Cleanup for l15 threadid

PrincetonUniversity · Aug 30, 2023 · 7aa28af · 7aa28af
1 parent 6aaec1b
commit 7aa28af
Show file tree

Hide file tree

Showing 10 changed files with 84 additions and 141 deletions.
diff --git a/piton/ariane_setup.sh b/piton/ariane_setup.sh
@@ -1,5 +1,4 @@
 #!/bin/bash
-# Modified by Barcelona Supercomputing Center on March 3rd, 2022
 # Copyright 2018 ETH Zurich and University of Bologna.
 # Copyright and related rights are licensed under the Solderpad Hardware
 # License, Version 0.51 (the "License"); you may not use this file except in
@@ -52,11 +51,8 @@ echo "make sure that you source this script in a bash shell in the root folder o
 
 if [ -z "$BASH" ] || [ ${0: -4} !=  "bash" ]
 then
-  #echo "not in bash ($0), aborting"
-  # Commentted out, as it is not really an issue and it may fail when running the CICD
-  echo "not in bash ($0)"
-  #return
-
+  echo "not in bash ($0), aborting"
+  return
 fi
 
 SCRIPTNAME=ariane_setup.sh
@@ -80,15 +76,18 @@ export ARIANE_ROOT=${PITON_ROOT}/piton/design/chip/tile/ariane/
 export CXX=g++ CC=gcc
 # customize this to a fast local disk
 
-export RISCV=~/scratch/`whoami`/riscv_install
-export VERILATOR_ROOT=~/scratch/`whoami`/verilator_4_104/
+if [ "$RISCV" ==  "" ]
+then
+  export RISCV=$HOME/scratch/riscv_install
+fi
+export VERILATOR_ROOT=$ARIANE_ROOT/tmp/verilator-4.014/
+
 # setup paths
 export PATH=$RISCV/bin:$VERILATOR_ROOT/bin:$PATH
 export LIBRARY_PATH=$RISCV/lib
-export LD_LIBRARY_PATH=$RISCV/lib:$LD_LIBRARY_PATH
-export C_INCLUDE_PATH=$RISCV/include:$VERILATOR_ROOT/include:$C_INCLUDE_PATH
-export CPLUS_INCLUDE_PATH=$RISCV/include:$VERILATOR_ROOT/include:$CPLUS_INCLUDE_PATH
-export MODELSIM_HOME=$HOME/scratch/questa_install/questasim
+export LD_LIBRARY_PATH=$RISCV/lib
+export C_INCLUDE_PATH=$RISCV/include:$VERILATOR_ROOT/include
+export CPLUS_INCLUDE_PATH=$RISCV/include:$VERILATOR_ROOT/include
 
 # source OpenPiton setup script
 # note: customize this script to reflect your tool setup

diff --git a/piton/design/chip/tile/l15/rtl/l15_mshr.v.pyv b/piton/design/chip/tile/l15/rtl/l15_mshr.v.pyv
@@ -163,24 +163,24 @@ integer i=0;
 always @ *
 begin
 
-	 for(i = 0; i < `L15_NUM_THREADS; i = i+1)
-	 begin
+    for(i = 0; i < `L15_NUM_THREADS; i = i+1)
+    begin
         tmp_vals[i] = 0;
         tmp_vals[i][`L15_MSHR_ID_IFILL] = ifill_val[i];
-	     tmp_vals[i][`L15_MSHR_ID_LD]    = ld_val[i];
-		  tmp_vals[i][`L15_MSHR_ID_ST]    = st_val[i];
+        tmp_vals[i][`L15_MSHR_ID_LD]    = ld_val[i];
+        tmp_vals[i][`L15_MSHR_ID_ST]    = st_val[i];
 
         tmp_st_address[i] = st_address[i];
         tmp_ld_address[i] = ld_address[i];
-		  tmp_st_way[i]     = st_way[i];
-		  tmp_st_state[i]   = st_state[i];
+        tmp_st_way[i]     = st_way[i];
+        tmp_st_state[i]   = st_state[i];
 
