Merge pull request #269 from ucb-bar/tweaks

Two tweaks to Rocket perf & QoR

csrc/emulator.cc

@@ -15,6 +15,7 @@
 extern dtm_t* dtm;
 static uint64_t trace_count = 0;
 bool verbose;
+bool done_reset;
 
 void handle_sigterm(int sig)
 {
@@ -89,6 +90,7 @@ int main(int argc, char** argv)
     tile->eval();
     tile->reset = 0;
   }
+  done_reset = true;
 
   while (!dtm->done() && !tile->io_success && trace_count < max_cycles) {
     tile->clk = 0;

csrc/verilator.h

@@ -6,6 +6,7 @@
 #include <stdio.h>
 
 extern bool verbose;
+extern bool done_reset;
 
 class VerilatedVcdFILE : public VerilatedVcdFile {
  public:

emulator/Makefrag-verilator

@@ -49,7 +49,9 @@ verilator/verilator-$(VERILATOR_VERSION).tar.gz:
 
 # Run Verilator to produce a fast binary to emulate this circuit.
 VERILATOR := $(INSTALLED_VERILATOR) --cc --exe
-VERILATOR_FLAGS := --top-module $(MODEL) +define+PRINTF_COND=\$$c\(\"verbose\"\) --assert \
+VERILATOR_FLAGS := --top-module $(MODEL) \
+  +define+PRINTF_COND=\$$c\(\"verbose\",\"\&\&\"\,\"done_reset\"\) \
+  +define+STOP_COND=\$$c\(\"done_reset\"\) --assert \
 	-Wno-STMTDLY --x-assign unique \
   -I$(base_dir)/vsrc \
   -O3 -CFLAGS "$(CXXFLAGS) -DVERILATOR -include $(base_dir)/csrc/verilator.h"

src/main/scala/rocket/dcache.scala

@@ -173,9 +173,9 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   when (s2_valid && (!s2_valid_hit || s2_update_meta)) { s1_nack := true }
 
   // exceptions
-  val misaligned = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes).misaligned
-  io.cpu.xcpt.ma.ld := s1_read && misaligned
-  io.cpu.xcpt.ma.st := s1_write && misaligned
+  val s1_storegen = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes)
+  io.cpu.xcpt.ma.ld := s1_read && s1_storegen.misaligned
+  io.cpu.xcpt.ma.st := s1_write && s1_storegen.misaligned
   io.cpu.xcpt.pf.ld := s1_read && tlb.io.resp.xcpt_ld
   io.cpu.xcpt.pf.st := s1_write && tlb.io.resp.xcpt_st
 
@@ -232,8 +232,8 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   // store->load RAW hazard detection
   val s1_idx = s1_req.addr(idxMSB, wordOffBits)
   val s1_raw_hazard = s1_read &&
-    ((pstore1_valid && pstore1_addr(idxMSB, wordOffBits) === s1_idx) ||
-     (pstore2_valid && pstore2_addr(idxMSB, wordOffBits) === s1_idx))
+    ((pstore1_valid && pstore1_addr(idxMSB, wordOffBits) === s1_idx && (pstore1_storegen.mask & s1_storegen.mask).orR) ||
+     (pstore2_valid && pstore2_addr(idxMSB, wordOffBits) === s1_idx && (pstore2_storegen_mask & s1_storegen.mask).orR))
   when (s1_valid && s1_raw_hazard) { s1_nack := true }
 
   metaWriteArb.io.in(0).valid := (s2_valid_hit && s2_update_meta) || (s2_victimize && !s2_victim_dirty)

src/main/scala/rocket/fpu.scala

@@ -12,8 +12,8 @@ import cde.{Parameters, Field}
 
 case class FPUConfig(
   divSqrt: Boolean = true,
-  sfmaLatency: Int = 2,
-  dfmaLatency: Int = 3
+  sfmaLatency: Int = 3,
+  dfmaLatency: Int = 4
 )
 
 object FPConstants