8 files changed, 327 insertions, 362 deletions

diff --git a/apps/CMakeLists.txt b/apps/CMakeLists.txt
index 80a506f..85457a1 100644
--- a/apps/CMakeLists.txt
+++ b/apps/CMakeLists.txt
@@ -7,5 +7,9 @@ link_directories(${UFODECODE_LIBRARY_DIRS})
 add_executable(xilinx xilinx.c)
 target_link_libraries(xilinx pcilib rt)
 
-add_executable(xilinx2 xilinx2.c)
-target_link_libraries(xilinx2 pcilib rt)
+add_executable(pio_test pio_test.c)
+target_link_libraries(pio_test pcilib rt)
+
+add_executable(compare_to_value compare_to_value.c)
+
+add_executable(heb_strip_bad_values heb_strip_bad_values.c)
diff --git a/apps/compare_to_value.c b/apps/compare_to_value.c
new file mode 100644
index 0000000..75ad353
--- /dev/null
+++ b/apps/compare_to_value.c
@@ -0,0 +1,66 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+
+int main(int argc, char *argv[]) {
+    long i, j, size, num;
+    size_t count = 0, total = 0;
+    int offset = 0, toread = 1, toskip = 0;
+    uint32_t value;
+    uint32_t *buf;
+    
+    if ((argc != 4)&&(argc != 7)) {
+	printf("Usage: %s <file> <dwords> <value> [offset_dwords read_dwords skip_dwords] \n", argv[0]);
+	exit(0);
+    }
+    
+    FILE *f = fopen(argv[1], "r");
+    if (!f) {
+	printf("Can't open %s\n", argv[1]);
+	exit(1);
+    }
+    
+    size = atol(argv[2]);
+    if (size <= 0) {
+	printf("Can't parse size %s\n", argv[2]);
+	exit(1);
+    }
+    
+    if (sscanf(argv[3], "%x", &value) != 1) {
+	printf("Can't parse register %s\n", argv[3]);
+	exit(1);
+    }
+
+    buf = malloc(size * sizeof(uint32_t));
+    if (!buf) {
+	printf("Can't allocate %lu bytes of memory\n", size * sizeof(uint32_t));
+	exit(1);
+    }
+    
+    if (argc == 7) {
+	offset = atoi(argv[4]);
+	toread = atoi(argv[5]);
+	toskip = atoi(argv[6]);
+    }
+    
+
+    num = fread(buf, 4, size, f);
+    if (num != size) {
+	printf("Only %lu of %lu dwords in the file\n", num, size);
+	exit(1);
+    }
+    fclose(f);
+    
+    for (i = offset; i < size; i += toskip) {
+	for (j = 0; j < toread; j++, i++) {
+	    total++;
+	    if (buf[i] != value) {
+		count++;
+	    }
+	}
+    }
+    free(buf);
+    
+    printf("%lu of %lu is wrong\n", count, total);
+    return 0;
+}
diff --git a/apps/counters.sh b/apps/counters.sh
deleted file mode 100755
index e3ba0f5..0000000
--- a/apps/counters.sh
+++ /dev/null
@@ -1,117 +0,0 @@
-#! /bin/bash
-
-BAR=0
-USE=1
-ITERATIONS=1
-TLP_SIZE=32
-BUFFER_SIZE=8
-
-function pci {
-    PCILIB_PATH=`pwd`/..
-    LD_LIBRARY_PATH="$PCILIB_PATH" $PCILIB_PATH/pci $*
-}
-
-
-function reset {
-    pci -b $BAR -w 0 1
-    usleep 1000
-    pci -b $BAR -w 0 0
-    pci -b $BAR -w 4 0
-}
-
-function read_cfg {
-#    echo $1 1>&2
-    pci -a config -r 0x$1 | awk '{ print $2; }'
