Replace cout usages with logger calls

Author: jackbondpreston-arm

gloo/allreduce_bcube.cc

@@ -0,0 +1,74 @@
+#include "gloo/allreduce_bcube.h"
+
+#include <string_view>
+
+#include "gloo/common/log.h"
+
+namespace gloo {
+
+template <typename T>
+void AllreduceBcube<T>::printElems(T* p, int count, int start) {
+  /* Early return if log level is not high enough, to prevent expensive code
+   * running. */
+  if (!spdlog::should_log(spdlog::level::trace))
+    return;
+
+  const std::size_t alignedStart = (start / wordsPerLine) * wordsPerLine;
+  fmt::memory_buffer line{};
+
+  /* Logs/flushes the line buffer - starting a new line */
+  auto printLine = [&]() {
+    if (!line.size())
+      return;
+    std::string_view sv{line.data(), line.size()};
+    GLOO_TRACE("{}", sv);
+    line.clear();
+  };
+
+  for (std::size_t x = alignedStart; x < start + count; ++x) {
+    if (x % wordsPerLine == 0) {
+      if (x != alignedStart)
+        printLine();
+      fmt::format_to(
+          std::back_inserter(line), "{} {:05}: ", fmt::ptr(&p[x]), x);
+    } else if (x % wordsPerSection == 0) {
+      fmt::format_to(std::back_inserter(line), "- ");
+    }
+
+    if (x < start)
+      fmt::format_to(std::back_inserter(line), "..... ");
+    else
+      fmt::format_to(std::back_inserter(line), "{:05} ", p[x]);
+  }
+  printLine();
+}
+
+template <typename T>
+void AllreduceBcube<T>::printStageBuffer(const std::string& msg) {
+  if (printCheck(myRank_)) {
+    GLOO_TRACE("rank ({}) {}:", myRank_, msg);
+    printElems(&ptrs_[0][0], totalNumElems_);
+  }
+}
+
+template <typename T>
+void AllreduceBcube<T>::printStepBuffer(
+    const std::string& stage,
+    int step,
+    int srcRank,
+    int destRank,
+    T* p,
+    int count,
+    int start) {
+  if (printCheck(myRank_)) {
+    GLOO_TRACE(
+        "{}: step ({}) srcRank ({}) -> destRank ({})",
+        stage,
+        step,
+        srcRank,
+        destRank);
+    printElems(p, count, start);
+  }
+}
+
+} // namespace gloo

gloo/allreduce_bcube.h

@@ -13,15 +13,12 @@
 #include <string.h>
 #include <algorithm>
 #include <cstring>
-#include <iomanip>
-#include <iostream>
 #include <memory>
 #include <string>
 #include <unordered_map>
 #include <vector>
 
 #include "gloo/algorithm.h"
-#include "gloo/common/error.h"
 #include "gloo/context.h"
 
 /**
@@ -537,48 +534,18 @@ class AllreduceBcube : public Algorithm {
   static bool printCheck(int /*rank*/) {
     return false;
   }
-  /**
-   * Prints a break given the offset of an element about to be printed
-   * @param p Pointer to the elements
-   * @param x The current offset to the pointer to words
-   */
-  static void printBreak(T* p, int x) {
-    if (0 == x % wordsPerLine) {
-      std::cout << std::endl
-                << &p[x] << " " << std::setfill('0') << std::setw(5) << x
-                << ": ";
-    } else if (0 == x % wordsPerSection) {
-      std::cout << "- ";
-    }
-  }
   /**
    * Pretty prints a list of elements
    * @param p Pointer to the elements
    * @param count The number of elements to be printed
    * @param start The offset from which to print
    */
-  static void printElems(T* p, int count, int start = 0) {
-    auto alignedStart = (start / wordsPerLine) * wordsPerLine;
-    for (int x = alignedStart; x < start + count; ++x) {
-      printBreak(p, x);
-      if (x < start) {
-        std::cout << "..... ";
-      } else {
-        std::cout << std::setfill('0') << std::setw(5) << p[x] << " ";
-      }
-    }
-  }
+  static void printElems(T* p, int count, int start);
   /**
    * Prints contents in the ptrs array at a particular stage
    * @param msg Custom message to be printed
    */
-  void printStageBuffer(const std::string& msg) {
-    if (printCheck(myRank_)) {
-      std::cout << "rank (" << myRank_ << ") " << msg << ": ";
-      printElems(&ptrs_[0][0], totalNumElems_);
-      std::cout << std::endl;
-    }
-  }
+  void printStageBuffer(const std::string& msg);
 
