Merge remote-tracking branch 'origin/dev' into chisel6

CHIPYARD.hash

@@ -1 +1 @@
-ef3409f87ff2988fa862ea48c995d2c27c93c7a2
+b4aae0ddfdc5aaced32e0df90b633eab5b8327ca

src/main/scala/gemmini/AccumulatorScale.scala

@@ -397,7 +397,7 @@ object AccumulatorScale {
  val neg_q_iexp = neg(q)
  val z_iexp = (neg_q_iexp * qln2_inv).asUInt.do_>>(16).asTypeOf(q) // q is non-positive
  val z_iexp_saturated = Wire(z_iexp.cloneType)
- z_iexp_saturated := Mux((5 until 16).map(z_iexp.asUInt(_)).reduce(_ | _), 32.S, z_iexp)
+ z_iexp_saturated := Mux((5 until 16).map(z_iexp.asUInt(_)).reduce(_ | _), 32.S.asTypeOf(z_iexp), z_iexp)
  val qp_iexp = q.mac(z_iexp, qln2).withWidthOf(q)
  val q_poly_iexp = qc.mac(qp_iexp + qb, qp_iexp + qb).withWidthOf(q)
  // we dont want a rounding shift

src/main/scala/gemmini/ConfigsFP.scala

@@ -109,8 +109,30 @@ object GemminiFPConfigs {
  mvin_scale_acc_args = Some(ScaleArguments((t: Float, u: Float) => t * u, 4, Float(8, 24), -1, identity = "1.0", c_str="((x) * (scale))")),
  )
 
-}
 
+ val chipFP32Config = FP32DefaultConfig.copy(sp_capacity=CapacityInKilobytes(32), acc_capacity=CapacityInKilobytes(8), dataflow=Dataflow.WS,
+ acc_scale_args = Some(ScaleArguments((t: Float, u: Float) => {t}, 1, Float(8, 24), -1, identity = "1.0",
+ c_str = "((x))"
+ )),
+ mvin_scale_args = Some(ScaleArguments((t: Float, u: Float) => t * u, 3, Float(8, 24), -1, identity = "1.0", c_str="((x) * (scale))")),
+ mvin_scale_acc_args=None,
+ acc_singleported=false,
+ acc_sub_banks = 1,
+ acc_banks = 2,
+ mesh_output_delay = 2,
+ tile_latency = 1,
+ acc_latency = 3,
+ ex_read_from_acc=false,
+ ex_write_to_spad=false,
+ has_training_convs = false,
+ hardcode_d_to_garbage_addr = true,
+ acc_read_full_width = false,
+ max_in_flight_mem_reqs = 16,
+ headerFileName = "gemmini_params_fp32.h",
+ num_counter = 0,
+ clock_gate = true 
+ )
+}
 
 //===========FP32 Default Config=========
 class GemminiFP32DefaultConfig extends Config((site, here, up) => {
@@ -123,6 +145,18 @@ class GemminiFP32DefaultConfig extends Config((site, here, up) => {
  )
 })
 
+class ChipFP32GemminiConfig[T <: Data : Arithmetic, U <: Data, V <: Data](
+ gemminiConfig: GemminiArrayConfig[T,U,V] = GemminiFPConfigs.chipFP32Config
+) extends Config((site, here, up) => {
+ case BuildRoCC => up(BuildRoCC) ++ Seq(
+ (p: Parameters) => {
+ implicit val q = p
+ val gemmini = LazyModule(new Gemmini(gemminiConfig))
+ gemmini
+ }
+ )
+})
+
 
 //===========FP16 Default Config=========
 class GemminiFP16DefaultConfig extends Config((site, here, up) => {

src/main/scala/gemmini/VectorScalarMultiplier.scala

@@ -198,6 +198,8 @@ object VectorScalarMultiplier {
  ) = {
  assert(!is_acc || is_mvin)
  val vsm = Module(new VectorScalarMultiplier(scale_args, cols, t, tag_t))
- (vsm.io.req, vsm.io.resp)
+ val vsm_in_q = Module(new Queue(chiselTypeOf(vsm.io.req.bits), 2))
+ vsm.io.req <> vsm_in_q.io.deq
+ (vsm_in_q.io.enq, vsm.io.resp) 
  }
 }