-}
-
-function parse_config {
-    info=0x`pci -b $BAR -r 0 | awk '{ print $2; }'`
-    model=`printf "%X" $((info>>24))`
-    if [ $model -eq 14 ]; then
-	model="Xilinx Virtex-6"
-    else
-	model="Xilinx $model"
-    fi
-    version=$(((info >> 8) & 0xFF))
-    data_width=$((16 * (2 ** ((info >> 16) & 0xF))))
-    
-    echo "$model, build $version, $data_width bits"
-
-
-    next=`read_cfg 34 | cut -c 7-8`
-
-    while [ $next -ne 0 ]; do
-	cap=`read_cfg $next`
-	capid=`echo $cap | cut -c 7-8`
-	if [ $capid -eq 10 ]; then
-	    addr=`printf "%X" $((0x$next + 12))`
-	    pcie_link1=`read_cfg $addr`
-	    addr=`printf "%X" $((0x$next + 16))`
-	    pcie_link2=`read_cfg $addr`
-
-	    link_speed=$((((0x$pcie_link2 & 0xF0000) >> 16)))
-	    link_width=$((((0x$pcie_link2 & 0x3F00000) >> 20)))
-
-	    dev_link_speed=$((((0x$pcie_link1 & 0xF))))
-	    dev_link_width=$((((0x$pcie_link1 & 0x3F0) >> 4)))
-	fi
-	next=`echo $cap | cut -c 5-6`
-    done
-
-    echo "Link: PCIe gen$link_speed x$link_width"
-    if [ $link_speed -ne $dev_link_speed -o $link_width -ne $dev_link_width ]; then
-	echo " * But device capable of gen$dev_link_speed x$dev_link_width"
-    fi
-    
-    info=0x`read_cfg 40`
-    max_tlp=$((2 ** (5 + ((info & 0xE0) >> 5))))
-    echo "TLP: 32 dwords (transfering 32 TLP per request)"
-    if [ $max_tlp -ne $TLP_SIZE ]; then
-	echo " * But device is able to transfer TLP up to $max_tlp bytes"
-    fi
-    
-    # 2500 MT/s, but PCIe gen1 and gen2 uses 10 bit encoding
-    speed=$((link_width * link_speed * 2500 / 10))
-}
-
-reset
-parse_config
-
-pci --enable-irq
-pci --acknowledge-irq
-
-# TLP size
-pci -b $BAR -w 0x0C 0x`echo "obase=16; $TLP_SIZE" | bc`
-# TLP count
-pci -b $BAR -w 0x10 0x`echo "obase=16; $BUFFER_SIZE * 1024 * 1024 / $TLP_SIZE / 4" | bc`
-# Data
-pci -b $BAR -w 0x14 0x13131313
-
-bus="80000000"
-dmaperf=0
-for i in `seq 1 $ITERATIONS`; do
-  for addr in $bus; do 
-    pci -b $BAR -w 0x08 0x$addr
-
-#Trigger
-    pci -b $BAR -w 0x04 0x01
-    pci --wait-irq
-
-    status=`pci -b $BAR -r 0x04 | awk '{print $2; }' | cut -c 5-8`
-    if [ $status != "0101" ]; then
-	echo "Read failed, invalid status: $status"
-    fi
-
-    dmaperf=$((dmaperf + 0x`pci -b $BAR -r 0x28 | awk '{print $2}'`))
-    reset
-  done
-done
-
-pci --free-kernel-memory $USE
-pci --disable-irq
-
-echo
-# Don't ask me about this formula
-echo "Performance reported by FPGA: $(($BUFFER_SIZE * 1024 * 1024 * ITERATIONS * $speed / $dmaperf / 8)) MB/s"
-
-#pci -b $BAR  -r 0 -s 32
diff --git a/apps/heb_strip_bad_values.c b/apps/heb_strip_bad_values.c
new file mode 100644
index 0000000..e04a53c
--- /dev/null
+++ b/apps/heb_strip_bad_values.c
@@ -0,0 +1,98 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+