   /**
    * Prints specified buffer during a step
@@ -596,14 +563,7 @@ class AllreduceBcube : public Algorithm {
       int destRank,
       T* p,
       int count,
-      int start = 0) {
-    if (printCheck(myRank_)) {
-      std::cout << stage << ": step (" << step << ") " << "srcRank (" << srcRank
-                << ") -> " << "destRank (" << destRank << "): ";
-      printElems(p, count, start);
-      std::cout << std::endl;
-    }
-  }
+      int start = 0);
   /**
    * Get all the peers of node with specified rank
    * @param rank Rank of the node for which peers are needed

gloo/cuda_allreduce_bcube.cc

@@ -8,11 +8,13 @@
 
 #include "gloo/cuda_allreduce_bcube.h"
 
+#include "gloo/common/log.h"
 #include "gloo/cuda_collectives_device.h"
 #include "gloo/cuda_collectives_host.h"
 #include "gloo/cuda_private.h"
 
 #include <sstream>
+#include <string_view>
 #ifndef _WIN32
 #include <unistd.h>
 #endif
@@ -226,35 +228,48 @@ bool CudaAllreduceBcube<T, W>::printCheck(int /* rank */) {
   return false;
 }
 
-template <typename T, typename W>
-void CudaAllreduceBcube<T, W>::printBreak(T* p, int x) {
-  if (0 == x % wordsPerLine) {
-    std::cout << std::endl
-              << &p[x] << " " << std::setfill('0') << std::setw(5) << x << ": ";
-  } else if (0 == x % wordsPerSection) {
-    std::cout << "- ";
-  }
-}
-
 template <typename T, typename W>
 void CudaAllreduceBcube<T, W>::printElems(T* p, int count, int start) {
-  auto alignedStart = (start / wordsPerLine) * wordsPerLine;
-  for (int x = alignedStart; x < start + count; ++x) {
-    printBreak(p, x);
-    if (x < start) {
-      std::cout << "..... ";
-    } else {
-      std::cout << std::setfill('0') << std::setw(5) << p[x] << " ";
+  /* Early return if log level is not high enough, to prevent expensive code
+   * running. */
+  if (!spdlog::should_log(spdlog::level::trace))
+    return;
+
+  const std::size_t alignedStart = (start / wordsPerLine) * wordsPerLine;
+  fmt::memory_buffer line{};
+
+  /* Logs/flushes the line buffer - starting a new line */
+  auto printLine = [&]() {
+    if (!line.size())
+      return;
+    std::string_view sv{line.data(), line.size()};
+    GLOO_TRACE("{}", sv);
+    line.clear();
+  };
+
+  for (std::size_t x = alignedStart; x < start + count; ++x) {
+    if (x % wordsPerLine == 0) {
+      if (x != alignedStart)
+        printLine();
+      fmt::format_to(
+          std::back_inserter(line), "{} {:05}: ", fmt::ptr(&p[x]), x);
+    } else if (x % wordsPerSection == 0) {
+      fmt::format_to(std::back_inserter(line), "- ");
     }
+
+    if (x < start)
+      fmt::format_to(std::back_inserter(line), "..... ");
+    else
+      fmt::format_to(std::back_inserter(line), "{:05} ", p[x]);
   }
+  printLine();
 }
 
 template <typename T, typename W>
 void CudaAllreduceBcube<T, W>::printStageBuffer(const std::string& msg) {
   if (printCheck(myRank_)) {
-    std::cout << "rank (" << myRank_ << ") " << msg << ": ";
+    GLOO_TRACE("rank ({}) {}:", myRank_, msg);
     printElems(&scratch_[0], totalNumElems_);
-    std::cout << std::endl;
   }
 }
 