-		  mshr_pipe_vals_s1[(`L15_NUM_MSHRID_PER_THREAD*(i+1))-1 -: `L15_NUM_MSHRID_PER_THREAD] = tmp_vals[i];
-		  mshr_pipe_ld_address[(`L15_PADDR_WIDTH*(i+1))-1 -: `L15_PADDR_WIDTH]  = tmp_ld_address[i];
-		  mshr_pipe_st_address[(`L15_PADDR_WIDTH*(i+1))-1 -: `L15_PADDR_WIDTH]  = tmp_st_address[i];
-		  mshr_pipe_st_way_s1[(2*(i+1))-1 -: 2]   = tmp_st_way[i];
-    	  mshr_pipe_st_state_s1[(`L15_MESI_TRANS_STATE_WIDTH*(i+1))-1 -: `L15_MESI_TRANS_STATE_WIDTH] = tmp_st_state[i];
-	 end
+        mshr_pipe_vals_s1[(`L15_NUM_MSHRID_PER_THREAD*(i+1))-1 -: `L15_NUM_MSHRID_PER_THREAD] = tmp_vals[i];
+        mshr_pipe_ld_address[(`L15_PADDR_WIDTH*(i+1))-1 -: `L15_PADDR_WIDTH]  = tmp_ld_address[i];
+        mshr_pipe_st_address[(`L15_PADDR_WIDTH*(i+1))-1 -: `L15_PADDR_WIDTH]  = tmp_st_address[i];
+        mshr_pipe_st_way_s1[(2*(i+1))-1 -: 2]   = tmp_st_way[i];
+        mshr_pipe_st_state_s1[(`L15_MESI_TRANS_STATE_WIDTH*(i+1))-1 -: `L15_MESI_TRANS_STATE_WIDTH] = tmp_st_state[i];
+    end
 
 
     // S1 read
@@ -361,11 +361,11 @@ always @ (posedge clk)
 begin
     if (!rst_n)
     begin
-	 	for(i = 0; i < `L15_NUM_THREADS; i = i+1)
-	 	begin
-		  st_homeid[i] <= '0;
-        ld_homeid[i] <= '0;
-		end
+        for(i = 0; i < `L15_NUM_THREADS; i = i+1)
+        begin
+            st_homeid[i] <= {PACKET_HOME_ID_WIDTH{1'b0}};
+            ld_homeid[i] <= {PACKET_HOME_ID_WIDTH{1'b0}};
+        end
     end
     else
     begin

diff --git a/piton/design/chip/tile/l15/rtl/l15_pipeline.v.pyv b/piton/design/chip/tile/l15/rtl/l15_pipeline.v.pyv
@@ -404,27 +404,15 @@ begin
     // predecode_mshr_read_address_s1 = mshr_pipe_address_s1;
     predecode_mshr_read_homeid_s1 = mshr_pipe_readres_homeid_s1;
 