+int main(int argc, char *argv[]) {
+    long i, num;
+    size_t count = 0, total = 0, size;
+    int offset = 3, toread = 1, toskip = 3;
+    uint32_t value;
+    uint32_t *buf;
+    uint32_t expected = 0;
+    uint32_t blocks = 0, status_good = 0;
+    
+    char fixed[4096];
+    struct stat st_buf;
+    
+    if ((argc != 2)&&(argc != 5)) {
+	printf("Usage: %s <file> [offset_dwords read_dwords skip_dwords] \n", argv[0]);
+	exit(0);
+    }
+    
+    FILE *f = fopen(argv[1], "r");
+    if (!f) {
+	printf("Can't open %s\n", argv[1]);
+	exit(1);
+    }
+
+    stat(argv[1], &st_buf);
+    size = st_buf.st_size / sizeof(uint32_t);
+
+
+    buf = malloc(size * sizeof(uint32_t));
+    if (!buf) {
+	printf("Can't allocate %lu bytes of memory\n", size * sizeof(uint32_t));
+	exit(1);
+    }
+    
+    if (argc == 5) {
+	offset = atoi(argv[2]);
+	toread = atoi(argv[3]);
+	toskip = atoi(argv[4]);
+    }
+    
+
+    num = fread(buf, 4, size, f);
+    if (num != size) {
+	printf("Failed to read %lu dwords, only %lu read\n", size, num);
+	exit(1);
+    }
+    fclose(f);
+    
+    sprintf(fixed, "%s.fixed", argv[1]);
+    f = fopen(fixed, "w");
+    if (!f) {
+	printf("Failed to open %s for output\n", fixed);
+	exit(1);
+    }
+    
+    expected = (buf[offset]>>24) + 2;
+    for (i = 1; i < size; i += (toread + toskip)) {
+	total++;
+	    
+	value = buf[i + offset] >> 24;
+//	printf("0x%lx: value (%lx) = expected (%lx)\n", i + offset, value, expected);
+	if (value == expected) {
+	    if (!status_good) {
+		status_good = 1;
+		blocks++;
+	    }
+	    fwrite(&buf[i], 4, toread + toskip, f);
+	    expected += 2;
+	    if (expected == 0xb8) 
+		expected = 0;
+	} else if ((!status_good)&&(value == 0)&&((i + toread + toskip)< size)) {
+	    value = buf[i + offset + toread + toskip] >> 24;
+	    if (value == 2) {
+		status_good = 1;
+		blocks++;
+		fwrite(&buf[i], 4, toread + toskip, f);
+		expected = 2;
+	    } else {
+		count++;
+	    }
+	} else {
+	    printf("0x%lx: value (%x) = expected (%x)\n", (i + offset)*sizeof(uint32_t), value, expected);
+	    status_good = 0;
+	    count++;
+	}
+    }
+    fclose(f);
+    free(buf);
+    
+    printf("%lu of %lu is wrong\n", count, total);
+    return 0;
+}
diff --git a/apps/load.sh b/apps/load.sh
deleted file mode 100755
index 6456c63..0000000
--- a/apps/load.sh
+++ /dev/null
@@ -1,3 +0,0 @@
-#! /bin/bash
-
-echo "10ee 6028" >   /sys/bus/pci/drivers/pciDriver/new_id
diff --git a/apps/pio_test.c b/apps/pio_test.c
new file mode 100644
index 0000000..84439ee
--- /dev/null
+++ b/apps/pio_test.c
@@ -0,0 +1,96 @@
+#define _BSD_SOURCE
+#define _POSIX_C_SOURCE 199309L
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <stdarg.h>
+#include <time.h>
