Add functionality to memcpy-perf benchmark

Add delay option to purposely hurt bandwidth
Convert several hardcoded parameters into command-line options (keeping
defaults)
    - start/end sizes
    - number of samples to be collected
    - dummy mode
Add dummy mode, approximate replication of CPU load without hitting
memory. Possibly useful for comparison. Uses L2 instead of memory.
All added command-line options are optional arguments, keeps existing
behavior

Added new calculation in benchmark output that shows time spent doing
the memory operation vs. time spent during the delay (if present)

Change-Id: I11d1f600afa23f49b922417babe6c15cfa058b5d

1 files changed, 103 insertions, 18 deletions

diff --git a/memcpy-perf/memcpy-perf.cpp b/memcpy-perf/memcpy-perf.cpp
index 20d060b8..2dfd900d 100644
--- a/memcpy-perf/memcpy-perf.cpp
+++ b/memcpy-perf/memcpy-perf.cpp
@@ -7,14 +7,20 @@
 #include <memory>
 #include <cmath>
 #include <string>
+#include <thread>
+
+#define CACHE_HIT_SIZE 1 << 17
 
 using namespace std;
 
-const size_t size_start = 64;
-const size_t size_end = 16 * (1ull << 20);
-const size_t samples = 2048;
+size_t size_start = 64;
+size_t size_end = 16 * (1ull << 20);
+size_t samples = 2048;
 size_t size_per_test = 64 * (1ull << 20);
 size_t tot_sum = 0;
+size_t delay = 0;
+float speed = 0;
+bool dummy = false;
 
 void __attribute__((noinline)) memcpy_noinline(void *dst, void *src, size_t size);
 void __attribute__((noinline)) memset_noinline(void *dst, int value, size_t size);
@@ -26,21 +32,64 @@ enum BenchType {
     SumBench,
 };
 
+static void usage(char* p) {
+    printf("Usage: %s <test> <options>\n"
+           "<test> is one of the following:\n"
+           "  --memcpy\n"
+           "  --memset\n"
+           "  --sum\n"
+           "<options> are optional and apply to all tests:\n"
+           "  --dummy\n"
+           "    Simulates cpu-only load of a test. Guaranteed to use L2\n"
+           "    instead.  Not supported on --sum test.\n"
+           "  --delay DELAY_DIVISOR\n"
+           "  --start START_SIZE_MB\n"
+           "    --end END_SIZE_MB (requires start, optional)\n"
+           "  --samples NUM_SAMPLES\n"
+           , p);
+}
+
 int main(int argc, char *argv[])
 {
-    BenchType type;
+    BenchType type = MemcpyBench;
     if (argc <= 1) {
-        cerr << "memcpy_perf [--memcpy|--memset|--sum]" << endl;
+        usage(argv[0]);
         return 0;
     }
-    if (string(argv[1]) == string("--memcpy")) {
-        type = MemcpyBench;
-    } else if (string(argv[1]) == string("--memset")) {
-        type = MemsetBench;
-    } else if (string(argv[1]) == string("--sum")) {
-        type = SumBench;
-    } else {
-        type = MemcpyBench;
+    for (int i = 1; i < argc; i++) {
+      if (string(argv[i]) == string("--memcpy")) {
+         type = MemcpyBench;
+      } else if (string(argv[i]) == string("--memset")) {
+         type = MemsetBench;
+      } else if (string(argv[i]) == string("--sum")) {
+         type = SumBench;
+      } else if (string(argv[i]) == string("--dummy")) {
+         dummy = true;
+      } else if (i + 1 < argc) {
+          if (string(argv[i]) == string("--delay")) {
+             delay = atoi(argv[++i]);
+          } else if (string(argv[i]) == string("--start")) {
+             size_start = atoi(argv[++i]) * (1ull << 20);
+             size_end = size_start;
+          } else if (string(argv[i]) == string("--end")) {
+             size_t end = atoi(argv[++i]) * (1ull << 20);
+             if (end > size_start && i > 3
+                 && string(argv[i-3]) == string("--start")) {
+                 size_end = end;
+             } else {
+                 printf("Cannot specify --end without --start.\n");
+                 return 0;
+             }
+          } else if (string(argv[i]) == string("--samples")) {
+             samples = atoi(argv[++i]);
+          } else {
+             printf("Unknown argument %s\n", argv[i]);
+             return 0;
+          }
+       } else {
+          printf("The %s option requires a single argument.\n", argv[i]);
+          return 0;
+       }
     }
 
     unique_ptr<uint8_t[]> src(new uint8_t[size_end]);
@@ -54,8 +103,10 @@ int main(int argc, char *argv[])
     //cout << "src: " << (uintptr_t)src.get() << endl;
     //cout << "dst: " <<  (uintptr_t)dst.get() << endl;
 
