From fd1528795e7bbc8d7a59885a4387807abbc23f96 Mon Sep 17 00:00:00 2001
From: Peng Xiao <pengxiao@outlook.com>
Date: Sat, 13 Apr 2013 11:39:13 +0800
Subject: [PATCH] Pass query type T into kernel

---
 modules/ocl/src/brute_force_matcher.cpp     | 36 ++++++++++++++++++++++-------
 modules/ocl/src/opencl/brute_force_match.cl | 16 ++++++++-----
 2 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/modules/ocl/src/brute_force_matcher.cpp b/modules/ocl/src/brute_force_matcher.cpp
index c76ba33..7edc594 100644
--- a/modules/ocl/src/brute_force_matcher.cpp
+++ b/modules/ocl/src/brute_force_matcher.cpp
@@ -64,6 +64,8 @@ namespace cv
 
 static const int OPT_SIZE = 100;
 
+static const char * T_ARR [] = {"uchar", "char", "ushort", "short", "int", "float", "double"};
+
 template < int BLOCK_SIZE, int MAX_DESC_LEN/*, typename Mask*/ >
 void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
                          const oclMat &trainIdx, const oclMat &distance, int distType)
@@ -78,7 +80,9 @@ void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", distType, block_size, m_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", 
+        T_ARR[query.depth()], distType, block_size, m_size);
 
     if(globalSize[0] != 0)
     {
@@ -119,7 +123,9 @@ void match(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d", distType, block_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d", 
+        T_ARR[query.depth()], distType, block_size);
 
     if(globalSize[0] != 0)
     {
@@ -162,7 +168,9 @@ void matchUnrolledCached(const oclMat &query, const oclMat &train, float maxDist
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", distType, block_size, m_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", 
+        T_ARR[query.depth()], distType, block_size, m_size);
 
     if(globalSize[0] != 0)
     {
@@ -202,7 +210,9 @@ void radius_match(const oclMat &query, const oclMat &train, float maxDistance, c
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d", distType, block_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d", 
+        T_ARR[query.depth()], distType, block_size);
 
     if(globalSize[0] != 0)
     {
@@ -300,7 +310,9 @@ void knn_matchUnrolledCached(const oclMat &query, const oclMat &train, const ocl
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", distType, block_size, m_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", 
+        T_ARR[query.depth()], distType, block_size, m_size);
 
     if(globalSize[0] != 0)
     {
@@ -334,7 +346,9 @@ void knn_match(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d -D BLOCK_SIZE=%d", distType, block_size);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d", 
+        T_ARR[query.depth()], distType, block_size);
 
     if(globalSize[0] != 0)
     {
@@ -368,7 +382,10 @@ void calcDistanceUnrolled(const oclMat &query, const oclMat &train, const oclMat
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d", distType);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d", 
+        T_ARR[query.depth()], distType, block_size, m_size);
+
     if(globalSize[0] != 0)
     {
         args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
@@ -401,7 +418,10 @@ void calcDistance(const oclMat &query, const oclMat &train, const oclMat &/*mask
     vector< pair<size_t, const void *> > args;
 
     char opt [OPT_SIZE] = "";
-    sprintf(opt, "-D DIST_TYPE=%d", distType);
+    sprintf(opt, 
+        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d", 
+        T_ARR[query.depth()], distType, block_size);
+
     if(globalSize[0] != 0)
     {
         args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
diff --git a/modules/ocl/src/opencl/brute_force_match.cl b/modules/ocl/src/opencl/brute_force_match.cl
index 7446c77..db13249 100644
--- a/modules/ocl/src/opencl/brute_force_match.cl
+++ b/modules/ocl/src/opencl/brute_force_match.cl
@@ -47,6 +47,10 @@
 #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics:enable
 #define MAX_FLOAT 3.40282e+038f
 
+#ifndef T
+#define T float
+#endif
+
 #ifndef BLOCK_SIZE
 #define BLOCK_SIZE 16
 #endif
@@ -54,7 +58,11 @@
 #define MAX_DESC_LEN 64
 #endif
 
-int bit1Count(float x)
+#ifndef DIST_TYPE
+#define DIST_TYPE 0
+#endif
+
+int bit1Count(int x)
 {
     int c = 0;
     int ix = (int)x;
@@ -63,13 +71,9 @@ int bit1Count(float x)
         c += ix & 0x1;
         ix >>= 1;
     }
-    return (float)c;
+    return c;
 }
 
-#ifndef DIST_TYPE
-#define DIST_TYPE 0
-#endif
-
 #if   (DIST_TYPE == 0)
 #define DIST(x, y) fabs((x) - (y))
 #elif (DIST_TYPE == 1)
-- 
2.7.4