+#include <sched.h>
+#include <sys/time.h>
+
+#include "pcilib.h"
+#include "irq.h"
+#include "kmem.h"
+
+#define DEVICE "/dev/fpga0"
+//#define REALTIME
+
+#define BAR PCILIB_BAR0
+#define BITS 32
+#define MASK ((1ll << BITS) - 1)
+
+
+#define WR(addr, value) { *(uint32_t*)(bar + addr) = value; }
+#define RD(addr, value) { value = *(uint32_t*)(bar + addr); }
+
+unsigned long long bits[BITS];
+
+int main(int argc, char *argv[]) {
+    uint32_t i, j;
+    pcilib_t *pci;
+    uint32_t reg, value, diff, errors;
+    void* volatile bar;
+    
+    unsigned long long attempts = 0, failures = 0;
+
+    if (argc < 2) {
+	printf("Usage: %s <register>\n", argv[0]);
+	exit(0);
+    }
+    
+    if (sscanf(argv[1], "%x", &reg) != 1) {
+	printf("Can't parse register %s\n", argv[1]);
+	exit(1);
+    }
+    
+#ifdef REALTIME
+    pid_t pid;
+    struct sched_param sched = {0};
+
+    pid = getpid();
+    sched.sched_priority = sched_get_priority_min(SCHED_FIFO);
+    if (sched_setscheduler(pid, SCHED_FIFO, &sched))
+	printf("Warning: not able to get real-time priority\n");
+#endif /* REALTIME */
+
+    pci = pcilib_open(DEVICE, PCILIB_MODEL_DETECT);
+    if (!pci) {
+	printf("pcilib_open\n");
+	exit(1);
+    }
+
+    bar = pcilib_map_bar(pci, BAR);
+    if (!bar) {
+	pcilib_close(pci);
+	printf("map bar\n");
+	exit(1);
+    }
+    
+    for (i = 0; 1; i++) {
+	WR(reg, (i%MASK));
+	RD(reg, value);
+	
+	attempts++;
+	if (value != (i%MASK)) {
+	    failures++;
+
+	    diff = value ^ (i%MASK);
+	    for (errors = 0, j = 0; j < BITS; j++)
+		if (diff&(1<<j)) errors++;
+	    bits[errors]++;
+		
+	    //printf("written: %x, read: %x\n", i, value);
+	}
+
+	if ((i % 1048576 ) == 0) {
+	    printf("Attempts %llu, failures %llu (", attempts, failures);
+	    for (j = 1; j < BITS; j++)
+		printf("%llu ", bits[j]);
+	    printf(")\n");
+	}
+
+    }
+
+    pcilib_unmap_bar(pci, BAR, bar);
+    pcilib_close(pci);
+}
diff --git a/apps/xilinx.c b/apps/xilinx.c
index 1ec31d7..757c388 100644
--- a/apps/xilinx.c
+++ b/apps/xilinx.c
@@ -2,6 +2,7 @@
 #define _POSIX_C_SOURCE 199309L
 #include <stdio.h>
 #include <stdlib.h>
+#include <string.h>
 #include <unistd.h>
 #include <stdarg.h>
 #include <time.h>
@@ -15,23 +16,27 @@
 #define DEVICE "/dev/fpga0"
 #define BAR PCILIB_BAR0
 #define USE PCILIB_KMEM_USE(PCILIB_KMEM_USE_USER, 1)
-#define BUFFERS 16
-#define ITERATIONS 16384
+#define STATIC_REGION 0x80000000 //  to reserve 512 MB at the specified address, add "memmap=512M$2G" to kernel parameters
+#define BUFFERS 1
+#define ITERATIONS 100
+#define TLP_SIZE 64
+#define HUGE_PAGE 4096	// number of pages per huge page
 #define PAGE_SIZE 4096	// other values are not supported in the kernel
 #define TIMEOUT 100000
 
