merged CCP with master

CHANGES.md

@@ -4,6 +4,7 @@
 * Binary file IO now supported.
 * Better handled missing input files.
 * Added CREDITS.md
+* Chains-on-chains partition available
 
 1.1.1
 =====

CMakeLists.txt

@@ -102,7 +102,7 @@ endif()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-file(GLOB SPLATT_SOURCES src/*.c ${MPI_SOURCES})
+file(GLOB SPLATT_SOURCES src/*.c src/ccp/*.c ${MPI_SOURCES})
 
 # Generate splatt library
 add_subdirectory(lib)

src/ccp/ccp.c

@@ -0,0 +1,182 @@
+
+
+/******************************************************************************
+ * INCLUDES
+ *****************************************************************************/
+
+#include "ccp.h"
+#include "../timer.h"
+
+
+/******************************************************************************
+ * PRIVATE FUNCTIONS
+ *****************************************************************************/
+
+static idx_t nprobes = 0;
+
+
+/**
+* @brief Perform a linear search on an array for a value.
+*
+* @param weights The array to search.
+* @param left The lower bound to begin at.
+* @param right The upper (exclusive) bound of items.
+* @param target The target value.
+*
+* @return The index j, where weights[j] <= target && weights[j+1] > target.
+*/
+static idx_t p_linear_search(
+    idx_t const * const weights,
+    idx_t const left,
+    idx_t const right,
+    idx_t const target)
+{
+  for(idx_t x=left; x < right-1; ++x) {
+    if(weights[x+1] > target) {
+      return x;
+    }
+  }
+
+  return right;
+}
+
+
+/**
+* @brief Perform a binary search on an array for a value.
+*
+* @param weights The array to search.
+* @param left The lower bound to begin at.
+* @param right The upper (exclusive) bound of items.
+* @param target The target value.
+*
+* @return The index j, where weights[j] <= target && weights[j+1] > target.
+*/
+static idx_t p_binary_search(
+    idx_t const * const weights,
+    idx_t left,
+    idx_t right,
+    idx_t const target)
+{
+  while((right - left) > 8) {
+    idx_t mid = left + ((right - left) / 2);
+
+    if(weights[mid] <= target && weights[mid+1] > target) {
+      return mid;
+    }
+
+    if(weights[mid] < target) {
+      left = mid + 1;
+    } else {
+      right = mid;
+    }
+  }
+
+  return p_linear_search(weights, left, right, target);
+}
+
+
+
+static idx_t p_eps_rb_partition_1d(
+    idx_t * const weights,
+    idx_t const nitems,
+    idx_t * const parts,
+    idx_t const nparts,
+    idx_t const eps)
+{
+  idx_t lower = weights[nitems-1] / nparts;
+  idx_t upper = weights[nitems-1];
+
+  do {
+    idx_t mid = lower + ((upper - lower) / 2);
+    if(lprobe(weights, nitems, parts, nparts, mid)) {
+      upper = mid;
+    } else {
+      lower = mid+1;
+    }
+  } while(upper > lower + eps);
+
+  return upper;
+}
+
+
+
+
+/******************************************************************************
+ * PUBLIC FUNCTIONS
+ *****************************************************************************/
+
+idx_t partition_1d(
+    idx_t * const weights,
+    idx_t const nitems,
+    idx_t * const parts,
+    idx_t const nparts)
+{
+  timer_start(&timers[TIMER_PART]);
+  prefix_sum_inc(weights, nitems);
+
+  nprobes = 0;
+
+  /* use recursive bisectioning with 0 tolerance to get exact solution */
+  idx_t bottleneck = p_eps_rb_partition_1d(weights, nitems, parts, nparts, 0);
+
+  timer_stop(&timers[TIMER_PART]);
+  return bottleneck;
+}
+
+
+
+bool lprobe(
+    idx_t const * const weights,
+    idx_t const nitems,
+    idx_t * const parts,
+    idx_t const nparts,
+    idx_t const bottleneck)
+{
+  idx_t p=0;
+  parts[p++] = 0;
+  idx_t bsum = bottleneck;
+
+  idx_t const wtotal = weights[nitems-1];
+
+  idx_t step = nitems / nparts;
+  while(p < nparts && bsum < wtotal) {
+    while(step < nitems && weights[step] < bsum) {
+      step += nitems / nparts;
+    }
+    parts[p] = p_binary_search(weights, step - (nitems/nparts), SS_MIN(step, nitems),
+        bsum);
+    bsum = weights[parts[p]] + bottleneck;
+    ++p;
+  }
+
+  parts[p] = nitems;
+
+  ++nprobes;
+  return bsum >= wtotal;
+}
+
+
+
+void prefix_sum_inc(
+    idx_t * const weights,
+    idx_t const nitems)
+{
+  for(idx_t x=1; x < nitems; ++x) {
+    weights[x] += weights[x-1];
+  }
+}
+
+
+
+void prefix_sum_exc(
+    idx_t * const weights,
+    idx_t const nitems)
+{
+  idx_t saved = weights[0];
+  weights[0] = 0;
+  for(idx_t x=1; x < nitems; ++x) {
+    idx_t const tmp = weights[x];
+    weights[x] = weights[x-1] + saved;
+    saved = tmp;
+  }
+}

