add interface between scale factor mem and L1

Author: Amanda Shi

src/main/scala/gemmini/ConfigsFP.scala

@@ -273,8 +273,8 @@ object GemminiMxFPConfigs {
                                  Seq.fill(4) {MxFloat(8, 8, 4, true, false)},
 
                                  // 16x16 mesh with varying precisions
-    scaleMem_data_width = 128,
-    scaleMem_bank_entries = 8192,
+    // scaleMem_data_width = 128,
+    // scaleMem_write_data_addr_width = 32,
     scaleSize = 32,
     enable_lut = true,
     mvin_scale_args = None,

src/main/scala/gemmini/Controller.scala

@@ -180,12 +180,12 @@ class GemminiModule[T <: Data: Arithmetic, U <: Data, V <: Data]
   val mx_requantizer = Option.when(outer.config.use_mx_scaling && outer.config.requantizer.isDefined && outer.config.lut.isDefined) {
     val q = outer.config.requantizer.get
     val l = outer.config.lut.get
-   
+    
     Module(new MxRequantizer(
       sp_data_width = outer.config.sp_width,
       sp_addr_width = log2Ceil(outer.config.sp_bank_entries),
-      scaleMem_data_width = outer.config.scaleMem_data_width,
-      scaleMem_addr_width = log2Ceil(outer.config.scaleMem_bank_entries),
+      scaleMem_data_width = outer.config.scale_mem.get.ScaleMemWriteDataWidth,
+      scaleMem_addr_width = outer.config.scale_mem.get.ScaleMemWriteAddrWidth,
       scaleSize = outer.config.scaleSize,
       scaleMembasewrite = 0, // TODO: add this into the instruction
       lutConfig = l,
@@ -217,7 +217,7 @@ class GemminiModule[T <: Data: Arithmetic, U <: Data, V <: Data]
       val lut0 = Flipped(Decoupled(new QuantLutWriteBundle(l)))
       val lut1 = Flipped(Decoupled(new QuantLutWriteBundle(l)))
       val lut2 = Flipped(Decoupled(new QuantLutWriteBundle(l)))
-      //val scaleFactorOut = Decoupled(new ScalingFactorWriteReq(scaleMem_addr_width, scaleMem_data_width)) 
+      //val scaleFactorOut = Decoupled(new ScalingFactorWriteReq(s.ScaleMemWriteAddrWidth, s.ScaleMemWriteAddrWidth)) 
     })
 
 

src/main/scala/gemmini/GemminiConfigs.scala

@@ -56,8 +56,8 @@ case class GemminiArrayConfig[T <: Data : Arithmetic, U <: Data, V <: Data](
                                                                              acc_latency: Int = 2,
 
 
-                                                                             scaleMem_data_width: Int = 128,
-                                                                             scaleMem_bank_entries: Int = 8192,
+                                                                            //  scaleMem_write_data_width: Int = 128,
+                                                                            //  scaleMem_write_data_addr_width: Int = 32,
                                                                              scaleSize: Int = 32,
 
                                                                              dma_maxbytes: Int = 64, // TODO get this from cacheblockbytes

src/main/scala/gemmini/MxConfigFragments.scala

@@ -10,6 +10,8 @@ case class GemminiScalingFactorMemConfig(
   subbanksPerBank: Int = 2,
   gpuInputWidthBytes: Int = 8,
   numBanks: Int = 8,
+  ScaleMemWriteDataWidth: Int = 256,
+  ScaleMemWriteAddrWidth: Int = 32,
 ) {
   def depth: Int = (sizeInBytes / (subbankLineSizeInBytes) / numBanks).toInt
   def bankWidthBytes = subbankLineSizeInBytes * subbanksPerBank

src/main/scala/gemmini/MxRequantizer.scala

@@ -346,11 +346,11 @@ class MxRequantizer[T <: Data: Arithmetic](
       scale_buffer_full := true.B
     }.otherwise {
       scale_write_counter := scale_write_counter + 1.U
-      when(io.scaleMem_write.ready){
+      when(io.scaleMem_write.fire){
         scale_buffer_full := false.B
       }
     }
-  }.elsewhen(io.scaleMem_write.ready){
+  }.elsewhen(io.scaleMem_write.fire){
       scale_buffer_full := false.B
   }