 /* IRQs are slow for some reason. REALTIME mode is slower. Adding delays does not really help,
 otherall we have only 3 checks in average. Check ready seems to be not needed and adds quite 
 much extra time */
-//#define USE_IRQ
+#define USE_IRQ
 //#define CHECK_READY
 //#define REALTIME
 //#define ADD_DELAYS
+#define CHECK_RESULT
 
 //#define WR(addr, value) { val = value; pcilib_write(pci, BAR, addr, sizeof(val), &val); }
 //#define RD(addr, value) { pcilib_read(pci, BAR, addr, sizeof(val), &val); value = val; }
-#define WR(addr, value) { *(uint32_t*)(bar + addr) = value; }
-#define RD(addr, value) { value = *(uint32_t*)(bar + addr); }
+#define WR(addr, value) { *(uint32_t*)(bar + addr + offset) = value; }
+#define RD(addr, value) { value = *(uint32_t*)(bar + addr + offset); }
 
 static void fail(const char *msg, ...) {
     va_list va;
@@ -63,7 +68,7 @@ void hpsleep(size_t ns) {
 
 int main() {
     int err;
-    int i, j;
+    long i, j;
     pcilib_t *pci;
     pcilib_kmem_handle_t *kbuf;
     uint32_t status;
@@ -72,12 +77,15 @@ int main() {
     void* volatile bar;
     uintptr_t bus_addr[BUFFERS];
 
+    pcilib_bar_t bar_tmp = BAR; 
+    uintptr_t offset = 0;
+
     pcilib_kmem_flags_t clean_flags = PCILIB_KMEM_FLAG_HARDWARE|PCILIB_KMEM_FLAG_PERSISTENT|PCILIB_KMEM_FLAG_EXCLUSIVE;
 
 #ifdef ADD_DELAYS
     long rpt = 0, rpt2 = 0;
     size_t best_time;
-    best_time = 1000000000L * PAGE_SIZE / (4L * 1024 * 1024 * 1024);
+    best_time = 1000000000L * HUGE_PAGE * PAGE_SIZE / (4L * 1024 * 1024 * 1024);
 #endif /* ADD_DELAYS */
 
 #ifdef REALTIME
@@ -99,22 +107,46 @@ int main() {
 	fail("map bar");
     }
 
+    pcilib_detect_address(pci, &bar_tmp, &offset, 1);
+
+	// Reset
+    WR(0x00, 1)
+    usleep(1000);
+    WR(0x00, 0)
+
     pcilib_enable_irq(pci, PCILIB_IRQ_TYPE_ALL, 0);
     pcilib_clear_irq(pci, PCILIB_IRQ_SOURCE_DEFAULT);
 
     pcilib_clean_kernel_memory(pci, USE, clean_flags);
 
+#ifdef STATIC_REGION
+    kbuf = pcilib_alloc_kernel_memory(pci, PCILIB_KMEM_TYPE_REGION_C2S, BUFFERS, HUGE_PAGE * PAGE_SIZE, STATIC_REGION, USE, 0);
+#else /* STATIC_REGION */
+    kbuf = pcilib_alloc_kernel_memory(pci, PCILIB_KMEM_TYPE_DMA_C2S_PAGE, BUFFERS, HUGE_PAGE * PAGE_SIZE, 4096, USE, 0);
+#endif /* STATIC_REGION */
 
-    kbuf = pcilib_alloc_kernel_memory(pci, PCILIB_KMEM_TYPE_DMA_C2S_PAGE, BUFFERS, PAGE_SIZE, 4096, USE, 0);
+    if (!kbuf) {
+	printf("KMem allocation failed\n");
+	exit(0);
+    }
+
+
+#ifdef CHECK_RESULT    
+    volatile uint32_t *ptr0 = pcilib_kmem_get_block_ua(pci, kbuf, 0);
 
+    memset((void*)ptr0, 0, (HUGE_PAGE * PAGE_SIZE));
     
