incorporate vortex2

src/main/scala/radiance/memory/VortexCache.scala

@@ -16,15 +16,18 @@ case class VortexL1Config(
     coreTagWidth: Int,
     writeInfoReqQSize: Int,
     mshrSize: Int,
-    l2ReqSourceGenSize: Int,
+    memSideSourceIds: Int,
-    uncachedAddrSets: Seq[AddressSet],
+    uncachedAddrSets: Seq[AddressSet]
-    icacheInstAddrSets: Seq[AddressSet]
 ) {
   def coreTagPlusSizeWidth: Int = {
     log2Ceil(wordSize) + coreTagWidth
   }
+  // NOTE: This assertion depends on the fact that the Vortex cache is
+  // configured to have 1 bank, and that it uses MSHR id as the tag of
+  // memory-side requests.  Otherwise, it will append bank id to the tag as
+  // well and break this requirement.
   require(
-    mshrSize == l2ReqSourceGenSize,
+    mshrSize == memSideSourceIds,
     "MSHR size must match the number of sourceIds to downstream."
   )
 }
@@ -37,18 +40,15 @@ object defaultVortexL1Config
       coreTagWidth = 8,
       writeInfoReqQSize = 16,
       mshrSize = 8,
-      l2ReqSourceGenSize = 8,
+      memSideSourceIds = 8,
-      uncachedAddrSets = Seq(AddressSet(0x2000000L, 0xffL)),
+      // Don't cache CLINT region to ensure coherent access
-      icacheInstAddrSets = Seq(AddressSet(0x80000000L, 0xfffffffL))
+      uncachedAddrSets = Seq(AddressSet(0x2000000L, 0xffffL))
     )
 
 class VortexL1Cache(config: VortexL1Config)(implicit p: Parameters)
     extends LazyModule {
-  // icache bank
+  val banks = Seq.tabulate(config.numBanks) { bankId =>
-  val icache_bank = LazyModule(new VortexBank(config, 0, isICache = true))
+    // helps with name mangling in Verilog
-
-  // dcache banks
-  val dcache_banks = Seq.tabulate(config.numBanks) { bankId =>
     val bank = LazyModule(new VortexBank(config, bankId))
     bank
   }
@@ -61,15 +61,13 @@ class VortexL1Cache(config: VortexL1Config)(implicit p: Parameters)
   // core-side crossbar that arbitrates core requests to banks
   protected val bankXbar = LazyModule(new TLXbar)
   bankXbar.node :=* coresideNode
-  dcache_banks.foreach { _.coalToVxCacheNode :=* bankXbar.node }
+  banks.foreach { _.coresideNode :=* bankXbar.node }
-  passThrough.coalToVxCacheNode :=* bankXbar.node
+  passThrough.coresideNode :=* bankXbar.node
-  icache_bank.coalToVxCacheNode :=* bankXbar.node
 
   // master node that exposes to and drives the downstream
   val masterNode = TLIdentityNode()
-  dcache_banks.foreach { masterNode := _.vxCacheToL2Node }
+  banks.foreach { masterNode := _.vxCacheToL2Node }
   masterNode := passThrough.vxCacheToL2Node
-  masterNode := icache_bank.vxCacheToL2Node
 
   lazy val module = new LazyModuleImp(this)
 }
@@ -101,7 +99,12 @@ class VortexBankPassThrough(config: VortexL1Config)(implicit p: Parameters)
       clients = Seq(
         TLMasterParameters.v1(
           name = "VortexBank",
-          sourceId = IdRange(0, 1 << (log2Ceil(config.l2ReqSourceGenSize) + 5)),
+          sourceId = IdRange(
+            0,
+            1 << (log2Ceil(
+              config.memSideSourceIds
+            ) + 5 /*FIXME: give more sourceId so that passthrough doesn't block; hacky*/ )
+          ),
           supportsProbe = TransferSizes(1, config.wordSize),
           supportsGet = TransferSizes(1, config.wordSize),
           supportsPutFull = TransferSizes(1, config.wordSize),
@@ -111,14 +114,14 @@ class VortexBankPassThrough(config: VortexL1Config)(implicit p: Parameters)
     )
   )
 
-  val coalToVxCacheNode = TLManagerNode(managerParam)
+  val coresideNode = TLManagerNode(managerParam)
   val vxCacheFetchNode = TLClientNode(clientParam)
   val vxCacheToL2Node = TLIdentityNode()
   vxCacheToL2Node := TLWidthWidget(config.cacheLineSize) := vxCacheFetchNode
 
-  // the implementation to make everything a pass through
+  // passthrough logic
   lazy val module = new LazyModuleImp(this) {
-    val (upstream, _) = coalToVxCacheNode.in(0)
+    val (upstream, _) = coresideNode.in(0)
     val (downstream, _) = vxCacheFetchNode.out(0)
 
     downstream.a <> upstream.a
@@ -129,28 +132,22 @@ class VortexBankPassThrough(config: VortexL1Config)(implicit p: Parameters)
 class VortexBank(
     config: VortexL1Config,
     bankId: Int,
-    isICache: Boolean = false
 )(implicit p: Parameters)
     extends LazyModule {
   // Generate AddressSet by excluding Addr we don't want
   def generateAddressSets(): Seq[AddressSet] = {
-    if (isICache) {
+    // suppose have 4 bank
-      config.icacheInstAddrSets
+    // base for bank 1: ...000000|01|0000
-      // Seq(AddressSet(0x00000000L, 0xFFFFFFFFL))
+    // mask for bank 1;    111111|00|1111
-    } else {
+    val base = 0x00000000L | (bankId * config.wordSize)
-      // suppose have 4 bank
+    val mask = 0xffffffffL ^ ((config.numBanks - 1) * config.wordSize)
-      // base for bank 1: ...000000|01|0000
-      // mask for bank 1;    111111|00|1111
-      val mask = 0xffffffffL ^ ((config.numBanks - 1) * config.wordSize)
-      val base = 0x00000000L | (bankId * config.wordSize)
 
-      val excludeSets = (config.uncachedAddrSets ++ config.icacheInstAddrSets)
+    val excludeSets = config.uncachedAddrSets
-      var remainingSets: Seq[AddressSet] = Seq(AddressSet(base, mask))
+    var remainingSets: Seq[AddressSet] = Seq(AddressSet(base, mask))
-      for (excludeSet <- excludeSets) {
+    for (excludeSet <- excludeSets) {
-        remainingSets = remainingSets.flatMap(_.subtract(excludeSet))
+      remainingSets = remainingSets.flatMap(_.subtract(excludeSet))
-      }
-      remainingSets
     }
+    remainingSets
   }
 
   // Slave node to upstream
@@ -177,7 +174,7 @@ class VortexBank(
       clients = Seq(
         TLMasterParameters.v1(
           name = "VortexBank",
-          sourceId = IdRange(0, config.l2ReqSourceGenSize),
+          sourceId = IdRange(0, config.memSideSourceIds),
           supportsProbe = TransferSizes(1, config.wordSize),
           supportsGet = TransferSizes(1, config.wordSize),
           supportsPutFull = TransferSizes(1, config.wordSize),
@@ -187,7 +184,8 @@ class VortexBank(
     )
   )
 
-  val coalToVxCacheNode = TLManagerNode(managerParam)
+  // Core -> VxCache
+  val coresideNode = TLManagerNode(managerParam)
   val vxCacheToL2Node = TLIdentityNode()
   val vxCacheFetchNode = TLClientNode(clientParam)
 
@@ -203,7 +201,7 @@ class VortexBankImp(
     config: VortexL1Config
 ) extends LazyModuleImp(outer) {
   val vxCache = Module(
-    new VX_cache(
+    new VX_cache_top(
       WORD_SIZE = config.wordSize,
       CACHE_LINE_SIZE = config.cacheLineSize,
       CORE_TAG_WIDTH = config.coreTagPlusSizeWidth,
@@ -250,7 +248,7 @@ class VortexBankImp(
 
