diff --git a/src/memory/memory.h b/src/memory/memory.h
index 3566d22e3..552ed43d8 100644
--- a/src/memory/memory.h
+++ b/src/memory/memory.h
@@ -72,6 +72,14 @@ extern void (*writememd[0x10000])(void);
 
 #endif
 
+#ifdef __arm__
+    void __aeabi_memcpy4(void*, void*, uint32_t);		//this is present in libc6 on Raspberry PI
+    #define MEMCPY4(...) __aeabi_memcpy4(__VA_ARGS__)
+#else
+    #define MEMCPY4(...) memcpy(__VA_ARGS__)
+#endif
+
+
 static void masked_write(uint32_t* dst, uint32_t value, uint32_t mask)
 {
     *dst = (*dst & ~mask) | (value & mask);
diff --git a/src/pi/pi_controller.c b/src/pi/pi_controller.c
index 9c12cc15a..49b6f6d09 100644
--- a/src/pi/pi_controller.c
+++ b/src/pi/pi_controller.c
@@ -126,9 +126,17 @@ static void dma_pi_write(struct pi_controller* pi)
     dram = (uint8_t*)pi->ri->rdram.dram;
     rom = pi->cart_rom.rom;
 
-    for(i = 0; i < longueur; ++i)
+    // Check for 32-bit word alignment for faster copying
+    if (((dram_address | rom_address | longueur) & 3) != 0)
     {
-        dram[(dram_address+i)^S8] = rom[(rom_address+i)^S8];
+        for(i = 0; i < longueur; ++i)
+        {
+            dram[(dram_address+i)^S8] = rom[(rom_address+i)^S8];
+        }
+    }
+    else
+    {
+        MEMCPY4(&dram[dram_address], &rom[rom_address], longueur);
     }
 
     invalidate_r4300_cached_code(0x80000000 + dram_address, longueur);
diff --git a/src/rsp/rsp_core.c b/src/rsp/rsp_core.c
index d6aa0744f..374844c06 100644
--- a/src/rsp/rsp_core.c
+++ b/src/rsp/rsp_core.c
@@ -40,20 +40,35 @@ static void dma_sp_write(struct rsp_core* sp)
     unsigned int length = ((l & 0xfff) | 7) + 1;
     unsigned int count = ((l >> 12) & 0xff) + 1;
     unsigned int skip = ((l >> 20) & 0xfff);
- 
+
     unsigned int memaddr = sp->regs[SP_MEM_ADDR_REG] & 0xfff;
     unsigned int dramaddr = sp->regs[SP_DRAM_ADDR_REG] & 0xffffff;
 
     unsigned char *spmem = (unsigned char*)sp->mem + (sp->regs[SP_MEM_ADDR_REG] & 0x1000);
     unsigned char *dram = (unsigned char*)sp->ri->rdram.dram;
 
-    for(j=0; j<count; j++) {
-        for(i=0; i<length; i++) {
-            spmem[memaddr^S8] = dram[dramaddr^S8];
-            memaddr++;
-            dramaddr++;
+    // Check for 32-bit word alignment for faster copying
+    if (((memaddr | dramaddr | length | skip) & 3) != 0)
+    {
+        for(j=0; j<count; j++)
+        {
+            for(i=0; i<length; i++)
+            {
+                spmem[memaddr^S8] = dram[dramaddr^S8];
+                memaddr++;
+                dramaddr++;
+            }
+            dramaddr+=skip;
+        }
+    }
+    else
+    {
+        for(j=0; j<count; j++)
+        {
+            MEMCPY4(&spmem[memaddr], &dram[dramaddr], length);
+            memaddr += length;
+            dramaddr += skip + length;
         }
-        dramaddr+=skip;
     }
 }
 
@@ -73,13 +88,28 @@ static void dma_sp_read(struct rsp_core* sp)
     unsigned char *spmem = (unsigned char*)sp->mem + (sp->regs[SP_MEM_ADDR_REG] & 0x1000);
     unsigned char *dram = (unsigned char*)sp->ri->rdram.dram;
 
-    for(j=0; j<count; j++) {
-        for(i=0; i<length; i++) {
-            dram[dramaddr^S8] = spmem[memaddr^S8];
-            memaddr++;
-            dramaddr++;
+    // Check for 32-bit word alignment for faster copying
+    if (((dramaddr | memaddr | skip | length) & 3) != 0)
+    {
+        for(j=0; j<count; j++)
+        {
+            for(i=0; i<length; i++)
+            {
+                dram[dramaddr^S8] = spmem[memaddr^S8];
+                memaddr++;
+                dramaddr++;
+            }
+            dramaddr+=skip;
+        }
+    }
+    else
+    {
+        for(j=0; j<count; j++)
+        {
+            MEMCPY4(&dram[dramaddr], &spmem[memaddr], length);
+            memaddr += length;
+            dramaddr += length + skip;
         }
-        dramaddr+=skip;
     }
 }
 
@@ -307,7 +337,6 @@ void do_SP_Task(struct rsp_core* sp)
             add_interupt_event(SP_INT, 4000/*500*/);
         sp->r4300->mi.regs[MI_INTR_REG] &= ~MI_INTR_SP;
         sp->regs[SP_STATUS_REG] &= ~0x300;
-        
     }
     else
     {