-    WR(0x00, 1)
-    usleep(1000);
-    WR(0x00, 0)
+    for (i = 0; i < (HUGE_PAGE * PAGE_SIZE / 4); i++) {
+	if (ptr0[i] != 0) break;
+    }
+    if (i < (HUGE_PAGE * PAGE_SIZE / 4)) {
+	printf("Initialization error in position %lu, value = %x\n", i * 4, ptr0[i]);
+    }
+#endif /* CHECK_RESULT */
+
     WR(0x04, 0)
-    
-    WR(0x0C, 0x20)
-    WR(0x10, (PAGE_SIZE / 0x80))
+    WR(0x0C, TLP_SIZE)
+    WR(0x10, (HUGE_PAGE * (PAGE_SIZE / (4 * TLP_SIZE))))
     WR(0x14, 0x13131313)
 
     for (j = 0; j < BUFFERS; j++ ) {
@@ -163,17 +195,31 @@ int main() {
     }
     gettimeofday(&end, NULL);
 
+#ifdef CHECK_RESULT    
+    pcilib_kmem_sync_block(pci, kbuf, PCILIB_KMEM_SYNC_FROMDEVICE, 0);
+
+    for (i = 0; i < (HUGE_PAGE * PAGE_SIZE / 4); i++) {
+//	printf("%lx ", ptr0[i]);
+	if (ptr0[i] != 0x13131313) break;
+    }
+    if (i < (HUGE_PAGE * PAGE_SIZE / 4)) {
+	printf("Error in position %lu, value = %x\n", i * 4, ptr0[i]);
+    }
+#endif /* CHECK_RESULT */
+
     pcilib_free_kernel_memory(pci, kbuf,  0);
     pcilib_disable_irq(pci, 0);
     pcilib_unmap_bar(pci, BAR, bar);
     pcilib_close(pci);
 
     run_time = (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_usec - start.tv_usec);
-    size = (long long int)ITERATIONS * BUFFERS * PAGE_SIZE;
+    size = (long long int)ITERATIONS * BUFFERS * HUGE_PAGE * PAGE_SIZE;
 
     printf("%.3lf GB/s: transfered %zu bytes in %zu us using %u buffers\n", 1000000. * size / run_time / 1024 / 1024 / 1024, size, run_time, BUFFERS);
 