@@ -268,10 +283,13 @@ void CudaAllreduceBcube<T, W>::printStepBuffer(
     int count,
     int start) {
   if (printCheck(myRank_)) {
-    std::cout << stage << ": step (" << step << ") " << "srcRank (" << srcRank
-              << ") -> " << "destRank (" << destRank << "): ";
+    GLOO_TRACE(
+        "{}: step ({}) srcRank ({}) -> destRank ({}):",
+        stage,
+        step,
+        srcRank,
+        destRank);
     printElems(p, count, start);
-    std::cout << std::endl;
   }
 }
 

gloo/transport/buffer.h

@@ -16,13 +16,9 @@ namespace transport {
 class Buffer {
  public:
   explicit Buffer(int slot, void* ptr, size_t size)
-      : slot_(slot), ptr_(ptr), size_(size), debug_(false) {}
+      : slot_(slot), ptr_(ptr), size_(size) {}
   virtual ~Buffer() = 0;
 
-  virtual void setDebug(bool debug) {
-    debug_ = debug;
-  }
-
   virtual void send(size_t offset, size_t length, size_t roffset = 0) = 0;
 
   // Send entire buffer by default
@@ -37,7 +33,6 @@ class Buffer {
   int slot_;
   void* ptr_;
   size_t size_;
-  bool debug_;
 };
 
 } // namespace transport

gloo/transport/ibverbs/buffer.cc

@@ -189,11 +189,7 @@ void Buffer::send(size_t offset, size_t length, size_t roffset) {
 
     checkErrorState();
 
-    if (debug_) {
-      std::cout << "[" << getpid() << "] ";
-      std::cout << "send " << length << " bytes";
-      std::cout << std::endl;
-    }
+    GLOO_TRACE("send {} bytes", length);
 
     // Increment number of sends in flight
     sendPending_++;
@@ -208,19 +204,11 @@ void Buffer::handleCompletion(int rank, struct ibv_wc* wc) {
   std::unique_lock<std::mutex> lock(m_);
 
   if (wc->opcode & IBV_WC_RECV) {
-    if (debug_) {
-      std::cout << "[" << getpid() << "] ";
-      std::cout << "recv " << wc->byte_len << " bytes";
-      std::cout << std::endl;
-    }
+    GLOO_TRACE("recv {} bytes", wc->byte_len);
     recvCompletions_++;
     recvCv_.notify_one();
   } else if (wc->opcode == IBV_WC_RDMA_WRITE) {
-    if (debug_) {
-      std::cout << "[" << getpid() << "] ";
-      std::cout << "send complete";
-      std::cout << std::endl;
-    }
+    GLOO_TRACE("send complete");
     sendCompletions_++;
     sendPending_--;
     sendCv_.notify_one();

gloo/transport/tcp/buffer.cc

@@ -6,16 +6,15 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include "gloo/transport/tcp/buffer.h"
-
 #include <string.h>
 #include <sys/syscall.h>
 #include <unistd.h>
-
 #include <iostream>
 
-#include "gloo/common/error.h"
 #include "gloo/common/enforce.h"
+#include "gloo/common/error.h"
+#include "gloo/common/log.h"
+#include "gloo/transport/tcp/buffer.h"
 
 namespace gloo {
 namespace transport {
@@ -126,12 +125,7 @@ void Buffer::send(size_t offset, size_t length, size_t roffset) {
   // to support this.
   GLOO_ENFORCE_LE(offset + length, size_);
 
-  if (debug_) {
-    std::cout << "[" << getpid() << ": " << syscall(SYS_gettid) << "] ";
-    std::cout << "send " << length << " bytes";
-    std::cout << " to " << pair_->peer().str();
-    std::cout << std::endl;
-  }
+  GLOO_TRACE("send {} bytes to {}", length, pair_->peer().str());
 
   op.preamble.nbytes = sizeof(op.preamble) + length;
   op.preamble.opcode = Op::SEND_BUFFER;

gloo/types.h

@@ -339,4 +339,12 @@ inline float cpu_half2float(float16 h) {
   return *(float*)rp;
 }
 
+/**
+ * @brief Format helper for spdlog/fmt - convert to a float32 and format as
+ * usual
+ */
+inline float format_as(float16 v) {
+  return cpu_half2float(v);
+}
+
 } // namespace gloo