   // Translate TL request from Coalescer to requests for VX_cache
   def TLReq2VXReq = {
-    val (tlInFromCoal, _) = outer.coalToVxCacheNode.in.head
+    val (tlInFromCoal, _) = outer.coresideNode.in.head
 
     // coal -> vxCache
     tlInFromCoal.a.ready :=
@@ -327,12 +325,12 @@ class VortexBankImp(
     tlInFromCoal.d.bits.data := vxCache.io.core_rsp_data
   }
 
-  // Since Vortex L1 is a write-through cache, it doesn't bookkeep writes and
+  // Since Vortex L1 is a write-through cache, it doesn't bookkeep writes in
-  // therefore doesn't allocate a new UUID for write requests.  We use a
+  // its MSHR and therefore doesn't allocate a new tag id for write requests.
-  // separate source ID allocator to solve this.
+  // We use a separate source ID allocator to solve this.
   val sourceGen = Module(
     new NewSourceGenerator(
-      log2Ceil(config.l2ReqSourceGenSize),
+      log2Ceil(config.memSideSourceIds),
       metadata = Some(UInt(32.W)),
       ignoreInUse = false
     )
@@ -389,70 +387,82 @@ class VortexBankImp(
   VXReq2TLReq
 }
 
-class VX_cache(
+class VX_cache_top(
-    CACHE_ID: Int = 0, // seems to be only used for debug trace prints
+    // these values should match the default settings in Verilog
+    // TODO: INSTANCE_ID
     CACHE_SIZE: Int = 16384 / 4, // <FIXME, divided by 4 for faster simulation
     CACHE_LINE_SIZE: Int = 16,
-    NUM_PORTS: Int = 1,
+    NUM_WAYS: Int = 4,
-    WORD_SIZE: Int =
+    // for single-bank configuration, set NUM_REQS = 1 and instead set
-      16, // hack - one "word" is enough to satisfy all 4 warps after decoalescing.
+    // WORD_SIZE to something wider than 4
-    CREQ_SIZE: Int = 0,
+    WORD_SIZE: Int = 16,
     CRSQ_SIZE: Int = 2,
-    MSHR_SIZE: Int = 8,
+    MSHR_SIZE: Int = 16,
     MRSQ_SIZE: Int = 0,
     MREQ_SIZE: Int = 4,
     WRITE_ENABLE: Int = 1,
+    UUID_WIDTH: Int = 0, // FIXME: should be different for debug
     CORE_TAG_WIDTH: Int =
-      10, // source ID ranges from 0 to 1 << 10, we need to allocate upper bits to save size
+      16, // source ID ranges from 0 to 1 << 10, we need to allocate upper bits to save size
-    CORE_TAG_ID_BITS: Int =
+    CORE_OUT_REG : Int = 0,
-      5, // no idea what this is, just match it with default L1 dcache
+    MEM_OUT_REG : Int = 0,
-    BANK_ADDR_OFFSET: Int = 0,
-    NC_ENABLE: Int = 0, // NC_ENABLE=1 means the cache becomes a passthrough
-    WORD_ADDR_WIDTH: Int = 28, // 16 byte "word" = 4 bits
-    MEM_TAG_WIDTH: Int =
-      14, // Elaborated value is also completely different from (32 - log2Ceil(CACHE_LINE_SIZE)). This should match with sourceIds on client node associated with this cache
-    MEM_ADDR_WIDTH: Int = 28 // 16 byte cache line = 4 bits
 ) extends BlackBox(
       Map(
-        "CACHE_ID" -> CACHE_ID,
+        // NOTE: NUM_REQS is analogous to SIMD width, whereas NUM_BANKS is the
-        "NUM_REQS" -> 1, // force to instantiate single bank by setting NUM_REQS to 1
+        // actual number of banks.  VX_cache.sv instantiates VX_stream_xbar
+        // that arbitrates the higher NUM_REQS into NUM_BANKS.  Since we do
+        // that logic ourselves using TL units, fix those params to 1 for the
+        // Verilog side.
+        "NUM_REQS" -> 1,
         "CACHE_SIZE" -> CACHE_SIZE,
-        "CACHE_LINE_SIZE" -> CACHE_LINE_SIZE,
+        "LINE_SIZE" -> CACHE_LINE_SIZE,
-        "NUM_PORTS" -> NUM_PORTS,
+        // NUM_BANKS is set to 1 to treat a whole VX_cache_top instance as a
+        // single bank
+        "NUM_BANKS" -> 1,
+        "NUM_WAYS" -> NUM_WAYS,
         "WORD_SIZE" -> WORD_SIZE,
-        "CREQ_SIZE" -> CREQ_SIZE,
         "CRSQ_SIZE" -> CRSQ_SIZE,
         "MSHR_SIZE" -> MSHR_SIZE,
         "MRSQ_SIZE" -> MRSQ_SIZE,
         "MREQ_SIZE" -> MREQ_SIZE,
         "WRITE_ENABLE" -> WRITE_ENABLE,
-        "CORE_TAG_WIDTH" -> CORE_TAG_WIDTH,
+        "UUID_WIDTH" -> UUID_WIDTH,
-        "CORE_TAG_ID_BITS" -> CORE_TAG_ID_BITS,
+        "TAG_WIDTH" -> CORE_TAG_WIDTH,
-        "MEM_TAG_WIDTH" -> MEM_TAG_WIDTH,
+        "CORE_OUT_REG" -> CORE_OUT_REG,
-        "BANK_ADDR_OFFSET" -> BANK_ADDR_OFFSET,
+        "MEM_OUT_REG" -> MEM_OUT_REG,
-        "NC_ENABLE" -> NC_ENABLE
+        // Although VX_cache_top exposes it as a parameter, MEM_TAG_WIDTH is
+        // not really configurable -- it is set to be a concatenation of the
+        // MSHR id and cache bank id.  Instead of trying to configure it from
+        // Chisel side, we try to figure out its value that's elaborated in the
+        // Verilog side and configure the Chisel io width correspondingly.
+        // "MEM_TAG_WIDTH" -> MEM_TAG_WIDTH
       )
     )
     with HasBlackBoxResource {
 
+  def memTagWidth(mshrSize: Int, numBanks: Int): Int =
+    log2Ceil(mshrSize) + log2Ceil(numBanks)
+  val MEM_TAG_WIDTH = memTagWidth(MSHR_SIZE, 1/* NUM_BANKS */)
+
+  // These logic is fixed in VX_cache_define.vh
+  val memAddrWidth = 32 // FIXME hardcoded
+  val cacheWordAddrWidth = 32 - log2Ceil(WORD_SIZE)
+  val cacheMemAddrWidth = 32 - log2Ceil(CACHE_LINE_SIZE)
+
   val io = IO(new Bundle {
     val clk = Input(Clock())
     val reset = Input(Reset())
 
-    // We should be able to turn the following into TileLink easily
-
     // CACHE <> CORE
     val core_req_valid = Input(Bool())
     val core_req_rw = Input(Bool())
-    val core_req_addr = Input(UInt(WORD_ADDR_WIDTH.W))
     val core_req_byteen = Input(UInt(WORD_SIZE.W))
+    val core_req_addr = Input(UInt(cacheWordAddrWidth.W))
     val core_req_data = Input(UInt((WORD_SIZE * 8).W))
     val core_req_tag = Input(UInt(CORE_TAG_WIDTH.W))
     val core_req_ready = Output(Bool())
 
     val core_rsp_valid = Output(Bool()) // 1 bit wide
-    val core_rsp_tmask =
-      Output(Bool()) // 1 bit wide, probably can ignore (check waveform)
     val core_rsp_data = Output(UInt((WORD_SIZE * 8).W))
     val core_rsp_tag = Output(UInt(CORE_TAG_WIDTH.W))
     val core_rsp_ready = Input(Bool())
@@ -461,7 +471,7 @@ class VX_cache(
     val mem_req_valid = Output(Bool())
     val mem_req_rw = Output(Bool())
     val mem_req_byteen = Output(UInt(CACHE_LINE_SIZE.W))
-    val mem_req_addr = Output(UInt(MEM_ADDR_WIDTH.W))
+    val mem_req_addr = Output(UInt(cacheMemAddrWidth.W))
     val mem_req_data = Output(UInt((CACHE_LINE_SIZE * 8).W))
     val mem_req_tag = Output(UInt(MEM_TAG_WIDTH.W))
     val mem_req_ready = Input(Bool())
@@ -472,133 +482,15 @@ class VX_cache(
     val mem_rsp_ready = Output(Bool())
   })
 