 # ifdef ADD_DELAYS
     printf("Repeats: %lf, %lf\n",1. * rpt / (ITERATIONS * BUFFERS), 1. * rpt2 / (ITERATIONS * BUFFERS));
 #endif /* USE_IRQ */	    
+
+
 }
diff --git a/apps/xilinx2.c b/apps/xilinx2.c
deleted file mode 100644
index 757c388..0000000
--- a/apps/xilinx2.c
+++ /dev/null
@@ -1,225 +0,0 @@
-#define _BSD_SOURCE
-#define _POSIX_C_SOURCE 199309L
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <unistd.h>
-#include <stdarg.h>
-#include <time.h>
-#include <sched.h>
-#include <sys/time.h>
-
-#include "pcilib.h"
-#include "irq.h"
-#include "kmem.h"
-
-#define DEVICE "/dev/fpga0"
-#define BAR PCILIB_BAR0
-#define USE PCILIB_KMEM_USE(PCILIB_KMEM_USE_USER, 1)
-#define STATIC_REGION 0x80000000 //  to reserve 512 MB at the specified address, add "memmap=512M$2G" to kernel parameters
-#define BUFFERS 1
-#define ITERATIONS 100
-#define TLP_SIZE 64
-#define HUGE_PAGE 4096	// number of pages per huge page
-#define PAGE_SIZE 4096	// other values are not supported in the kernel
-#define TIMEOUT 100000
-
-/* IRQs are slow for some reason. REALTIME mode is slower. Adding delays does not really help,
-otherall we have only 3 checks in average. Check ready seems to be not needed and adds quite 
-much extra time */
-#define USE_IRQ
-//#define CHECK_READY
-//#define REALTIME
-//#define ADD_DELAYS
-#define CHECK_RESULT
-
-//#define WR(addr, value) { val = value; pcilib_write(pci, BAR, addr, sizeof(val), &val); }
-//#define RD(addr, value) { pcilib_read(pci, BAR, addr, sizeof(val), &val); value = val; }
-#define WR(addr, value) { *(uint32_t*)(bar + addr + offset) = value; }
-#define RD(addr, value) { value = *(uint32_t*)(bar + addr + offset); }
-
-static void fail(const char *msg, ...) {
-    va_list va;
-    
-    va_start(va, msg);
-    vprintf(msg, va);
-    va_end(va);
-    printf("\n");
-
-    exit(-1);
-}
-
-void hpsleep(size_t ns) {
-    struct timespec wait, tv;
-
-    clock_gettime(CLOCK_REALTIME, &wait);
-
-    wait.tv_nsec += ns;
-    if (wait.tv_nsec > 999999999) {
-	wait.tv_sec += 1;
-	wait.tv_nsec = 1000000000 - wait.tv_nsec;
-    }
-
-    do {
-	clock_gettime(CLOCK_REALTIME, &tv);
-    } while ((wait.tv_sec > tv.tv_sec)||((wait.tv_sec == tv.tv_sec)&&(wait.tv_nsec > tv.tv_nsec)));
-}
-
-
-int main() {
-    int err;
-    long i, j;
-    pcilib_t *pci;
-    pcilib_kmem_handle_t *kbuf;
-    uint32_t status;
-    struct timeval start, end;
-    size_t size, run_time;
-    void* volatile bar;
-    uintptr_t bus_addr[BUFFERS];
-
-    pcilib_bar_t bar_tmp = BAR; 
-    uintptr_t offset = 0;
-
-    pcilib_kmem_flags_t clean_flags = PCILIB_KMEM_FLAG_HARDWARE|PCILIB_KMEM_FLAG_PERSISTENT|PCILIB_KMEM_FLAG_EXCLUSIVE;
-
-#ifdef ADD_DELAYS
-    long rpt = 0, rpt2 = 0;
-    size_t best_time;
-    best_time = 1000000000L * HUGE_PAGE * PAGE_SIZE / (4L * 1024 * 1024 * 1024);
-#endif /* ADD_DELAYS */
-
-#ifdef REALTIME
-    pid_t pid;
-    struct sched_param sched = {0};
-
-    pid = getpid();
-    sched.sched_priority = sched_get_priority_min(SCHED_FIFO);
-    if (sched_setscheduler(pid, SCHED_FIFO, &sched))
-	printf("Warning: not able to get real-time priority\n");
-#endif /* REALTIME */
-
-    pci = pcilib_open(DEVICE, PCILIB_MODEL_DETECT);
-    if (!pci) fail("pcilib_open");
-
-    bar = pcilib_map_bar(pci, BAR);
-    if (!bar) {
-	pcilib_close(pci);
-	fail("map bar");
-    }
-