-    for (double cur_pow = start_pow; cur_pow <= end_pow; cur_pow += pow_inc) {
-        chrono::time_point<chrono::high_resolution_clock> copy_start, copy_end;
+    for (double cur_pow = start_pow; cur_pow <= end_pow && samples > 0;
+            cur_pow += pow_inc) {
+        chrono::time_point<chrono::high_resolution_clock>
+            copy_start, copy_end, pre_wait;
 
         size_t cur_size = (size_t)pow(10.0, cur_pow);
         size_t iter_per_size = size_per_test / cur_size;
@@ -65,9 +116,21 @@ int main(int argc, char *argv[])
             case MemsetBench: {
                 memcpy_noinline(src.get(), dst.get(), cur_size);
                 memset_noinline(dst.get(), 0xdeadbeef, cur_size);
+                size_t hit_size = CACHE_HIT_SIZE;
                 copy_start = chrono::high_resolution_clock::now();
                 for (int i = 0; i < iter_per_size; i++) {
-                    memset_noinline(dst.get(), 0xdeadbeef, cur_size);
+                    if (!dummy) {
+                        memset_noinline(dst.get(), 0xdeadbeef, cur_size);
+                    } else {
+                        while (hit_size < cur_size) {
+                            memset_noinline
+                                (dst.get(), 0xdeadbeef, CACHE_HIT_SIZE);
+                            hit_size += 1 << 17;
+                        }
+                    }
+                    if (delay != 0)
+                        this_thread::sleep_for(chrono
+                            ::nanoseconds(size_per_test / delay));
                 }
                 copy_end = chrono::high_resolution_clock::now();
                 break;
@@ -75,9 +138,21 @@ int main(int argc, char *argv[])
             case MemcpyBench: {
                 memcpy_noinline(dst.get(), src.get(), cur_size);
                 memcpy_noinline(src.get(), dst.get(), cur_size);
+                size_t hit_size = CACHE_HIT_SIZE;
                 copy_start = chrono::high_resolution_clock::now();
                 for (int i = 0; i < iter_per_size; i++) {
-                    memcpy_noinline(dst.get(), src.get(), cur_size);
+                    if (!dummy) {
+                        memcpy_noinline(dst.get(), src.get(), cur_size);
+                    } else {
+                        while (hit_size < cur_size) {
+                            memcpy_noinline
+                                (dst.get(), src.get(), CACHE_HIT_SIZE);
+                            hit_size += CACHE_HIT_SIZE;
+                        }
+                    }
+                    if (delay != 0)
+                        this_thread::sleep_for(chrono
+                            ::nanoseconds(size_per_test / delay));
                 }
                 copy_end = chrono::high_resolution_clock::now();
                 break;
@@ -88,6 +163,9 @@ int main(int argc, char *argv[])
                 copy_start = chrono::high_resolution_clock::now();
                 for (int i = 0; i < iter_per_size; i++) {
                     s += sum(src.get(), cur_size);
+                    if (delay != 0)
+                        this_thread::sleep_for(chrono
+                            ::nanoseconds(size_per_test / delay));
                 }
                 copy_end = chrono::high_resolution_clock::now();
                 tot_sum += s;
@@ -95,11 +173,18 @@ int main(int argc, char *argv[])
             }
         }
 
+        samples--;
         double ns_per_copy = chrono::duration_cast<chrono::nanoseconds>(copy_end - copy_start).count() / double(iter_per_size);
         double gb_per_sec = ((double)cur_size / (1ull<<30)) / (ns_per_copy / 1.0E9);
         if (type == MemcpyBench)
             gb_per_sec *= 2.0;
-        cout << "size: " << cur_size << ", perf: " << gb_per_sec << "GB/s, iter: " << iter_per_size << endl;
+        double percent_waiting = 0;
+        if (delay != 0) {
+            percent_waiting = (size_per_test / delay) / ns_per_copy * 100;
+        }
+        cout << "size: " << cur_size << ", perf: " << gb_per_sec
+             << "GB/s, iter: " << iter_per_size << ", \% time spent waiting: "
+             << percent_waiting << endl;
     }
     return 0;
 }