-  addResource("/vsrc/vortex/hw/rtl/VX_dispatch.sv")
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_bank.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_issue.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_bypass.sv")
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_data.sv")
   addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_warp_sched.sv")
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_init.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sat.sv")
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_mshr.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_stride.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_lerp.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_addr.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_mem.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_format.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sampler.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_wrap.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_scope.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_fpu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_scoreboard.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_writeback.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_muldiv.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_decode.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_ibuffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_icache_stage.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_trace_instr.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpu_types.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_config.vh")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_lzc.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_fifo_queue.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_scan.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_find_first.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_multiplier.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_bits_remove.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_pipe_register.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_priority_encoder.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_reset_relay.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_popcount.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_bits_insert.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_skid_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_fixed_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_shift_register.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_index_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_onehot_encoder.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_matrix_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_dp_ram.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_axi_adapter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_elastic_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_rr_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_sp_ram.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_demux.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_serial_div.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_fair_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_pending_size.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_csr_data.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_cache_arb.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_ipdom_stack.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpr_stage.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_execute.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_fetch.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_alu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_platform.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_commit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_pipeline.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_lsu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_csr_unit.sv")
-  addResource("/vsrc/vortex/hw/VX_config.h")
-  addResource("/vsrc/vortex/sim/common/rvfloats.h")
-  addResource("/vsrc/vortex/sim/common/rvfloats.cpp")
-  addResource("/csrc/softfloat/include/internals.h")
-  addResource("/csrc/softfloat/include/primitives.h")
-  addResource("/csrc/softfloat/include/primitiveTypes.h")
-  addResource("/csrc/softfloat/include/softfloat.h")
-  addResource("/csrc/softfloat/include/softfloat_types.h")
-  addResource("/csrc/softfloat/RISCV/specialize.h")
-  addResource("/vsrc/vortex/hw/dpi/float_dpi.cpp")
-  addResource("/vsrc/vortex/hw/dpi/float_dpi.vh")
-  addResource("/vsrc/vortex/hw/dpi/util_dpi.cpp")
-  addResource("/vsrc/vortex/hw/dpi/util_dpi.vh")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_dpi.sv")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_types.vh")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_icache_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dcache_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_join_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ifetch_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_cache_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_memsys_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpr_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_decode_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_writeback_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_pipeline_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpr_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_cmt_to_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_csr_to_alu_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ifetch_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_alu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_csr_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ibuffer_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_branch_ctl_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dcache_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_icache_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_lsu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_wstall_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_mem_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fpu_to_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_commit_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_warp_ctl_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fetch_to_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_tex_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_mem_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fpu_req_if.sv")
-  // addResource("/vsrc/vortex/hw/rtl/cache/VX_shared_mem.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_core_rsp_merge.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_tag_access.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_core_req_bank_sel.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_bank.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_data_access.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_flush_ctrl.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_nc_bypass.sv")
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_miss_resrv.sv")
   addResource("/vsrc/vortex/hw/rtl/cache/VX_cache.sv")
-
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_tags.sv")
+  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_top.sv")
 }
 
 // <FIXME> Delete the following NewSourceGenerator when merging with origin/graphics

src/main/scala/radiance/subsystem/Configs.scala

@@ -82,7 +82,11 @@ class WithVortexL1Banks(nBanks: Int = 4) extends Config ((site, _, up) => {
   }
 })
 
-class WithCoalescer(nNewSrcIds: Int = 8) extends Config((site, _, up) => {
+// When `enable` is false, we still elaborate Coalescer, but it acts as a
+// pass-through logic that always outputs un-coalesced requests.  This is
+// useful for when we want to keep the generated wire and net names the same
+// to e.g. compare waveforms.
+class WithCoalescer(nNewSrcIds: Int = 8, enable : Boolean = true) extends Config((site, _, up) => {
   case CoalescerKey => {
     val (nLanes, numOldSrcIds) = up(SIMTCoreKey, site) match {
       case Some(param) => (param.nLanes, param.nSrcIds)
@@ -104,6 +108,7 @@ class WithCoalescer(nNewSrcIds: Int = 8) extends Config((site, _, up) => {
       
     // Note: this config chooses a single-sized coalescing logic by default.
     Some(DefaultCoalescerConfig.copy(
+      enable       = enable,
       numLanes     = nLanes,
       numOldSrcIds = numOldSrcIds,
       numNewSrcIds = nNewSrcIds,
@@ -150,4 +155,4 @@ class WithNCustomSmallRocketCores(
       crossing
     )) ++ prev
   }
-})
+})

src/main/scala/radiance/tile/VortexCore.scala

@@ -10,26 +10,26 @@ import org.chipsalliance.cde.config.Parameters
 import freechips.rocketchip.tile._
 
 class VortexBundleA(
-  sourceWidth: Int,
+  tagWidth: Int,
   dataWidth: Int
 ) extends Bundle {
   assert(dataWidth % 8 == 0)
   val opcode = UInt(3.W) // FIXME: hardcoded
   val size = UInt(4.W) // FIXME: hardcoded
-  val source = UInt(sourceWidth.W) // FIXME: hardcoded
+  val source = UInt(tagWidth.W) // FIXME: hardcoded
   val address = UInt(32.W) // FIXME: hardcoded
   val mask = UInt((dataWidth / 8).W) // FIXME: hardcoded
   val data = UInt(dataWidth.W) // FIXME: hardcoded
 }
 
 class VortexBundleD(
-  sourceWidth: Int,
+  tagWidth: Int,
   dataWidth: Int
 ) extends Bundle {
   assert(dataWidth % 8 == 0)
   val opcode = UInt(3.W) // FIXME: hardcoded
   val size = UInt(4.W) // FIXME: hardcoded
-  val source = UInt(sourceWidth.W) // FIXME: hardcoded
+  val source = UInt(tagWidth.W) // FIXME: hardcoded
   val data = UInt(dataWidth.W) // FIXME: hardcoded
 }
 
@@ -42,20 +42,60 @@ class VortexBundle(tile: VortexTile)(implicit p: Parameters) extends CoreBundle
   
