For benchmarking

ucb-bar · May 6, 2020 · 4298984 · 4298984
1 parent 503c605
commit 4298984
Show file tree

Hide file tree

Showing 4 changed files with 17 additions and 6 deletions.
diff --git a/onnxruntime/core/mlas/lib/systolic/systolic.cpp b/onnxruntime/core/mlas/lib/systolic/systolic.cpp
@@ -29,14 +29,14 @@ inline tiled_matmul_type_t get_accelerator_mode(int mode) {
 
 void SystolicMultiplyi8i8_i8(char accelerator_mode, bool relu, int dimI, int dimJ, int dimK,
                              const elem_t* in1, const elem_t* in2, elem_t* out, int divisor, __attribute__((unused)) float real_multiplier, const int32_t* bias) {
-  printf("Called into systolic matmul!\n");
+  //printf("Called into systolic matmul!\n");
   bool isPowerOf2 = divisor && !(divisor & (divisor - 1));
   if (!isPowerOf2) {
     throw std::runtime_error("Divisor passed to systolic matmul must be power of 2");
   }
   int shift = sizeof(int) * 8 - __builtin_clz(divisor) - 1;
 
-  printf("Using accelerated matmul with dimensions (%d, %d, %d)\n", dimI, dimJ, dimK);
+  //printf("Using accelerated matmul with dimensions (%d, %d, %d)\n", dimI, dimJ, dimK);
   tiled_matmul_auto(dimI, dimJ, dimK, in1, in2, bias, out, /*activation= */ relu,
                     shift, /*relu6_shift= */ 0, /* repeating_bias= */ 0, get_accelerator_mode(accelerator_mode));
 }
@@ -48,14 +48,14 @@ void SystolicMultiplyi8i8_i8(char accelerator_mode, bool relu,
                              elem_t* out, int strideOut,
                              int divisor, __attribute__((unused)) float real_multiplier,
                              const int32_t* bias, int strideBias, bool repeating_bias) {
-  printf("Called into systolic matmul!\n");
+  //printf("Called into systolic matmul!\n");
   bool isPowerOf2 = divisor && !(divisor & (divisor - 1));
   if (!isPowerOf2) {
     throw std::runtime_error("Divisor passed to systolic matmul must be power of 2");
   }
   int shift = sizeof(int) * 8 - __builtin_clz(divisor) - 1;
 
-  printf("Using accelerated matmul with dimensions (%d, %d, %d)\n", dimI, dimJ, dimK);
+  //printf("Using accelerated matmul with dimensions (%d, %d, %d)\n", dimI, dimJ, dimK);
   tiled_matmul_auto(dimI, dimJ, dimK,
                     strideIn1, strideIn2, strideBias, strideOut,
                     in1, in2, bias, out, /*activation= */ relu,

diff --git a/onnxruntime/core/mlas/lib/systolic/systolic_include.h b/onnxruntime/core/mlas/lib/systolic/systolic_include.h
@@ -719,4 +719,9 @@ void tiled_matmul_auto(size_t dim_I, size_t dim_J, size_t dim_K,
   tiled_matmul_auto(dim_I, dim_J, dim_K, dim_K, dim_J, dim_J, dim_J, A, B, D, C, act, shift, relu6_shift, repeating_bias, tiled_matmul_type);
 }
 
+__attribute__((constructor))
+void cleargemmini() {
+  gemmini_flush(0);
+}
+
 #endif  // SRC_MAIN_C_SYSTOLIC_H
diff --git a/onnxruntime/core/providers/systolic/nn/qlinearconv.cc b/onnxruntime/core/providers/systolic/nn/qlinearconv.cc
@@ -181,7 +181,7 @@ Status QLinearConv<StorageOrder::NHWC>::Compute(OpKernelContext* context) const
     auto col_data = alloc->Alloc(SafeInt<size_t>(sizeof(int8_t)) * col_buffer_size);
     col_buffer = BufferUniquePtr(col_data, BufferDeleter(alloc));
   } else {
-    printf("1x1 case!\n");
+    //printf("1x1 case!\n");
   }
 
   auto* col_buffer_data = static_cast<int8_t*>(col_buffer.get());
@@ -405,7 +405,7 @@ Status QLinearConv<StorageOrder::NCHW>::Compute(OpKernelContext* context) const
       col_buffer_shape.insert(col_buffer_shape.end(), output_dims.begin(), output_dims.end());
     }
   } else {
-    printf("1x1 case!\n");
+    //printf("1x1 case!\n");
   }
 
   auto* col_buffer_data = static_cast<int8_t*>(col_buffer.get());

diff --git a/systolic_runner/imagenet_runner/src/runner.cpp b/systolic_runner/imagenet_runner/src/runner.cpp
@@ -17,6 +17,7 @@
 #include "tensor_helper.h"
 #include "cmd_args.h"
 #include "labels.h"
+#include <sys/mman.h>
 
 const char* imagenet_labels[1000] = IMAGENET_LABELS;
 
@@ -28,6 +29,11 @@ unsigned long long read_cycles()
 }
 
 int main(int argc, char* argv[]) {
+    if (mlockall(MCL_CURRENT | MCL_FUTURE) != 0) {
+      perror("mlockall failed");
+      exit(1);
+    }
+
   setbuf(stdout, NULL);
   printf("Loaded runner program\n");