-    // mshr_val_array
-    //mshr_val_array[0] = mshr_pipe_vals_s1[`L15_NUM_MSHRID_PER_THREAD*1 - 1 -: `L15_NUM_MSHRID_PER_THREAD];
-    //mshr_st_state_array[0] = mshr_pipe_st_state_s1[`L15_MESI_TRANS_STATE_WIDTH*1 - 1 -: `L15_MESI_TRANS_STATE_WIDTH];
-    //mshr_st_address_array[0] = mshr_pipe_st_address[`L15_PADDR_WIDTH*1 - 1 -: `L15_PADDR_WIDTH];
-    //mshr_ld_address_array[0] = mshr_pipe_ld_address[`L15_PADDR_WIDTH*1 - 1 -: `L15_PADDR_WIDTH];
-    //mshr_st_way_array[0] = mshr_pipe_st_way_s1[2*1 - 1 -: 2];
-    //
-    //mshr_val_array[1] = mshr_pipe_vals_s1[`L15_NUM_MSHRID_PER_THREAD*2 - 1 -: `L15_NUM_MSHRID_PER_THREAD];
-    //mshr_st_state_array[1] = mshr_pipe_st_state_s1[`L15_MESI_TRANS_STATE_WIDTH*2 - 1 -: `L15_MESI_TRANS_STATE_WIDTH];
-    //mshr_st_address_array[1] = mshr_pipe_st_address[`L15_PADDR_WIDTH*2 - 1 -: `L15_PADDR_WIDTH];
-    //mshr_ld_address_array[1] = mshr_pipe_ld_address[`L15_PADDR_WIDTH*2 - 1 -: `L15_PADDR_WIDTH];
-    //mshr_st_way_array[1] = mshr_pipe_st_way_s1[2*2 - 1 -: 2];
-
-	 for(i = 0; i < `L15_NUM_THREADS; i = i+1)
-	 begin
-	 	 mshr_val_array[i]        = mshr_pipe_vals_s1[`L15_NUM_MSHRID_PER_THREAD*(i+1) - 1 -: `L15_NUM_MSHRID_PER_THREAD];
-    	 mshr_st_state_array[i]   = mshr_pipe_st_state_s1[`L15_MESI_TRANS_STATE_WIDTH*(i+1) - 1 -: `L15_MESI_TRANS_STATE_WIDTH];
-       mshr_st_address_array[i] = mshr_pipe_st_address[`L15_PADDR_WIDTH*(i+1) - 1 -: `L15_PADDR_WIDTH];
-       mshr_ld_address_array[i] = mshr_pipe_ld_address[`L15_PADDR_WIDTH*(i+1) - 1 -: `L15_PADDR_WIDTH];
-       mshr_st_way_array[i]     = mshr_pipe_st_way_s1[`L15_WAY_WIDTH*(i+1) - 1 -: `L15_WAY_WIDTH]; 
-	 end
+    // mshr_val_array	 
+    for(i = 0; i < `L15_NUM_THREADS; i = i+1)
+    begin
+        mshr_val_array[i]        = mshr_pipe_vals_s1[`L15_NUM_MSHRID_PER_THREAD*(i+1) - 1 -: `L15_NUM_MSHRID_PER_THREAD];
+        mshr_st_state_array[i]   = mshr_pipe_st_state_s1[`L15_MESI_TRANS_STATE_WIDTH*(i+1) - 1 -: `L15_MESI_TRANS_STATE_WIDTH];
+        mshr_st_address_array[i] = mshr_pipe_st_address[`L15_PADDR_WIDTH*(i+1) - 1 -: `L15_PADDR_WIDTH];
+        mshr_ld_address_array[i] = mshr_pipe_ld_address[`L15_PADDR_WIDTH*(i+1) - 1 -: `L15_PADDR_WIDTH];
+        mshr_st_way_array[i]     = mshr_pipe_st_way_s1[`L15_WAY_WIDTH*(i+1) - 1 -: `L15_WAY_WIDTH]; 
+    end
 
 end
 
@@ -450,10 +438,6 @@ reg predecode_tagcheck_matched_trd_ld_s1 [`L15_THREAD_ARRAY_MASK]; // each eleme
 reg predecode_tagcheck_matched_trd_st_s1 [`L15_THREAD_ARRAY_MASK];
 reg predecode_tagcheck_matched_lds_s1;
 reg predecode_tagcheck_matched_sts_s1;
-//reg predecode_tagcheck_matched_t0ld_s1;
-//reg predecode_tagcheck_matched_t0st_s1;
-//reg predecode_tagcheck_matched_t1ld_s1;
-//reg predecode_tagcheck_matched_t1st_s1;
 
 reg predecode_int_vec_dis_s1;
 reg predecode_tagcheck_matched_s1;
@@ -850,28 +834,26 @@ begin
 
     // TAG CHECKING
     predecode_partial_tag_s1[19:4] = pcxdecoder_l15_address[19:4]; // compare partial tag to save energy & timing
-	 predecode_tagcheck_matched_lds_s1 = 0;
-	 predecode_tagcheck_matched_sts_s1 = 0;
+    predecode_tagcheck_matched_lds_s1 = 0;
+    predecode_tagcheck_matched_sts_s1 = 0;
 