   // conditionally instantiate ports depending on whether we want to use VX_cache or not
   val imem = if (!tile.vortexParams.useVxCache) Some(Vec(1, new Bundle {
-    val a = Decoupled(new VortexBundleA(sourceWidth = 10, dataWidth = 32))
+    val a = Decoupled(new VortexBundleA(tagWidth = tile.imemTagWidth, dataWidth = 32))
-    val d = Flipped(Decoupled(new VortexBundleD(sourceWidth = 10, dataWidth = 32)))
+    val d = Flipped(Decoupled(new VortexBundleD(tagWidth = tile.imemTagWidth, dataWidth = 32)))
   })) else None
   val dmem = if (!tile.vortexParams.useVxCache) Some(Vec(tile.numLanes, new Bundle {
-    val a = Decoupled(new VortexBundleA(sourceWidth = 10, dataWidth = 32))
+    // val a = Decoupled(new VortexBundleA(tagWidth = tile.dmemTagWidth, dataWidth = 32))
-    val d = Flipped(Decoupled(new VortexBundleD(sourceWidth = 10, dataWidth = 32)))
+    // val d = Flipped(Decoupled(new VortexBundleD(tagWidth = dmemTagWidth, dataWidth = 32)))
+  })) else None
+  val smem = if (!tile.vortexParams.useVxCache) Some(Vec(tile.numLanes, new Bundle {
+    // val a = Decoupled(new VortexBundleA(tagWidth = tile.smemTagWidth, dataWidth = 32))
+    // val d = Flipped(Decoupled(new VortexBundleD(tagWidth = tile.smemTagWidth, dataWidth = 32)))
   })) else None
   val mem = if (tile.vortexParams.useVxCache) Some(new Bundle { 
-    val a = Decoupled(new VortexBundleA(sourceWidth = 15, dataWidth = 128))
+    val a = Decoupled(new VortexBundleA(tagWidth = 15, dataWidth = 128))
-    val d = Flipped(Decoupled(new VortexBundleD(sourceWidth = 15, dataWidth = 128)))
+    val d = Flipped(Decoupled(new VortexBundleD(tagWidth = 15, dataWidth = 128)))
     // val a = tile.memNode.out.head._1.a.cloneType
     // val d = Flipped(tile.memNode.out.head._1.d.cloneType)
   }) else None
 
+  // Chisel doesn't support 2-D array in BlackBox interface to Verilog, so
+  // everything needs to be 1-D flattened UInt with their widths configurable by numLanes.
+  //
+  // FIXME: hardcoded bitwidths
+  val dmem_a_ready = Input(UInt((tile.numLanes * 1).W))
+  val dmem_a_valid = Output(UInt((tile.numLanes * 1).W))
+  val dmem_a_bits_opcode = Output(UInt((tile.numLanes * 3).W))
+  val dmem_a_bits_size = Output(UInt((tile.numLanes * 4).W))
+  val dmem_a_bits_source = Output(UInt((tile.numLanes * tile.dmemTagWidth).W))
+  val dmem_a_bits_address = Output(UInt((tile.numLanes * 32).W))
+  val dmem_a_bits_mask = Output(UInt((tile.numLanes * 4).W))
+  val dmem_a_bits_data = Output(UInt((tile.numLanes * 32).W))
+
+  val dmem_d_valid = Input(UInt((tile.numLanes * 1).W))
+  val dmem_d_bits_opcode = Input(UInt((tile.numLanes * 3).W))
+  val dmem_d_bits_size = Input(UInt((tile.numLanes * 4).W))
+  val dmem_d_bits_source = Input(UInt((tile.numLanes * tile.dmemTagWidth).W))
+  val dmem_d_bits_data = Input(UInt((tile.numLanes * 32).W))
+  val dmem_d_ready = Output(UInt((tile.numLanes * 1).W))
+
+  val smem_a_ready = Input(UInt((tile.numLanes * 1).W))
+  val smem_a_valid = Output(UInt((tile.numLanes * 1).W))
+  val smem_a_bits_opcode = Output(UInt((tile.numLanes * 3).W))
+  val smem_a_bits_size = Output(UInt((tile.numLanes * 4).W))
+  val smem_a_bits_source = Output(UInt((tile.numLanes * tile.smemTagWidth).W))
+  val smem_a_bits_address = Output(UInt((tile.numLanes * 32).W))
+  val smem_a_bits_mask = Output(UInt((tile.numLanes * 4).W))
+  val smem_a_bits_data = Output(UInt((tile.numLanes * 32).W))
+
+  val smem_d_valid = Input(UInt((tile.numLanes * 1).W))
+  val smem_d_bits_opcode = Input(UInt((tile.numLanes * 3).W))
+  val smem_d_bits_size = Input(UInt((tile.numLanes * 4).W))
+  val smem_d_bits_source = Input(UInt((tile.numLanes * tile.smemTagWidth).W))
+  val smem_d_bits_data = Input(UInt((tile.numLanes * 32).W))
+  val smem_d_ready = Output(UInt((tile.numLanes * 1).W))
+
   // val fpu = Flipped(new FPUCoreIO())
   //val rocc = Flipped(new RoCCCoreIO(nTotalRoCCCSRs))
   //val trace = Output(new TraceBundle)
@@ -70,7 +110,12 @@ class Vortex(tile: VortexTile)(implicit p: Parameters)
       // Each Vortex core gets tied-off hartId of 0, 1, 2, 3, ...
       // The actual MHARTID read by the program is different by warp, not core;
       // see VX_csr_data that implements the read logic for CSR_MHARTID/GWID.
-      Map("CORE_ID" -> tile.tileParams.hartId)
+      Map(
+        "CORE_ID" -> tile.tileParams.hartId,
+        // TODO: can we get this as a parameter?
+        "BOOTROM_HANG100" -> 0x10100,
+        "NUM_THREADS" -> tile.numLanes
+      )
     )
     with HasBlackBoxResource {
   // addResource("/vsrc/vortex/hw/unit_tests/generic_queue/testbench.v")
@@ -90,87 +135,153 @@ class Vortex(tile: VortexTile)(implicit p: Parameters)
   // addResource("/vsrc/vortex/hw/syn/synopsys/models/memory/cln28hpc/rf2_32x128_wm1/rf2_32x128_wm1.v")
   // addResource("/vsrc/vortex/hw/syn/synopsys/models/memory/cln28hpc/rf2_32x128_wm1/vsim/rf2_32x128_wm1_tb.v")
   // addResource("/vsrc/vortex/hw/syn/modelsim/vortex_tb.v")
-  addResource("/vsrc/vortex/hw/rtl/VX_dispatch.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_issue.sv")
 
