fsqrt ok

libmodules/bfloat16nncore.py

@@ -201,8 +201,7 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
                 NextValue(self.Loader_Delay, self.Loader_Delay + 1), # Increment
             )
         )
-        """            
-            fpu.fsqrt
+        """                        
             fpu.fmadd
             fpu.fmsub
             fpu.fnmadd
@@ -212,14 +211,14 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
         """
         Loader_fsm.act("Loader_EXEC1",            
             NextValue(self.b16Status.storage[5], True), # Current status added
-            NextValue(fpu.fdiv, True), # This command requested            
+            NextValue(fpu.fsqrt, True), # This command requested            
             NextValue(fpu.fready, False), # Engage trigger
             NextState("Loader_EXEC2")            
         )
         Loader_fsm.act("Loader_EXEC2",
             NextValue(self.b16Status.storage[6], True), # Current status added
             If(fpu.fready,                
-                NextValue(fpu.fdiv, False), # Clear command request
+                NextValue(fpu.fsqrt, False), # Clear command request
                 NextValue(self.b16Result.storage, fpu.fresult[16:32]), # Pick result (little endian, high word!)
                 NextValue(self.b16Status.storage[15], True), # Indicate readyness ...
                 NextValue(self.bReady, True), # Indicate readyness (LED on!)

software/source/bfloat16nnlib.c

@@ -121,7 +121,7 @@ int key_eval(void)
 				for(i=0, ui32ptr = (uint32_t *)DRAMDATABASE;i<DRAMDATASIZE;i++) // Setup test data
 					*ui32ptr++ = i+1;		
 				ui16ptr = (uint16_t *)(DRAMDATABASE + 0 * sizeof(float));
-				*ui16ptr++ = f2ui16(1.0);
+				*ui16ptr++ = f2ui16(4.0);
 				*ui16ptr++ = f2ui16(2.0);	
 				*ui16ptr++ = f2ui16(3.0);
 				if(fpgaload((uint32_t *)DRAMDATABASE, 512)) {