-	 for(i = 0; i < `L15_NUM_THREADS; i = i+1)
-	 begin
-	 	 predecode_tagcheck_matched_trd_ld_s1[i] = mshr_val_array[i][`L15_MSHR_ID_LD] 
+    for(i = 0; i < `L15_NUM_THREADS; i = i+1)
+    begin
+        predecode_tagcheck_matched_trd_ld_s1[i] = mshr_val_array[i][`L15_MSHR_ID_LD] 
                                         && (predecode_partial_tag_s1[19:4] == mshr_ld_address_array[i][19:4]);		 
-	 	 predecode_tagcheck_matched_trd_st_s1[i] = mshr_val_array[i][`L15_MSHR_ID_ST] 
-		 											 && (pcxdecoder_l15_address[39:4] == mshr_st_address_array[i][39:4]);
+        predecode_tagcheck_matched_trd_st_s1[i] = mshr_val_array[i][`L15_MSHR_ID_ST] 
+                                        && (pcxdecoder_l15_address[39:4] == mshr_st_address_array[i][39:4]);
 
-    	 predecode_tagcheck_matched_lds_s1 = predecode_tagcheck_matched_trd_ld_s1[i] | predecode_tagcheck_matched_lds_s1;
-     	 predecode_tagcheck_matched_sts_s1 = predecode_tagcheck_matched_trd_st_s1[i] | predecode_tagcheck_matched_sts_s1; 
-
-		 if(predecode_tagcheck_matched_trd_st_s1[i] == 1)
-			 predecode_hit_stbuf_threadid_s1 = i;
+        predecode_tagcheck_matched_lds_s1 = predecode_tagcheck_matched_trd_ld_s1[i] | predecode_tagcheck_matched_lds_s1;
+        predecode_tagcheck_matched_sts_s1 = predecode_tagcheck_matched_trd_st_s1[i] | predecode_tagcheck_matched_sts_s1; 
 
-	 end
+        if(predecode_tagcheck_matched_trd_st_s1[i] == 1)
+            predecode_hit_stbuf_threadid_s1 = i;
+    end
 
-	 predecode_tagcheck_matched_s1 = predecode_tagcheck_matched_lds_s1 | predecode_tagcheck_matched_sts_s1;
+	predecode_tagcheck_matched_s1 = predecode_tagcheck_matched_lds_s1 | predecode_tagcheck_matched_sts_s1;
     // misc
     predecode_hit_stbuf_s1 = predecode_tagcheck_matched_sts_s1;
-    //predecode_hit_stbuf_threadid_s1 = predecode_tagcheck_matched_t1st_s1 ? 1'b1 : 1'b0;
     // note: only work with 2 threads for now; need to change the algo of mshr if need to increase the num of threads
 end
 
@@ -3326,31 +3308,35 @@ reg [`L15_UNPARAM_1_0] stbuf_way_s3; // wmt todo: move calculation to s2
 always @ *
 begin
 
-	 stbuf_compare_match_val_s3 = 0;
-	 stbuf_compare_lru_match_val_s3 = 0;
-	 for(i = 0; i < `L15_NUM_THREADS; i = i+1)
-	 begin
-		  `ifdef PITON_ASIC_RTL
-		 		stbuf_compare_address_match_s3[i] = mshr_st_address_array[i][10:4] == cache_index_s3;
-		  `else
-		 		stbuf_compare_address_match_s3[i] = mshr_st_address_array[i][39:4] == address_s3[39:4];
-		  `endif
+    stbuf_compare_match_val_s3 = 0;
+    stbuf_compare_lru_match_val_s3 = 0;
+    for(i = 0; i < `L15_NUM_THREADS; i = i+1)
+    begin
+        `ifdef PITON_ASIC_RTL
+            stbuf_compare_address_match_s3[i] = mshr_st_address_array[i][10:4] == cache_index_s3;
+        `else
+            stbuf_compare_address_match_s3[i] = mshr_st_address_array[i][39:4] == address_s3[39:4];
+        `endif
 
-		 	stbuf_compare_match_s3[i] = mshr_val_array[i][`L15_MSHR_ID_ST] 
-		 										&& (mshr_st_state_array[i] == `L15_MESI_TRANSITION_STATE_SM) 
-		 										&& (stbuf_compare_address_match_s3[i] == 1'b1);
+        stbuf_compare_match_s3[i] = mshr_val_array[i][`L15_MSHR_ID_ST] 
+                                    && (mshr_st_state_array[i] == `L15_MESI_TRANSITION_STATE_SM) 
+                                    && (stbuf_compare_address_match_s3[i] == 1'b1);
 