-  addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_define.vh")
+
-  
+  addResource("/vsrc/vortex/hw/rtl/VX_gpu_pkg.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_warp_sched.sv")
+
-  // addResource("/vsrc/vortex/hw/rtl/Vortex.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sat.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_stride.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_lerp.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_addr.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_mem.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_format.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sampler.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_wrap.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_scope.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_fpu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_scoreboard.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_writeback.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_muldiv.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_decode.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_ibuffer.sv")
   // addResource("/vsrc/vortex/hw/rtl/VX_cluster.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_icache_stage.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_trace_instr.vh")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpu_types.vh")
   addResource("/vsrc/vortex/hw/rtl/VX_config.vh")
-  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_mux.sv")
+  addResource("/vsrc/vortex/hw/VX_config.h")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_lzc.sv")
+  addResource("/vsrc/vortex/hw/rtl/VX_define.vh")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_fifo_queue.sv")
+  addResource("/vsrc/vortex/hw/rtl/VX_platform.vh")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_scan.sv")
+  addResource("/vsrc/vortex/hw/rtl/VX_scope.vh")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_find_first.sv")
+  // addResource("/vsrc/vortex/hw/rtl/VX_socket.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_multiplier.sv")
+  addResource("/vsrc/vortex/hw/rtl/VX_types.vh")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_bits_remove.sv")
+  // addResource("/vsrc/vortex/hw/rtl/Vortex.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_pipe_register.sv")
+
-  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_onehot_mux.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_alu_unit.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_priority_encoder.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_commit.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_reset_relay.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_core.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_popcount.sv")
+  // These are top modules used for unittests
+  // addResource("/vsrc/vortex/hw/rtl/core/VX_core_top.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_top.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_cluster_top.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_csr_data.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_csr_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_dcr_data.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_decode.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_dispatch.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_dispatch_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_execute.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_fetch.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_gather_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_ibuffer.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_int_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_ipdom_stack.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_issue.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_lsu_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_muldiv_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_operands.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_pending_instr.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_schedule.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_scoreboard.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_sfu_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_smem_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_split_join.sv")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_trace.vh")
+  addResource("/vsrc/vortex/hw/rtl/core/VX_wctl_unit.sv")
+
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_bank.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_bypass.sv")
+  // need to disable this if VX_cache_cluster_top is disabled, otherwise causes
+  // unconnected port error
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_cluster.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_data.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_define.vh")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_init.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_mshr.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_tags.sv")
+  // addResource("/vsrc/vortex/hw/rtl/cache/VX_cache_wrap.sv")
+
+  // gbar is only used in the socket/cluster hierarchy
+  // addResource("/vsrc/vortex/hw/rtl/mem/VX_gbar_arb.sv")
+  // addResource("/vsrc/vortex/hw/rtl/mem/VX_gbar_bus_if.sv")
+  // addResource("/vsrc/vortex/hw/rtl/mem/VX_gbar_unit.sv")
+  // mem_arb is used in VX_socket or VX_cache_cluster
+  // addResource("/vsrc/vortex/hw/rtl/mem/VX_mem_arb.sv")
+  addResource("/vsrc/vortex/hw/rtl/mem/VX_mem_bus_if.sv")
+  // addResource("/vsrc/vortex/hw/rtl/mem/VX_mem_perf_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/mem/VX_shared_mem.sv")
+  addResource("/vsrc/vortex/hw/rtl/mem/VX_smem_switch.sv")
+
+  // tex_unit missing in Vortex 2.0
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sat.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_stride.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_lerp.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_addr.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_mem.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_format.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_sampler.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_unit.sv")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_define.vh")
+  // addResource("/vsrc/vortex/hw/rtl/tex_unit/VX_tex_wrap.sv")
+
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_allocator.sv")
+  // addResource("/vsrc/vortex/hw/rtl/libs/VX_avs_adapter.sv")
+  // addResource("/vsrc/vortex/hw/rtl/libs/VX_axi_adapter.sv")
   addResource("/vsrc/vortex/hw/rtl/libs/VX_bits_insert.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_skid_buffer.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_bits_remove.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_fixed_arbiter.sv")
+  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_bypass_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_shift_register.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_cyclic_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_index_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_onehot_encoder.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_matrix_arbiter.sv")
   // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_divider.sv")
   addResource("/vsrc/vortex/hw/rtl/libs/VX_dp_ram.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_axi_adapter.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_elastic_adapter.sv")
   addResource("/vsrc/vortex/hw/rtl/libs/VX_elastic_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_rr_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_arbiter.sv")
-  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_bypass_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_sp_ram.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_demux.sv")
-  
-  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_index_queue.sv")
-  addResource("/vsrc/vortex/hw/rtl/libs/VX_serial_div.sv")
   addResource("/vsrc/vortex/hw/rtl/libs/VX_fair_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_define.vh")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_fifo_queue.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_csr_data.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_find_first.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_cache_arb.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_generic_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_ipdom_stack.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_index_buffer.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_gpr_stage.sv")
+  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_index_queue.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_execute.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_lzc.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_fetch.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_matrix_arbiter.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_alu_unit.sv")
+  // addResource("/vsrc/vortex/hw/rtl/libs/VX_mem_adapter.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_platform.vh")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_mem_rsp_sel.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_commit.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_mem_scheduler.sv")
-  
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_multiplier.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_pipeline.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_mux.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_lsu_unit.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_onehot_encoder.sv")
-  addResource("/vsrc/vortex/hw/rtl/VX_csr_unit.sv")
+  // unused addResource("/vsrc/vortex/hw/rtl/libs/VX_onehot_mux.sv")
-  // addResource("/vsrc/vortex/hw/rtl/Vortex_axi.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_pending_size.sv")
-  // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_div.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_pipe_register.sv")
-  addResource("/vsrc/vortex/hw/VX_config.h")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_popcount.sv")
-  addResource("/vsrc/vortex/sim/common/rvfloats.h")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_priority_arbiter.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_priority_encoder.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_reduce.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_reset_relay.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_rr_arbiter.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_scan.sv")
+  // These VX_scope_* seems to be used for FPGA debugging; if we leave them in,
+  // they cause elaboration errors
+  // addResource("/vsrc/vortex/hw/rtl/libs/VX_scope_switch.sv")
+  // addResource("/vsrc/vortex/hw/rtl/libs/VX_scope_tap.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_serial_div.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_serial_mul.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_shift_register.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_skid_buffer.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_sp_ram.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_arb.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_switch.sv")
+  addResource("/vsrc/vortex/hw/rtl/libs/VX_stream_xbar.sv")
+
+  addResource("/vsrc/vortex/hw/dpi/float_dpi.cpp")
+  addResource("/vsrc/vortex/hw/dpi/float_dpi.vh")
+  addResource("/vsrc/vortex/hw/dpi/util_dpi.cpp")
+  addResource("/vsrc/vortex/hw/dpi/util_dpi.vh")
+  // needed dpi cpp files
+  addResource("/vsrc/vortex/sim/common/bitmanip.h")
+  addResource("/vsrc/vortex/sim/common/mem.cpp")
+  addResource("/vsrc/vortex/sim/common/mem.h")
+  addResource("/vsrc/vortex/sim/common/mempool.h")
   addResource("/vsrc/vortex/sim/common/rvfloats.cpp")
+  addResource("/vsrc/vortex/sim/common/rvfloats.h")
+  addResource("/vsrc/vortex/sim/common/simobject.h")
+  addResource("/vsrc/vortex/sim/common/stringutil.h")
+  addResource("/vsrc/vortex/sim/common/util.cpp")
+  addResource("/vsrc/vortex/sim/common/util.h")
+  addResource("/vsrc/vortex/sim/common/uuid_gen.h")
+
   // addResource("/csrc/softfloat_archive.a")
   addResource("/csrc/softfloat/include/internals.h")
   addResource("/csrc/softfloat/include/primitives.h")
@@ -178,11 +289,22 @@ class Vortex(tile: VortexTile)(implicit p: Parameters)
   addResource("/csrc/softfloat/include/softfloat.h")
   addResource("/csrc/softfloat/include/softfloat_types.h")
   addResource("/csrc/softfloat/RISCV/specialize.h")
-  addResource("/vsrc/vortex/hw/dpi/float_dpi.cpp")
+
-  addResource("/vsrc/vortex/hw/dpi/float_dpi.vh")
+  // Vortex 2.0: fp_cores/ renamed to fpu/
-  addResource("/vsrc/vortex/hw/dpi/util_dpi.cpp")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_class.sv")
-  addResource("/vsrc/vortex/hw/dpi/util_dpi.vh")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_cvt.sv")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_dpi.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_define.vh")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_div.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_dpi.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_dsp.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_fma.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_fpnew.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_ncomp.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_pkg.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_rounding.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_sqrt.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_to_csr_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/fpu/VX_fpu_unit.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_rounding.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/altera/stratix10/dspba_delay_ver.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/altera/stratix10/acl_fsqrt.sv")
@@ -195,46 +317,32 @@ class Vortex(tile: VortexTile)(implicit p: Parameters)
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_class.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_fpnew.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_cvt.sv")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_define.vh")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_fma.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_ncomp.sv")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_fpga.sv")
-  addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_types.vh")
+  // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fpu_types.vh")
   // addResource("/vsrc/vortex/hw/rtl/fp_cores/VX_fp_sqrt.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_icache_rsp_if.sv")
+
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dcache_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_join_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ifetch_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_cache_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_memsys_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpr_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_decode_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_writeback_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_pipeline_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_gpr_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_cmt_to_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_csr_to_alu_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ifetch_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_alu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_csr_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ibuffer_if.sv")
   addResource("/vsrc/vortex/hw/rtl/interfaces/VX_branch_ctl_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dcache_rsp_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_commit_csr_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_icache_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_lsu_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_wstall_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_mem_rsp_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fpu_to_csr_if.sv")
   addResource("/vsrc/vortex/hw/rtl/interfaces/VX_commit_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_req_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_commit_sched_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dcr_bus_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_decode_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_decode_sched_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_dispatch_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_execute_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fetch_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_ibuffer_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_operands_if.sv")
+  // addResource("/vsrc/vortex/hw/rtl/interfaces/VX_pipeline_perf_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_sched_csr_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_schedule_if.sv")
+  // addResource("/vsrc/vortex/hw/rtl/interfaces/VX_sfu_csr_if.sv")
+  // addResource("/vsrc/vortex/hw/rtl/interfaces/VX_sfu_perf_if.sv")
   addResource("/vsrc/vortex/hw/rtl/interfaces/VX_warp_ctl_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_tex_rsp_if.sv")
+  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_writeback_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fetch_to_csr_if.sv")
+
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_perf_tex_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_mem_req_if.sv")
-  addResource("/vsrc/vortex/hw/rtl/interfaces/VX_fpu_req_if.sv")
   // addResource("/vsrc/vortex/hw/rtl/afu/vortex_afu.sv")
   // addResource("/vsrc/vortex/hw/rtl/afu/ccip_std_afu.sv")
   // addResource("/vsrc/vortex/hw/rtl/afu/vortex_afu.vh")
@@ -262,10 +370,9 @@ class Vortex(tile: VortexTile)(implicit p: Parameters)
     addResource("/vsrc/vortex/hw/rtl/VX_core.sv")
     addResource("/vsrc/vortex/hw/rtl/VX_core_wrapper.sv")
   } else {
-    addResource("/vsrc/vortex/hw/rtl/VX_pipeline_wrapper.sv")
+    addResource("/vsrc/vortex/hw/rtl/VX_core_wrapper.sv")
   }
 
