Improved cache maintenance

Invalidating and cleaning buffers exchanged between Core and remote CPU.

Change-Id: Icd9ce6c916422a6bbcdd42e31651a622240d0ce4

3 files changed, 55 insertions, 1 deletions

diff --git a/applications/inference_process/include/inference_process.hpp b/applications/inference_process/include/inference_process.hpp
index 53b9331..ec682d1 100644
--- a/applications/inference_process/include/inference_process.hpp
+++ b/applications/inference_process/include/inference_process.hpp
@@ -29,6 +29,9 @@ struct DataPtr {
     size_t size;
 
     DataPtr(void *data = nullptr, size_t size = 0);
+
+    void invalidate();
+    void clean();
 };
 
 struct InferenceJob {
@@ -46,6 +49,9 @@ struct InferenceJob {
                  const std::vector<DataPtr> &output,
                  const std::vector<DataPtr> &expectedOutput,
                  size_t numBytesToPrint);
+
+    void invalidate();
+    void clean();
 };
 
 class InferenceProcess {
diff --git a/applications/inference_process/src/inference_process.cc b/applications/inference_process/src/inference_process.cc
index 7743f8c..61db73d 100644
--- a/applications/inference_process/src/inference_process.cc
+++ b/applications/inference_process/src/inference_process.cc
@@ -87,6 +87,20 @@ bool copyOutput(const TfLiteTensor &src, InferenceProcess::DataPtr &dst) {
 namespace InferenceProcess {
 DataPtr::DataPtr(void *_data, size_t _size) : data(_data), size(_size) {}
 
+void DataPtr::invalidate() {
+#if defined(__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U)
+    printf("Invalidate. data=%p, size=%zu\n", data, size);
+    SCB_InvalidateDCache_by_Addr(reinterpret_cast<uint32_t *>(data), size);
+#endif
+}
+
+void DataPtr::clean() {
+#if defined(__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U)
+    printf("Clean. data=%p, size=%zu\n", data, size);
+    SCB_CleanDCache_by_Addr(reinterpret_cast<uint32_t *>(data), size);
+#endif
+}
+
 InferenceJob::InferenceJob() : numBytesToPrint(0) {}
 
 InferenceJob::InferenceJob(const string &_name,
@@ -99,6 +113,38 @@ InferenceJob::InferenceJob(const string &_name,
     networkModel(_networkModel), input(_input), output(_output), expectedOutput(_expectedOutput),
     numBytesToPrint(_numBytesToPrint) {}
 
+void InferenceJob::invalidate() {
+    networkModel.invalidate();
+
+    for (auto &it : input) {
+        it.invalidate();
+    }
+
+    for (auto &it : output) {
+        it.invalidate();
+    }
+
+    for (auto &it : expectedOutput) {
+        it.invalidate();
+    }
+}
+
+void InferenceJob::clean() {
+    networkModel.clean();
+
+    for (auto &it : input) {
+        it.clean();
+    }
+
+    for (auto &it : output) {
+        it.clean();
+    }
+
+    for (auto &it : expectedOutput) {
+        it.clean();
+    }
+}
+
 InferenceProcess::InferenceProcess() : lock(0) {}
 
 // NOTE: Adding code for get_lock & free_lock with some corrections from
diff --git a/applications/message_process/src/message_process.cc b/applications/message_process/src/message_process.cc
index 200d92b..c890399 100644
--- a/applications/message_process/src/message_process.cc
+++ b/applications/message_process/src/message_process.cc
@@ -201,7 +201,7 @@ bool MessageProcess::handleMessage() {
             return false;
         }
 
-        printf("InferenceReq. user_arg=0x%" PRIx64 ", network={0x%" PRIu32 ", %" PRIu32 "}",
+        printf("InferenceReq. user_arg=0x%" PRIx64 ", network={0x%" PRIx32 ", %" PRIu32 "}",
                req.user_arg,
                req.network.ptr,
                req.network.size);
@@ -241,8 +241,10 @@ bool MessageProcess::handleMessage() {
         vector<DataPtr> expectedOutput;
 
         InferenceJob job("job", networkModel, ifm, ofm, expectedOutput, -1);
+        job.invalidate();
 
         bool failed = inferenceProcess.runJob(job);
+        job.clean();
 
         sendInferenceRsp(req.user_arg, job.output, failed);
         break;