Index: apps/plugins/test_mem.c
===================================================================
--- apps/plugins/test_mem.c	(revision 25893)
+++ apps/plugins/test_mem.c	(working copy)
@@ -23,62 +23,135 @@
 
 PLUGIN_HEADER
 
-#define BUF_SIZE ((PLUGIN_BUFFER_SIZE-(10<<10)) / (sizeof(int)))
-#define LOOP_REPEAT 8
-static volatile int buf[BUF_SIZE];
+#define BUF_SIZE (1<<13) /* 32 KB = (1<<13)*sizeof(int) */
 
+#define LOOP_REPEAT_DRAM 256
+static volatile int buf_dram[BUF_SIZE];
+
+#if defined(USE_IRAM)
+#define LOOP_REPEAT_IRAM 512
+static volatile int buf_iram[BUF_SIZE] IBSS_ATTR;
+#endif
+
 /* (Byte per loop * loops * ticks per s / ticks)>>10 = KB per s */
-#define KB_PER_SEC(delta) (((BUF_SIZE*sizeof(buf[0])*LOOP_REPEAT*HZ)/delta) >> 10)
+#define KB_PER_SEC(cnt, delta) (((BUF_SIZE*sizeof(int)*cnt*HZ)/delta) >> 10)
 
+void write_test(volatile int *buf, int buf_size, int loop_cnt, 
+                int line, char *ramtype)
+{
+    int delta;
+    int last_tick = *rb->current_tick;
+    
+#if defined(CPU_ARM)
+    asm volatile (
+            "mov r6, %[loops]      \n"
+        ".outer_loop_write:        \n"
+            "mov r4, %[buf_p]      \n"
+            "mov r5, %[size]       \n"
+        ".inner_loop_write:        \n"
+            "ldmia r4!, {r0-r3}    \n"
+            "ldmia r4!, {r0-r3}    \n"
+            "subs r5, r5, #8       \n"
+            "bgt .inner_loop_write \n"
+            "subs r6, r6, #1       \n"
+            "bgt .outer_loop_write \n"
+            :
+            : [loops] "r" (loop_cnt), [size] "r" (buf_size), [buf_p] "r" (buf)
+            : "r0", "r1", "r2", "r3", "r4", "r5", "r6"
+        );
+#else
+    int i, j;
+    for(i = 0; i < loop_cnt; i++)
+    {
+        for (j = 0; j < buf_size; j+=4)
+        {
+            buf[j  ] = j;
+            buf[j+1] = j+1;
+            buf[j+2] = j+2;
+            buf[j+3] = j+3;
+        }
+    }
+#endif
+    delta = *rb->current_tick - last_tick;
+    delta = delta>0 ? delta : delta+1;
+
+    rb->screens[0]->putsf(0, line, "%s write: %6d KB/s (%2d ticks)", 
+                          ramtype, KB_PER_SEC(loop_cnt, delta), delta);
+}
+
+void read_test(volatile int *buf, int buf_size, int loop_cnt, 
+               int line, char *ramtype)
+{
+    int delta;
+    int last_tick = *rb->current_tick;
+    
+#if defined(CPU_ARM)
+    asm volatile (
+            "mov r0, #0           \n"
+            "mov r1, #1           \n"
+            "mov r2, #2           \n"
+            "mov r3, #3           \n"
+            "mov r6, %[loops]     \n"
+        ".outer_loop_read:        \n"
+            "mov r4, %[buf_p]     \n"
+            "mov r5, %[size]      \n"
+        ".inner_loop_read:        \n"
+            "stmia r4!, {r0-r3}   \n"
+            "stmia r4!, {r0-r3}   \n"
+            "subs r5, r5, #8      \n"
+            "bgt .inner_loop_read \n"
+            "subs r6, r6, #1      \n"
+            "bgt .outer_loop_read \n"
+            :
+            : [loops] "r" (loop_cnt), [size] "r" (buf_size), [buf_p] "r" (buf)
+            : "r0", "r1", "r2", "r3", "r4", "r5", "r6"
+        );
+#else
+    int i, j, x;
+    for(i = 0; i < loop_cnt; i++)
+    {
+        for(j = 0; j < buf_size; j+=4)
+        {
+            x = buf[j  ];
+            x = buf[j+2];
+            x = buf[j+3];
+            x = buf[j+4];
+        }
+    }
+#endif
+    delta = *rb->current_tick - last_tick;
+    delta = delta>0 ? delta : delta+1;
+
+    rb->screens[0]->putsf(0, line, "%s read : %6d KB/s (%2d ticks)", 
+                          ramtype, KB_PER_SEC(loop_cnt, delta), delta);
+}
+
+
 enum plugin_status plugin_start(const void* parameter)
 {
     (void)parameter;
     bool done = false;
     bool boost = false;
     int count = 0;
-    int last_tick = 0;
 
     rb->lcd_setfont(FONT_SYSFIXED);
 
     while (!done)
     {
-        unsigned i, j;
         int line = 0;
-        int x;
-        int delta;
-        last_tick = *rb->current_tick;
 
-        for(i = 0; i < LOOP_REPEAT; i++)
-        {
-            for (j = 0; j < BUF_SIZE; j+=4)
-            {
-                buf[j  ] = j;
-                buf[j+1] = j+1;
-                buf[j+2] = j+2;
-                buf[j+3] = j+3;
-            }
-        }
-        delta = *rb->current_tick - last_tick;
         rb->screens[0]->clear_display();
         rb->screens[0]->putsf(0, line++, "%s", boost?"boosted":"unboosted");
-        rb->screens[0]->putsf(0, line++, "bufsize: %u", BUF_SIZE*sizeof(buf[0]));
+        rb->screens[0]->putsf(0, line++, "bufsize: %u", BUF_SIZE*sizeof(int));
         rb->screens[0]->putsf(0, line++, "loop#: %d", ++count);
-        rb->screens[0]->putsf(0, line++, "write ticks: %2d (%5d KB/s)", delta,
-                              KB_PER_SEC(delta));
-        last_tick = *rb->current_tick;
-        for(i = 0; i < LOOP_REPEAT; i++)
-        {
-            for(j = 0; j < BUF_SIZE; j+=4)
-            {
-                x = buf[j  ];
-                x = buf[j+2];
-                x = buf[j+3];
-                x = buf[j+4];
-            }
-        }
-        delta = *rb->current_tick - last_tick;
-        rb->screens[0]->putsf(0, line++, "read ticks : %2d (%5d KB/s)", delta,
-                              KB_PER_SEC(delta));
+
+        read_test (buf_dram, BUF_SIZE, LOOP_REPEAT_DRAM, line++, "(DRAM)");
+        write_test(buf_dram, BUF_SIZE, LOOP_REPEAT_DRAM, line++, "(DRAM)");
+#if defined(USE_IRAM)
+        read_test (buf_iram, BUF_SIZE, LOOP_REPEAT_IRAM, line++, "(IRAM)");
+        write_test(buf_iram, BUF_SIZE, LOOP_REPEAT_IRAM, line++, "(IRAM)");
+#endif
+
         rb->screens[0]->update();
 
         switch (rb->get_action(CONTEXT_STD, TIMEOUT_NOBLOCK))