   val nTotalRoCCCSRs = 0
-  val coreBundle = new VortexBundle(tile)
+  val io = IO(new VortexBundle(tile))
-  val io = IO(coreBundle)
 }

src/main/scala/radiance/tile/VortexTile.scala

@@ -187,6 +187,18 @@ class VortexTile private (
       "We recommend setting nSrcIds to at least 16."
   )
 
+  val smemSourceWidth = 4 // FIXME: hardcoded
+
+  // TODO: parametrize
+  val numWarps = 4
+  val NW_WIDTH = (if (numWarps == 1) 1 else log2Ceil(numWarps))
+  val UUID_WIDTH = 44
+  val imemTagWidth = UUID_WIDTH + NW_WIDTH
+  val LSUQ_TAG_BITS = 4
+  val dmemTagWidth = UUID_WIDTH + LSUQ_TAG_BITS
+  // dmem and smem shares the same tag width, DCACHE_NOSM_TAG_WIDTH
+  val smemTagWidth = dmemTagWidth
+
   val imemNodes = Seq.tabulate(1) { i =>
     TLClientNode(
       Seq(
@@ -228,14 +240,37 @@ class VortexTile private (
       )
     )
   }
+
+  val smemNodes = Seq.tabulate(numLanes) { i =>
+    TLClientNode(
+      Seq(
+        TLMasterPortParameters.v1(
+          clients = Seq(
+            TLMasterParameters.v1(
+              sourceId = IdRange(0, 1 << smemSourceWidth),
+              name = s"Vortex Core ${vortexParams.hartId} SharedMem Lane $i",
+              requestFifo = true,
+              supportsProbe =
+                TransferSizes(1, lazyCoreParamsView.coreDataBytes),
+              supportsGet = TransferSizes(1, lazyCoreParamsView.coreDataBytes),
+              supportsPutFull =
+                TransferSizes(1, lazyCoreParamsView.coreDataBytes),
+              supportsPutPartial =
+                TransferSizes(1, lazyCoreParamsView.coreDataBytes)
+            )
+          )
+        )
+      )
+    )
+  }
+
   // combine outgoing per-lane dmemNode into 1 idenity node
   //
   // NOTE: We need TLWidthWidget here because there might be a data width
   // mismatch between Vortex's per-lane response and the system bus when we
   // don't instantiate either L1 or the coalescer.  This _should_ be optimized
-  // out when we instantiate coalescer which should handle data width conversion
+  // out when we instantiate either which should handle data width conversion
-  // internally (which it does by... using TLWidthWidget), but probably not
+  // internally (which it does by... using TLWidthWidget).
-  // the cleanest way to do this.
   val dmemAggregateNode = TLIdentityNode()
   dmemNodes.foreach { dmemAggregateNode := TLWidthWidget(4) := _ }
 
@@ -262,7 +297,7 @@ class VortexTile private (
   val coalescerNode = p(CoalescerKey) match {
     case Some(coalescerParam) => {
       val coal = LazyModule(
-        new CoalescingUnit(coalescerParam.copy(enable = true))
+        new CoalescingUnit(coalescerParam)
       )
       coal.cpuNode :=* dmemAggregateNode
       coal.aggregateNode // N+1 lanes
@@ -271,37 +306,56 @@ class VortexTile private (
   }
 
   // Conditionally instantiate L1 cache
-  val l1Node = p(VortexL1Key) match {
+  val (icacheNode, dcacheNode): (TLNode, TLNode) = p(VortexL1Key) match {
     case Some(vortexL1Config) => {
       println(
         s"============ Using Vortex L1 cache ================="
       )
-      require(
+      // require(
-        p(CoalescerKey).isDefined,
+      //   p(CoalescerKey).isDefined,
-        "Vortex L1 configuration currently only works when coalescer is also enabled."
+      //   "Vortex L1 configuration currently only works when coalescer is also enabled."
-      )
+      // )
 
-      val l1cache = LazyModule(new VortexL1Cache(vortexL1Config))
+      val icache = LazyModule(new VortexL1Cache(vortexL1Config))
-      // Connect L1 with imem_fetch_interface without XBar
+      val dcache = LazyModule(new VortexL1Cache(vortexL1Config))
-      // coalToVxCacheNode is a bad naming, it really means up steam of vxBank in whihc it takes input
+      // imemNodes.foreach { icache.coresideNode := TLWidthWidget(4) := _ }
-      // imemNodes.foreach { l1cache.icache_bank.coalToVxCacheNode := TLWidthWidget(4) := _ }
+      assert(imemNodes.length == 1) // FIXME
-      imemNodes.foreach { l1cache.coresideNode := TLWidthWidget(4) := _ }
+      icache.coresideNode := TLWidthWidget(4) := imemNodes(0)
-      l1cache.coresideNode :=* coalescerNode
+      // dmemNodes go through coalescerNode
-      l1cache.masterNode
+      dcache.coresideNode :=* coalescerNode
+      (icache.masterNode, dcache.masterNode)
     }
     case None => {
-      // Regardless of using coalescer or not, if we're not using L1, imemNode
+      val imemWideNode = TLIdentityNode()
-      // goes directly to tile exit xbar
+      assert(imemNodes.length == 1) // FIXME
-      // FIXME: unnatural, have L1 just handle dmem
+      imemWideNode := TLWidthWidget(4) := imemNodes(0)
-      imemNodes.foreach { tlMasterXbar.node := TLWidthWidget(4) := _ }
+      (imemWideNode, coalescerNode)
-      coalescerNode
     }
   }
 
+  // Instantiate sharedmem banks
+  //
+  // Instantiate the same number of banks as there are lanes.
+  // TODO: parametrize
+  val smemBanks = Seq.tabulate(numLanes) { bankId =>
+    // Banked-by-word (4 bytes)
+    // base for bank 1: ff...000000|01|00
+    // mask for bank 1; 00...111111|00|11
+    val base = 0xff000000L | (bankId * 4 /*wordSize*/ )
+    val mask = 0x00ffffffL ^ ((numLanes - 1) * 4 /*wordSize*/ )
+    LazyModule(new TLRAM(AddressSet(base, mask), beatBytes = 4 /*wordSize*/ ))
+  }
+  // smem lanes-to-banks crossbar
+  val smemXbar = LazyModule(new TLXbar)
+  smemNodes.foreach(smemXbar.node := _)
+  smemBanks.foreach(_.node := smemXbar.node)
+
   if (vortexParams.useVxCache) {
     tlMasterXbar.node := TLWidthWidget(16) := memNode
   } else {
-    tlMasterXbar.node :=* l1Node
+    // imemNodes.foreach { tlMasterXbar.node := TLWidthWidget(4) := _ }
+    tlMasterXbar.node :=* icacheNode
+    tlMasterXbar.node :=* dcacheNode
   }
 