-		 	stbuf_compare_lru_match_s3[i] = stbuf_compare_match_s3[i] && (mshr_st_way_array[i] == lru_way_s3);
+        stbuf_compare_lru_match_s3[i] = stbuf_compare_match_s3[i] && (mshr_st_way_array[i] == lru_way_s3);
 
-		 	if (stbuf_compare_match_s3[i] == 1)
+        if (stbuf_compare_match_s3[i] == 1)
+        begin
             stbuf_compare_threadid_s3 = i;
+        end
 
-		 	if (stbuf_compare_lru_match_s3[i] == 1)
+        if (stbuf_compare_lru_match_s3[i] == 1)
+        begin
             stbuf_compare_lru_threadid_s3 = i;
+        end
 
-		   stbuf_compare_match_val_s3     = stbuf_compare_match_s3[i]     | stbuf_compare_match_val_s3;
-		   stbuf_compare_lru_match_val_s3 = stbuf_compare_lru_match_s3[i] | stbuf_compare_lru_match_val_s3;
-	 end
+        stbuf_compare_match_val_s3     = stbuf_compare_match_s3[i]     | stbuf_compare_match_val_s3;
+        stbuf_compare_lru_match_val_s3 = stbuf_compare_lru_match_s3[i] | stbuf_compare_lru_match_val_s3;
+    end
 
     stbuf_way_s3 = mshr_st_way_array[stbuf_compare_threadid_s3];
     // stbuf_way_wmt_data_s3 = wmt_data_s3[stbuf_way_s3];

diff --git a/piton/design/chip/tile/l15/rtl/noc2decoder.v b/piton/design/chip/tile/l15/rtl/noc2decoder.v
@@ -104,11 +104,11 @@ begin
     // the threadid is encoded in the mshrid sent to L2, is the next L15_THREADID_WIDTH bits after the first L15_MSHR_ID_WIDTH bits
     noc2decoder_l15_threadid = noc2_mshrid[`L15_MSHR_ID_WIDTH+`L15_THREADID_WIDTH -1 -: `L15_THREADID_WIDTH];
 
-	 `ifdef NO_RTL_CSM
-    	noc2decoder_l15_hmc_fill = 1'b0; //noc2_mshrid[`MSG_MSHRID_WIDTH-1];
-	 `else
-    	noc2decoder_l15_hmc_fill = noc2_mshrid[`MSG_MSHRID_WIDTH-1];
-	 `endif
+    `ifdef NO_RTL_CSM
+        noc2decoder_l15_hmc_fill = 1'b0; //noc2_mshrid[`MSG_MSHRID_WIDTH-1];
+    `else
+        noc2decoder_l15_hmc_fill = noc2_mshrid[`MSG_MSHRID_WIDTH-1];
+    `endif
 
     noc2decoder_l15_l2miss = noc2_data[`MSG_L2_MISS];
     noc2decoder_l15_icache_type = noc2_data[`MSG_CACHE_TYPE];

diff --git a/piton/design/include/l15.h.pyv b/piton/design/include/l15.h.pyv
@@ -85,6 +85,9 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
       print("`define L15_WMT_ALIAS_WIDTH %d" % int(math.log(l15_set_count/l1d_set_count, 2)))
 
    print("`define L15_CACHELINE_WIDTH %d" % (L15_LINE_SIZE*8))
+
+   print("`define L15_NUM_THREADS     %d" % CONFIG_L15_NUM_THREADS)
+   print("`define L15_THREADID_WIDTH  %d" % int(math.log(CONFIG_L15_NUM_THREADS, 2)))
 %>
 
 `define L15_MESI_STATE_WIDTH 2
@@ -210,13 +213,6 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 `define L15_MSHR_ID_ST 2'd3
 