src/ccp/ccp.h

@@ -0,0 +1,58 @@
+#ifndef SPLATT_CCP_CCP_H
+#define SPLATT_CCP_CCP_H
+
+#include "../base.h"
+
+
+/******************************************************************************
+ * INCLUDES
+ *****************************************************************************/
+#include <stdbool.h>
+
+
+
+
+/******************************************************************************
+ * PUBLIC FUNCTIONS
+ *****************************************************************************/
+
+#define partition_1d splatt_partition_1d
+idx_t partition_1d(
+    idx_t * const weights,
+    idx_t const nitems,
+    idx_t * const parts,
+    idx_t const nparts);
+
+
+bool lprobe(
+    idx_t const * const weights,
+    idx_t const nitems,
+    idx_t * const parts,
+    idx_t const nparts,
+    idx_t const bottleneck);
+
+
+#define prefix_sum_inc splatt_prefix_sum_inc
+/**
+* @brief Compute an inclusive prefix sum: [3, 4, 5] -> [3, 7, 12].
+*
+* @param weights The numbers to sum.
+* @param nitems The number of items in 'weights'.
+*/
+void prefix_sum_inc(
+    idx_t * const weights,
+    idx_t const nitems);
+
+
+#define prefix_sum_exc splatt_prefix_sum_exc
+/**
+* @brief Compute an exclusive prefix sum: [3, 4, 5] -> [0, 3, 7].
+*
+* @param weights The numbers to sum.
+* @param nitems The number of items in 'weights'.
+*/
+void prefix_sum_exc(
+    idx_t * const weights,
+    idx_t const nitems);
+
+#endif

src/mpi/mpi_io.c

@@ -8,7 +8,6 @@
 #include "../util.h"
 
 
-
 /******************************************************************************
  * API FUNCTONS
  *****************************************************************************/
@@ -279,7 +278,6 @@ static sptensor_t * p_rearrange_by_part(
 }
 
 
-
 static void p_find_my_slices_1d(
   idx_t ** const ssizes,
   idx_t const nmodes,

src/timer.c

@@ -28,6 +28,7 @@ static char const * const timer_names[] = {
   [TIMER_MATMUL]    = "MAT MULT",
   [TIMER_ATA]       = "MAT A^TA",
   [TIMER_MATNORM]   = "MAT NORM",
+  [TIMER_PART]      = "PART1D",
   [TIMER_MISC]      = "MISC",
 #ifdef SPLATT_USE_MPI
   [TIMER_MPI]           = "MPI",

src/timer.h

@@ -42,6 +42,7 @@ typedef enum
   TIMER_ATA,
   TIMER_MATNORM,
   TIMER_IO,
+  TIMER_PART,
   TIMER_LVL2,   /* LEVEL 2 */
 #ifdef SPLATT_USE_MPI
   TIMER_MPI,