   /* below are copied from rocket */
@@ -457,95 +511,170 @@ class VortexTileModuleImp(outer: VortexTile) extends BaseTileModuleImp(outer) {
     outer.memNode.out(0)._1.a <> memTLAdapter.io.outReq
     memTLAdapter.io.outResp <> outer.memNode.out(0)._1.d
   } else {
-    val imemTLAdapter = Module(
+    def connectImem = {
-      new VortexTLAdapter(
+      val imemTLAdapter = Module(
-        outer.imemSourceWidth,
-        chiselTypeOf(core.io.imem.get(0).a.bits),
-        chiselTypeOf(core.io.imem.get(0).d.bits),
-        outer.imemNodes.head.out.head
-      )
-    )
-    // TODO: make imemNodes not a vector
-    imemTLAdapter.io.inReq <> core.io.imem.get(0).a
-    core.io.imem.get(0).d <> imemTLAdapter.io.inResp
-    outer.imemNodes(0).out(0)._1.a <> imemTLAdapter.io.outReq
-    imemTLAdapter.io.outResp <> outer.imemNodes(0).out(0)._1.d
-
-    // @perf: this would duplicate SourceGenerator table for every lane and eat
-    // up some area
-    val dmemTLBundles = outer.dmemNodes.map(_.out.head._1)
-    val dmemTLAdapters = Seq.tabulate(outer.numLanes) { _ =>
-      Module(
         new VortexTLAdapter(
-          outer.dmemSourceWidth,
+          outer.imemSourceWidth,
-          chiselTypeOf(core.io.dmem.get(0).a.bits),
+          chiselTypeOf(core.io.imem.get(0).a.bits),
-          chiselTypeOf(core.io.dmem.get(0).d.bits),
+          chiselTypeOf(core.io.imem.get(0).d.bits),
-          outer.dmemNodes(0).out.head
+          outer.imemNodes.head.out.head
         )
       )
+      // TODO: make imemNodes not a vector
+      imemTLAdapter.io.inReq <> core.io.imem.get(0).a
+      core.io.imem.get(0).d <> imemTLAdapter.io.inResp
+      outer.imemNodes(0).out(0)._1.a <> imemTLAdapter.io.outReq
+      imemTLAdapter.io.outResp <> outer.imemNodes(0).out(0)._1.d
     }
 
-    // Since the individual per-lane TL requests might come back out-of-sync between
+    def connectDmem = {
-    // the lanes, but Vortex core expects the per-lane responses to be synced,
+      // @perf: this would duplicate SourceGenerator table for every lane and eat
-    // we need to selectively fire responses that have the same source, and
+      // up some area
-    // delay others.
+      val dmemTLBundles = outer.dmemNodes.map(_.out.head._1)
-    //
+      val dmemTLAdapters = Seq.tabulate(outer.numLanes) { _ =>
-    // In order to do that, we pick a source from one of the valid lanes using e.g.
+        Module(
-    // an arbiter.  Then using the chosen source id, we
+          new VortexTLAdapter(
-    // - lie to core that response is not valid if source doesn't match picked, and
+            outer.dmemSourceWidth,
-    // - lie to downstream that core is not ready if source doesn't match picked.
+            new VortexBundleA(tagWidth = outer.dmemTagWidth, dataWidth = 32),
-    //
+            new VortexBundleD(tagWidth = outer.dmemTagWidth, dataWidth = 32),
-    // Note that we cannot do this filtering logic using TileLink source ID, because
+            outer.dmemNodes(0).out.head
-    // we're allocating source for each lane independently.  In that case, it's
+          )
-    // possible that lane 0's source matches lane 1/2/3's source by chance,
+        )
-    // even when they originated from different warps.  Using Vortex's dcache req tag
+      }
-    // solves this issue because they use a UUID that is unique across all requests
+
-    // in the program.
+      // Since the individual per-lane TL requests might come back out-of-sync between
-    //
+      // the lanes, but Vortex core expects the per-lane responses to be synced,
-    // TODO: A cleaner solution would be to simply do a synchronized allocation
+      // we need to selectively fire responses that have the same source, and
-    // of a same source id for all lanes.
+      // delay others.
-    val arb = Module(
+      //
-      new RRArbiter(
+      // In order to do that, we pick a source from one of the valid lanes using e.g.
-        core.io.dmem.get.head.d.bits.source.cloneType,
+      // an arbiter.  Then using the chosen source id, we
-        outer.numLanes
+      // - lie to core that response is not valid if source doesn't match picked, and
+      // - lie to downstream that core is not ready if source doesn't match picked.
+      //
+      // Note that we cannot do this filtering logic using TileLink source ID, because
+      // we're allocating source for each lane independently.  In that case, it's
+      // possible that lane 0's source matches lane 1/2/3's source by chance,
+      // even when they originated from different warps.  Using Vortex's dcache req tag
+      // solves this issue because they use a UUID that is unique across all requests
+      // in the program.
+      //
+      // TODO: A cleaner solution would be to simply do a synchronized allocation
+      // of a same source id for all lanes.
+      val arb = Module(
+        new RRArbiter(
+          // FIXME: should really be source on D channel
+          new VortexBundleA(tagWidth = outer.dmemTagWidth, dataWidth = 32).source.cloneType,
+          outer.numLanes
+        )
       )
-    )
+      arb.io.out.ready := true.B
-    arb.io.out.ready := true.B
+      val dmemBundles = dmemTLAdapters.map(_.io.inResp)
-    val dmemBundles = dmemTLAdapters.map(_.io.inResp)
+      (arb.io.in zip dmemBundles).foreach { case (arbIn, vxDmem) =>
-    (arb.io.in zip dmemBundles).foreach { case (arbIn, vxDmem) =>
+        arbIn.valid := vxDmem.valid
-      arbIn.valid := vxDmem.valid
+        arbIn.bits := vxDmem.bits.source
-      arbIn.bits := vxDmem.bits.source
+      }
-    }
+      val matchingSources = Wire(UInt(outer.numLanes.W))
-    val matchingSources = Wire(UInt(outer.numLanes.W))
+      matchingSources := dmemBundles
-    matchingSources := dmemBundles
+        .map(b =>
-      .map(b =>
+          // If there is no valid response pending across all lanes,
-        // If there is no valid response pending across all lanes,
+          // matchingSources should not filter out upstream ready signals, so
-        // matchingSources should not filter out upstream ready signals, so
+          // set it to all-1
-        // set it to all-1
+          !arb.io.out.valid || (b.bits.source === arb.io.out.bits)
-        !arb.io.out.valid || (b.bits.source === arb.io.out.bits)
+        )
-      )
+        .asUInt
-      .asUInt
 
