Merge pull request #40 from RadeonOpenCompute/xtndBufferInit

Extend buffer initialization choice

rocm_bandwidth_test.cpp

@@ -48,6 +48,7 @@
 #include <algorithm>
 #include <unistd.h>
 #include <cctype>
+#include <cmath>
 #include <sstream>
 #include <limits>
 
@@ -105,7 +106,11 @@ void RocmBandwidthTest::InitializeSrcBuffer(size_t size, void* buf_cpy,
   if (init_src_ == NULL) {
     err_ = hsa_amd_memory_pool_allocate(sys_pool_, size, 0, (void**)&init_src_);
     ErrorCheck(err_);
-    memset(init_src_, init_val_, size);
+    long double* src_buf = (long double*) init_src_;
+    uint32_t count = (size / sizeof(long double));
+    for (uint32_t idx = 0; idx < count; idx++) {
+      src_buf[idx] = (init_) ? init_val_ : sin(idx);
+    }
     err_ = hsa_signal_create(0, 0, NULL, &init_signal_);
     ErrorCheck(err_);
   }
@@ -136,7 +141,7 @@ bool RocmBandwidthTest::ValidateDstBuffer(size_t max_size, size_t curr_size, voi
   }
 
   // If Copy device is a Gpu setup buffer access
-  memset(validate_dst_, ~init_val_, curr_size);
+  memset(validate_dst_, ~(0x23), curr_size);
   hsa_device_type_t cpy_dev_type = agent_list_[cpy_dev_idx].device_type_;
   if (cpy_dev_type == HSA_DEVICE_TYPE_GPU) {
     AcquireAccess(cpy_agent, validate_dst_);
@@ -217,6 +222,9 @@ void RocmBandwidthTest::AllocateConcurrentCopyResources(bool bidir,
     dev_idx_list.push_back(src_dev_idx);
     dev_idx_list.push_back(dst_dev_idx);
   
+    // Initialize source buffers with data that could be verified
+    InitializeSrcBuffer(max_size, buf_src, src_dev_idx, src_dev);
+    
     // For bidirectional copies allocate buffers
     // and signal for reverse direction as well
     if (bidir) {
@@ -237,6 +245,9 @@ void RocmBandwidthTest::AllocateConcurrentCopyResources(bool bidir,
       dev_list.push_back(src_dev);
       dev_idx_list.push_back(dst_dev_idx);
       dev_idx_list.push_back(src_dev_idx);
+    
+      // Initialize source buffers with data that could be verified
+      InitializeSrcBuffer(max_size, buf_src, dst_dev_idx, dst_dev);
     }
   }
 }
@@ -526,15 +537,12 @@ void RocmBandwidthTest::RunCopyBenchmark(async_trans_t& trans) {
   }
 
   // Initialize source buffers with data that could be verified
-  if ((init_) || (validate_)) {
   InitializeSrcBuffer(max_size, buf_src_fwd,
                       src_dev_idx_fwd, src_agent_fwd);
-
   if (bidir) {
     InitializeSrcBuffer(max_size, buf_src_rev,
                         src_dev_idx_rev, src_agent_rev);
   }
-  }
 
   // Setup access to destination buffers for
   // both unidirectional and bidirectional copies
@@ -760,15 +768,15 @@ RocmBandwidthTest::RocmBandwidthTest(int argc, char** argv) : BaseTest() {
   validate_ = false;
   print_cpu_time_ = false;
   
-  // Set initial value to 0x23 in case
+  // Set initial value to 11.231926 in case
   // user does not have a preference
-  init_val_ = 0x23;
+  init_val_ = 11.231926;
   init_src_ = NULL;
   validate_dst_ = NULL;
 
   // Initialize version of the test
   version_.major_id = 2;
-  version_.minor_id = 2;
+  version_.minor_id = 3;
   version_.step_id = 0;
   version_.reserved = 0;
 
@@ -794,13 +802,15 @@ RocmBandwidthTest::~RocmBandwidthTest() {
   delete link_type_matrix_;
   delete link_weight_matrix_;
   delete active_agents_list_;
-  if (init_) {
+
+  if (init_src_ != NULL) {
     hsa_signal_destroy(init_signal_);
     hsa_amd_memory_pool_free(init_src_);
+  }
+
   if (validate_) {
     hsa_amd_memory_pool_free(validate_dst_);
   }
-  }
 }
 
 std::string RocmBandwidthTest::GetVersion() const {

rocm_bandwidth_test.hpp

@@ -517,7 +517,7 @@ class RocmBandwidthTest : public BaseTest {
 
   // Determines if user has requested validation
   bool validate_;
-  uint8_t init_val_;
+  long double init_val_;
 
   // Handles to buffer used to initialize and validate
   void* init_src_;

rocm_bandwidth_test_parse.cpp

@@ -46,15 +46,41 @@
 #include <assert.h>
 #include <algorithm>
 #include <sstream>
+#include <iomanip>
 #include <unistd.h>
+//#include <strings.h>
+
+// Parse option value string. The string has to be either
+// sin or cos literal
+// value as in example: -I sin or -I cos
+/*
+static bool ParseTrigValue(char* value_str, uint32_t&value) {
+ 
+  // Capture the option value string
+  std::cout << "Value of Trig: " << value_str << std::endl;
+  int32_t cmp = strncasecmp("sin", value_str, 3);
+  if (cmp == 0) {
+    value = 1;
+    return true;
+  }
+
+  cmp = strncasecmp("cos", value_str, 3);
+  if (cmp == 0) {
+    value = 2;
+    return true;
+  }
+  
+  return false;
+}
+*/
 
 // Parse option value string. The string has one decimal
-// value as in example: -i 0x33
-static bool ParseInitValue(char* value_str, uint8_t&value) {
+// value as in example: -i 11.231926
+static bool ParseInitValue(char* value_str, long double&value) {
  
   // Capture the option value string
-  uint32_t value_read = strtoul(value_str, NULL, 0);
-  return ((value = value_read) && (value_read > 255)) ? false : true;
+  value = strtold(value_str, NULL);
+  return true;
 }
 
 // Parse option value string. The string has one more decimal
@@ -456,7 +482,8 @@ void RocmBandwidthTest::ParseArguments() {
       case '?':
         std::cout << "Argument is illegal or needs value: " << '?' << std::endl;
         if ((optopt == 'b') || (optopt == 's') ||
-            (optopt == 'd') || (optopt == 'm') || (optopt == 'i')) {
+            (optopt == 'd') || (optopt == 'm') ||
+            (optopt == 'i') || (false)) {
           std::cout << "Error: Options -b -s -d -m -i -k and -K require argument" << std::endl;
         }
         print_help = true;

rocm_bandwidth_test_print.cpp

@@ -59,7 +59,7 @@ void RocmBandwidthTest::PrintHelpScreen() {
   std::cout << "\t -v    Run the test in validation mode" << std::endl;
   std::cout << "\t -l    Run test to collect Latency data" << std::endl;
   std::cout << "\t -c    Time the operation using CPU Timers" << std::endl;
-  std::cout << "\t -i    Initialize copy buffer with specified byte pattern" << std::endl;
+  std::cout << "\t -i    Initialize copy buffer with specified 'long double' pattern" << std::endl;
   std::cout << "\t -t    Prints system topology and allocatable memory info" << std::endl;
   std::cout << "\t -m    List of buffer sizes to use, specified in Megabytes" << std::endl;
   std::cout << "\t -b    List devices to use in bidirectional copy operations" << std::endl;