-    pcilib_detect_address(pci, &bar_tmp, &offset, 1);
-
-	// Reset
-    WR(0x00, 1)
-    usleep(1000);
-    WR(0x00, 0)
-
-    pcilib_enable_irq(pci, PCILIB_IRQ_TYPE_ALL, 0);
-    pcilib_clear_irq(pci, PCILIB_IRQ_SOURCE_DEFAULT);
-
-    pcilib_clean_kernel_memory(pci, USE, clean_flags);
-
-#ifdef STATIC_REGION
-    kbuf = pcilib_alloc_kernel_memory(pci, PCILIB_KMEM_TYPE_REGION_C2S, BUFFERS, HUGE_PAGE * PAGE_SIZE, STATIC_REGION, USE, 0);
-#else /* STATIC_REGION */
-    kbuf = pcilib_alloc_kernel_memory(pci, PCILIB_KMEM_TYPE_DMA_C2S_PAGE, BUFFERS, HUGE_PAGE * PAGE_SIZE, 4096, USE, 0);
-#endif /* STATIC_REGION */
-
-    if (!kbuf) {
-	printf("KMem allocation failed\n");
-	exit(0);
-    }
-
-
-#ifdef CHECK_RESULT    
-    volatile uint32_t *ptr0 = pcilib_kmem_get_block_ua(pci, kbuf, 0);
-
-    memset((void*)ptr0, 0, (HUGE_PAGE * PAGE_SIZE));
-    
-    for (i = 0; i < (HUGE_PAGE * PAGE_SIZE / 4); i++) {
-	if (ptr0[i] != 0) break;
-    }
-    if (i < (HUGE_PAGE * PAGE_SIZE / 4)) {
-	printf("Initialization error in position %lu, value = %x\n", i * 4, ptr0[i]);
-    }
-#endif /* CHECK_RESULT */
-
-    WR(0x04, 0)
-    WR(0x0C, TLP_SIZE)
-    WR(0x10, (HUGE_PAGE * (PAGE_SIZE / (4 * TLP_SIZE))))
-    WR(0x14, 0x13131313)
-
-    for (j = 0; j < BUFFERS; j++ ) {
-        bus_addr[j] = pcilib_kmem_get_block_ba(pci, kbuf, j);
-    }
-
-    gettimeofday(&start, NULL);
-
-    for (i = 0; i < ITERATIONS; i++) {
-	for (j = 0; j < BUFFERS; j++ ) {
-//	    uintptr_t ba = pcilib_kmem_get_block_ba(pci, kbuf, j);
-//	    WR(0x08, ba)
-	    WR(0x08, bus_addr[j]);
-	    WR(0x04, 0x01)
-
-#ifdef USE_IRQ
-	    err = pcilib_wait_irq(pci, PCILIB_IRQ_SOURCE_DEFAULT, TIMEOUT, NULL);
-	    if (err) printf("Timeout waiting for IRQ, err: %i\n", err);
-
-	    RD(0x04, status);
-	    if ((status&0xFFFF) != 0x101) printf("Invalid status %x\n", status);
-//	    WR(0x04, 0x00);
-#else /* USE_IRQ */
-# ifdef ADD_DELAYS
-//	    hpsleep(best_time);
-	    do {
-		rpt++;
-		RD(0x04, status);
-	    } while (status != 0x101);
-# else /* ADD_DELAYS */
-	    do {
-		RD(0x04, status);
-	    } while (status != 0x101);
-# endif /* ADD_DELAYS */
-#endif /* USE_IRQ */
-
-	    WR(0x00, 1)
-#ifdef CHECK_READY
-	    do {
-		rpt2++;
-		RD(0x04, status);
-	    } while (status != 0);
-#endif /* CHECK_READY */
-	    WR(0x00, 0)
-	}
-    }
-    gettimeofday(&end, NULL);
-
-#ifdef CHECK_RESULT    
-    pcilib_kmem_sync_block(pci, kbuf, PCILIB_KMEM_SYNC_FROMDEVICE, 0);
-
-    for (i = 0; i < (HUGE_PAGE * PAGE_SIZE / 4); i++) {
-//	printf("%lx ", ptr0[i]);
-	if (ptr0[i] != 0x13131313) break;
-    }
-    if (i < (HUGE_PAGE * PAGE_SIZE / 4)) {
-	printf("Error in position %lu, value = %x\n", i * 4, ptr0[i]);
-    }
-#endif /* CHECK_RESULT */
-
-    pcilib_free_kernel_memory(pci, kbuf,  0);
-    pcilib_disable_irq(pci, 0);
-    pcilib_unmap_bar(pci, BAR, bar);
-    pcilib_close(pci);
-
-    run_time = (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_usec - start.tv_usec);
-    size = (long long int)ITERATIONS * BUFFERS * HUGE_PAGE * PAGE_SIZE;
-
-    printf("%.3lf GB/s: transfered %zu bytes in %zu us using %u buffers\n", 1000000. * size / run_time / 1024 / 1024 / 1024, size, run_time, BUFFERS);
-
-# ifdef ADD_DELAYS
-    printf("Repeats: %lf, %lf\n",1. * rpt / (ITERATIONS * BUFFERS), 1. * rpt2 / (ITERATIONS * BUFFERS));
-#endif /* USE_IRQ */	    
-
-
-}