-    // make connection:
+      // make connection:
-    // VortexBundle <--> sourceId filter <--> VortexTLAdapter <--> dmemNodes
+      // VortexBundle <--> sourceId filter <--> VortexTLAdapter <--> dmemNodes
-    (core.io.dmem.get zip dmemTLAdapters) foreach { case (coreMem, tlAdapter) =>
+      // 
-      tlAdapter.io.inReq <> coreMem.a
+      // Chisel doesn't support 2-D array in BlackBox interface to Verilog, so
-      coreMem.d <> tlAdapter.io.inResp
+      // need to flatten everything.
-    }
+      dmemTLAdapters.zipWithIndex.foreach {
-    (core.io.dmem.get zip dmemTLAdapters).zipWithIndex.foreach {
+        case (tlAdapter, i) =>
-      case ((coreMem, tlAdapter), i) =>
+          // tlAdapter.io.inReq <> coreMem.a
-        coreMem.d.valid := tlAdapter.io.inResp.valid && matchingSources(i)
+          tlAdapter.io.inReq.valid := core.io.dmem_a_valid(i)
-        tlAdapter.io.inResp.ready := coreMem.d.ready && matchingSources(i)
+          tlAdapter.io.inReq.bits.opcode := core.io.dmem_a_bits_opcode(3 * (i + 1) - 1, 3 * i)
-    }
+          tlAdapter.io.inReq.bits.size := core.io.dmem_a_bits_size(4 * (i + 1) - 1, 4 * i)
-    (dmemTLAdapters zip dmemTLBundles) foreach { case (tlAdapter, tlOut) =>
+          tlAdapter.io.inReq.bits.source := core.io.dmem_a_bits_source(outer.dmemTagWidth * (i + 1) - 1, outer.dmemTagWidth * i)
-      tlOut.a <> tlAdapter.io.outReq
+          tlAdapter.io.inReq.bits.address := core.io.dmem_a_bits_address(32 * (i + 1) - 1, 32 * i)
-      tlAdapter.io.outResp <> tlOut.d
+          tlAdapter.io.inReq.bits.mask := core.io.dmem_a_bits_mask(4 * (i + 1) - 1, 4 * i)
+          tlAdapter.io.inReq.bits.data := core.io.dmem_a_bits_data(32 * (i + 1) - 1, 32 * i)
+      }
+      core.io.dmem_a_ready := dmemTLAdapters.map(_.io.inReq.ready).asUInt
+
+      core.io.dmem_d_valid := dmemTLAdapters.map(_.io.inResp.valid).asUInt
+      core.io.dmem_d_bits_opcode := dmemTLAdapters.map(_.io.inResp.bits.opcode).asUInt
+      core.io.dmem_d_bits_size := dmemTLAdapters.map(_.io.inResp.bits.size).asUInt
+      core.io.dmem_d_bits_source := dmemTLAdapters.map(_.io.inResp.bits.source).asUInt
+      core.io.dmem_d_bits_data := dmemTLAdapters.map(_.io.inResp.bits.data).asUInt
+
+      // override response channel with matchingSources
+      val dmem_d_valid_vec = Wire(Vec(outer.numLanes, Bool()))
+      dmemTLAdapters.zipWithIndex.foreach {
+        case (tlAdapter, i) =>
+          dmem_d_valid_vec(i) := tlAdapter.io.inResp.valid && matchingSources(i)
+          tlAdapter.io.inResp.ready := core.io.dmem_d_ready(i) && matchingSources(i)
+      }
+      core.io.dmem_d_valid := dmem_d_valid_vec.asUInt
+
+      (dmemTLAdapters zip dmemTLBundles) foreach { case (tlAdapter, tlOut) =>
+        tlOut.a <> tlAdapter.io.outReq
+        tlAdapter.io.outResp <> tlOut.d
+      }
+
+      outer.dmemAggregateNode.out.foreach { bo =>
+        dontTouch(bo._1.a)
+        dontTouch(bo._1.d)
+      }
     }
 
-    outer.dmemAggregateNode.out.foreach { bo =>
+    def connectSmem = {
-      dontTouch(bo._1.a)
+      // @perf: this would duplicate SourceGenerator table for every lane and eat
-      dontTouch(bo._1.d)
+      // up some area
+      val smemTLBundles = outer.smemNodes.map(_.out.head._1)
+      val smemTLAdapters = Seq.tabulate(outer.numLanes) { _ =>
+        Module(
+          new VortexTLAdapter(
+            outer.smemSourceWidth,
+            new VortexBundleA(tagWidth = outer.smemTagWidth, dataWidth = 32),
+            new VortexBundleD(tagWidth = outer.smemTagWidth, dataWidth = 32),
+            outer.smemNodes(0).out.head
+          )
+        )
+      }
+
+      smemTLAdapters.zipWithIndex.foreach {
+        case (tlAdapter, i) =>
+          // tlAdapter.io.inReq <> coreMem.a
+          tlAdapter.io.inReq.valid := core.io.smem_a_valid(i)
+          tlAdapter.io.inReq.bits.opcode := core.io.smem_a_bits_opcode(3 * (i + 1) - 1, 3 * i)
+          tlAdapter.io.inReq.bits.size := core.io.smem_a_bits_size(4 * (i + 1) - 1, 4 * i)
+          tlAdapter.io.inReq.bits.source := core.io.smem_a_bits_source(outer.smemTagWidth * (i + 1) - 1, outer.smemTagWidth * i)
+          tlAdapter.io.inReq.bits.address := core.io.smem_a_bits_address(32 * (i + 1) - 1, 32 * i)
+          tlAdapter.io.inReq.bits.mask := core.io.smem_a_bits_mask(4 * (i + 1) - 1, 4 * i)
+          tlAdapter.io.inReq.bits.data := core.io.smem_a_bits_data(32 * (i + 1) - 1, 32 * i)
+      }
+      core.io.smem_a_ready := smemTLAdapters.map(_.io.inReq.ready).asUInt
+
+      core.io.smem_d_valid := smemTLAdapters.map(_.io.inResp.valid).asUInt
+      core.io.smem_d_bits_opcode := smemTLAdapters.map(_.io.inResp.bits.opcode).asUInt
+      core.io.smem_d_bits_size := smemTLAdapters.map(_.io.inResp.bits.size).asUInt
+      core.io.smem_d_bits_source := smemTLAdapters.map(_.io.inResp.bits.source).asUInt
+      core.io.smem_d_bits_data := smemTLAdapters.map(_.io.inResp.bits.data).asUInt
+      smemTLAdapters.zipWithIndex.foreach {
+        case (tlAdapter, i) =>
+          tlAdapter.io.inResp.ready := core.io.smem_d_ready(i)
+      }
+
+      (smemTLAdapters zip smemTLBundles) foreach { case (tlAdapter, tlOut) =>
+        tlOut.a <> tlAdapter.io.outReq
+        tlAdapter.io.outResp <> tlOut.d
+      }
     }
+
+    connectImem
+    connectDmem
+    connectSmem
   }
 
   // TODO: generalize for useVxCache
@@ -566,7 +695,7 @@ class VortexTLAdapter(
     val inResp = Decoupled(inRespT)
     val outResp = chiselTypeOf(outTL._1.d)
   })
-  val edge = outTL._2
+  val (bundle, edge) = outTL
   val sourceGen = Module(
     new SourceGenerator(
       newSourceWidth,
@@ -587,14 +716,20 @@ class VortexTLAdapter(
   io.outReq.bits.size := io.inReq.bits.size
   io.outReq.bits.source := io.inReq.bits.source
   io.outReq.bits.address := io.inReq.bits.address
-  // generate TL-correct mask
+  // Get requires contiguous mask; only copy core's potentially-partial mask
-  io.outReq.bits.mask := edge.mask(io.inReq.bits.address, io.inReq.bits.size)
+  // when writing
+  io.outReq.bits.mask := Mux(
+    edge.hasData(io.outReq.bits),
+    io.inReq.bits.mask,
+    // generate TL-correct mask
+    edge.mask(io.inReq.bits.address, io.inReq.bits.size)
+  )
   io.outReq.bits.data := io.inReq.bits.data
   io.outReq.bits.corrupt := 0.U
   io.inReq.ready := io.outReq.ready
   // VortexBundleD <> TLBundleD
-  // Do not reply to write requests; Vortex core does not expect ack on writes
+  // Filtering out write requests is handled inside the wrapper Verilog
-  io.inResp.valid := io.outResp.valid && edge.hasData(io.outResp.bits)
+  io.inResp.valid := io.outResp.valid
   io.inResp.bits.opcode := io.outResp.bits.opcode
   io.inResp.bits.size := io.outResp.bits.size
   io.inResp.bits.source := io.outResp.bits.source