test lut mmio

Author: Amanda Shi

src/main/scala/gemmini/Controller.scala

@@ -257,11 +257,11 @@ class GemminiModule[T <: Data: Arithmetic, U <: Data, V <: Data]
         mx_sel(bank) := (ex_controller.io.output_MxFormat === 1.U)
       }
     }
-    if(outer.config.lut.isDefined){
-      mx_requantizer.get.io.lut0_write := DontCare
-      mx_requantizer.get.io.lut1_write := DontCare
-      mx_requantizer.get.io.lut2_write := DontCare
-    }
+    // if(outer.config.lut.isDefined){
+    //   mx_requantizer.get.io.lut0_write := DontCare
+    //   mx_requantizer.get.io.lut1_write := DontCare
+    //   mx_requantizer.get.io.lut2_write := DontCare
+    // }
   }
 
   /* 

src/main/scala/gemmini/MxConfigFragments.scala

@@ -29,14 +29,14 @@ case class GemminiRequantizerConfig(
   minOutputBits: Int = 4,
   maxOutputBits: Int = 8,
   outputIdBits: Int = 3,
-  lutUpdateRegularityW : Int = 128, 
+  lutUpdateRegularityW : Int = 128,  // means how many elements update once the lut
   lutUpdateRegularityActIn : Int = 128,
   lutUpdateRegularityActOut : Int = 128,  
 )
 
 case class GemminiLUTConfig(
   numBits: Int = 96,
-  numEntries: Int = 32,
+  numEntries: Int = 1,
   numTables: Int = 3,
   rdataWidth: Int = 6,
   raddrWidth: Int = 4, 

src/main/scala/gemmini/QuantLut.scala

@@ -62,15 +62,17 @@ class QuantLut(
     when(lutCache_act_in_flag === false.B){
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_act_in_0(lane)(entry) := io.lut_write_act_in.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_act_in_0(lane)(entry) := io.lut_write_act_in.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_act_in_0(lane)(entry) := io.lut_write_act_in.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
       }
       lutCache_act_in_flag := ~lutCache_act_in_flag
       lutCache_act_in_buffer_0_read_enable := true.B
     }
     }.otherwise {
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_act_in_1(lane)(entry) := io.lut_write_act_in.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_act_in_1(lane)(entry) := io.lut_write_act_in.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_act_in_1(lane)(entry) := io.lut_write_act_in.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
       }
       lutCache_act_in_flag := ~lutCache_act_in_flag
       lutCache_act_in_buffer_1_read_enable := true.B
@@ -109,15 +111,17 @@ class QuantLut(
     when(lutCache_weight_flag === false.B){
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_weight_0(lane)(entry) := io.lut_write_weight.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_weight_0(lane)(entry) := io.lut_write_weight.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_weight_0(lane)(entry) := io.lut_write_weight.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
         }
       }
       lutCache_weight_flag := ~lutCache_weight_flag
       lutCache_weight_buffer_0_read_enable := true.B
     }.otherwise {
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_weight_1(lane)(entry) := io.lut_write_weight.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_weight_1(lane)(entry) := io.lut_write_weight.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_weight_1(lane)(entry) := io.lut_write_weight.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
         }
       }
       lutCache_weight_flag := ~lutCache_weight_flag
@@ -155,15 +159,17 @@ class QuantLut(
     when(lutCache_act_out_flag === false.B){
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_act_out_0(lane)(entry) := io.lut_write_act_out.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_act_out_0(lane)(entry) := io.lut_write_act_out.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_act_out_0(lane)(entry) := io.lut_write_act_out.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
         }
       }
       lutCache_act_out_flag := ~lutCache_act_out_flag
       lutCache_act_out_buffer_0_read_enable := true.B
     }.otherwise {
       for (lane <- 0 until 32) {
         for (entry <- 0 until 16) {
-          lutCache_act_out_1(lane)(entry) := io.lut_write_act_out.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          //lutCache_act_out_1(lane)(entry) := io.lut_write_act_out.bits.data(lane)((entry+1)*rdataWidth-1, entry*rdataWidth)
+          lutCache_act_out_1(lane)(entry) := io.lut_write_act_out.bits.data(0)((entry+1)*rdataWidth-1, entry*rdataWidth)
         }
       }
       lutCache_act_out_flag := ~lutCache_act_out_flag