 // if NO_RTL_CSM is defined L15_NUM_THREADS could go up to 64, otherwise 32.
-<%
-   import math
-   import pyhplib
-   from pyhplib import *
-   print("`define L15_NUM_THREADS     %d" % L15_NUM_THREADS)
-   print("`define L15_THREADID_WIDTH  %d" % int(math.log(L15_NUM_THREADS, 2)))
-%>
 `define L15_THREADID_MASK `L15_THREADID_WIDTH-1:0
 `define L15_THREAD_ARRAY_MASK `L15_NUM_THREADS-1:0
 `define L15_NUM_MSHRID_PER_THREAD 4

diff --git a/piton/tools/bin/pyhplib.py b/piton/tools/bin/pyhplib.py
@@ -32,8 +32,6 @@
 MAX_X = 8;
 MAX_Y = 8;
 
-
-
 PITON_X_TILES = int(os.environ.get('PITON_X_TILES', '-1'))
 #print "//x_tiles:", num_tiles
 
@@ -74,6 +72,7 @@
 # cache configurations
 CONFIG_L15_SIZE = int(os.environ.get('CONFIG_L15_SIZE', '8192'))
 CONFIG_L15_ASSOCIATIVITY = int(os.environ.get('CONFIG_L15_ASSOCIATIVITY', '4'))
+CONFIG_L15_NUM_THREADS = int(os.environ.get('CONFIG_L15_NUM_THREADS', '2'))
 CONFIG_L1D_SIZE = int(os.environ.get('CONFIG_L1D_SIZE', '8192'))
 CONFIG_L1D_ASSOCIATIVITY = int(os.environ.get('CONFIG_L1D_ASSOCIATIVITY', '4'))
 CONFIG_L1I_SIZE = int(os.environ.get('CONFIG_L1I_SIZE', '16384'))
@@ -89,8 +88,6 @@
 L15_LINE_SIZE = 16
 L2_LINE_SIZE = 64
 
-L15_NUM_THREADS = int(os.environ.get('L15_NUM_THREADS', '2'))
-
 #########################################################
 # BRAM configurations
 #########################################################

diff --git a/piton/tools/src/sims/manycore.config b/piton/tools/src/sims/manycore.config
@@ -107,14 +107,12 @@
 #ifdef FLIST_ORAM
     -flist=$DV_ROOT/design/chip/tinyoram/rtl/Flist.oram
     -config_rtl=ORAM_ON"
-    -config_rtl=NO_RTL_CSM
     -sim_run_args=+oram"
 #endif
 
     // No scan chains
     -config_rtl=NO_SCAN
-	 -config_rtl=NO_RTL_CSM
-
+
     -config_l1i_size=16384       // default
     -config_l1i_associativity=4     // default
 

diff --git a/piton/tools/src/sims/sims,2.0 b/piton/tools/src/sims/sims,2.0
@@ -2637,6 +2637,7 @@ sub parse_args
     "config_l1d_associativity",
     "config_l15_size",
     "config_l15_associativity",
+    "config_l15_num_threads",
     "config_l2_size",
     "config_l2_associativity",
     );
@@ -2651,12 +2652,6 @@ sub parse_args
     }
   }
 
-  GetOptions (\%opt,"l15_num_threads=s",
-    );
-    if ($opt{l15_num_threads}) {
-        $ENV{L15_NUM_THREADS} = $opt{l15_num_threads};
-    }
-
   $ENV{PROTOSYN_RUNTIME_DESIGN_PATH} = $ENV{DV_ROOT} . "/verif/env/manycore";
   $ENV{PROTOSYN_RUNTIME_BOARD} = "";
 

diff --git a/piton/tools/verilator/mcs_map_info.h b/piton/tools/verilator/mcs_map_info.h
diff --git a/piton/tools/verilator/my_top.cpp b/piton/tools/verilator/my_top.cpp
@@ -38,7 +38,7 @@ Vcmp_top* top;
 VerilatedVcdC* tfp;
 #endif
 
-extern "C" void metro_mpi_init_jbus_model_call(const char *str, int oram);
+extern "C" void init_jbus_model_call(char *str, int oram);
 
 // This is a 64-bit integer to reduce wrap over issues and
 // // allow modulus. You can also use a double, if you wish.