3) == 0);
+
+                // a given lane's atomic can't cross cache lines
+                assert(!misaligned_acc);
+
                 req = std::make_shared<Request>(vaddr, sizeof(T), 0,
                     gpuDynInst->computeUnit()->masterId(), 0,
                     gpuDynInst->wfDynId,