Merge pull request #39 from pulp-platform/priv-fpu

Added private FPUs.
pulp-platform · Jan 26, 2024 · 2c9ff46 · 2c9ff46
2 parents bd2cd62 + 388261e
commit 2c9ff46
Show file tree

Hide file tree

Showing 10 changed files with 388 additions and 269 deletions.
diff --git a/Bender.local b/Bender.local
@@ -1,4 +1,5 @@
 overrides:
+  hci                 : { git: "https://github.com/pulp-platform/hci.git"                 , rev: 3cb3d99b2cebfeed55cb6ab9d98fce7b99e97cb9 } # branch: master
   axi                 : { git: "https://github.com/pulp-platform/axi.git"                 , version: =0.39.1-beta                         }
   register_interface  : { git: "https://github.com/pulp-platform/register_interface.git"  , rev: 19163bb5191d2669a8cbc267cdd4ce8e60f20746 } # branch: master
   cluster_interconnect: { git: "https://github.com/pulp-platform/cluster_interconnect.git", rev: 89e1019d64a86425211be6200770576cbdf3e8b3 } # branch: assertion-fix
diff --git a/Bender.lock b/Bender.lock
@@ -105,7 +105,7 @@ packages:
     dependencies:
     - common_cells
   hci:
-    revision: 78fb8fc8a6f2c376554562c47755b4d0febaba25
+    revision: 3cb3d99b2cebfeed55cb6ab9d98fce7b99e97cb9
     version: null
     source:
       Git: https://github.com/pulp-platform/hci.git
@@ -133,8 +133,8 @@ packages:
     dependencies:
     - tech_cells_generic
   hwpe-stream:
-    revision: ddc154424187dff42a8fcec946c768ceb13f13de
-    version: 1.6.4
+    revision: 389bd7fb1975d2df1546910c5f220c668122e646
+    version: 1.6.5
     source:
       Git: https://github.com/pulp-platform/hwpe-stream.git
     dependencies:
@@ -216,7 +216,7 @@ packages:
     - common_cells
     - common_verification
   riscv:
-    revision: a1dcae35edae6092ddbf92c424690cb903b678d5
+    revision: c760db14dbd6cc3ec3b8ae8274df2eac7225bcac
     version: null
     source:
       Git: git@github.com:AlSaqr-platform/riscv_nn.git

diff --git a/Bender.yml b/Bender.yml
@@ -21,17 +21,16 @@ dependencies:
   idma:                   { git: "https://github.com/pulp-platform/iDMA.git", rev: 437ffa9dac5dea0daccfd3e8ae604d4f6ae2cdf1 } # branch: master
   hier-icache:            { git: "https://github.com/pulp-platform/hier-icache.git", rev: "a971e364bf8090cf77fafad995b480c1ac7ea4e0" } # branch: yt/carfield
   cluster_peripherals:    { git: "https://github.com/pulp-platform/cluster_peripherals.git", rev: c015839816938a790c8da5fd5829cfc536f1ca9c } # branch: yt/return-reg
-  # fpu_interco:            { git: "https://github.com/pulp-platform/fpu_interco.git", rev: "4aec4b68424947b0c4cf25fd7c4b907cb9ec3dfa" } # branch: yt/carfield
   axi:                    { git: "https://github.com/pulp-platform/axi.git", version: =0.39.1-beta  }
   axi_slice:              { git: "https://github.com/pulp-platform/axi_slice.git", version: 1.1.4 } # deprecated, replaced by axi_cut (in axi repo)
   timer_unit:             { git: "https://github.com/pulp-platform/timer_unit.git", version: 1.0.2 }
   common_cells:           { git: "https://github.com/pulp-platform/common_cells.git", version: 1.21.0 }
   tech_cells_generic:     { git: "https://github.com/pulp-platform/tech_cells_generic.git", version: 0.2.3 }
-  riscv:                  { git: "git@github.com:AlSaqr-platform/riscv_nn.git", rev: a1dcae35edae6092ddbf92c424690cb903b678d5 } # branch: yt/hmr
+  riscv:                  { git: "git@github.com:AlSaqr-platform/riscv_nn.git", rev: astral-v1.0 }
   cv32e40p:               { git: "https://github.com/pulp-platform/cv32e40p.git", rev: e863f576699815b38cc9d80dbdede8ed5efd5991 } # `michaero/safety-island-clic` branch
   ibex:                   { git: "https://github.com/pulp-platform/ibex.git", rev: "pulpissimo-v6.1.2" }
   scm:                    { git: "https://github.com/pulp-platform/scm.git", rev: 74426dee36f28ae1c02f7635cf844a0156145320 } # branch: yt/bump-clkgating
-  hci:                    { git: "https://github.com/pulp-platform/hci.git", rev: b2e6f391aa6c10c03f45b693d80a0aaddecf169b } # branch: master
+  hci:                    { git: "https://github.com/pulp-platform/hci.git", rev: 3cb3d99b2cebfeed55cb6ab9d98fce7b99e97cb9 } # branch: test_mode_fix
   register_interface:     { git: "https://github.com/pulp-platform/register_interface.git", rev: 19163bb5191d2669a8cbc267cdd4ce8e60f20746  } # branch: master
   common_cells:           { git: "https://github.com/pulp-platform/common_cells.git",  version: 1.29.0 }
   redundancy_cells:       { git: "https://github.com/pulp-platform/redundancy_cells.git", rev: 32023555679cfdb8a0a073ad4c17fc3a5d1ddea5 } # branch: yt/rapidrecovery

diff --git a/Makefile b/Makefile
@@ -4,10 +4,11 @@
 
 ROOT_DIR      = $(strip $(shell dirname $(realpath $(lastword $(MAKEFILE_LIST)))))
 
+QUESTA ?= questa-2022.3
 GIT ?= git
 BENDER ?= bender
-VSIM ?= vsim
-VOPT ?= vopt
+VSIM ?= $(QUESTA) vsim
+VOPT ?= $(QUESTA) vopt
 top_level ?= pulp_cluster_tb
 dpi-library ?= work-dpi
 library ?= work
@@ -76,11 +77,11 @@ Bender.lock:
 
 ## Clone pulp-runtime as SW stack
 pulp-runtime:
-	git clone git@github.com:pulp-platform/pulp-runtime.git -b yt/carfield $@
+	git clone git@github.com:pulp-platform/pulp-runtime.git -b astral $@
 
 ## Clone regression tests for bare-metal verification
 regression-tests:
-	git clone git@github.com:pulp-platform/regression_tests.git -b yt/carfield $@
+	git clone git@github.com:pulp-platform/regression_tests.git -b astral $@
 
 ########################
 # Build and simulation #
@@ -92,30 +93,22 @@ sim_clean:
 
 scripts/compile.tcl: | Bender.lock
 	$(call generate_vsim, $@, $(bender_defs) $(bender_targs),..)
-
-# compile the elfloader.cpp
-$(dpi-library)/%.o: tb/dpi/%.cc $(dpi_hdr)
-	mkdir -p $(dpi-library)
-	$(CXX) -shared -fPIC -std=c++0x -Bsymbolic $(CFLAGS) -c $< -o $@
-
-$(dpi-library)/cl_dpi.so: $(dpi)
-	$(CXX) -shared -m64 -o $(dpi-library)/cl_dpi.so $? -L$(RISCV)/lib -L$(SPIKE_ROOT)/lib -Wl,-rpath,$(RISCV)/lib -Wl,-rpath,$(SPIKE_ROOT)/lib -lfesvr
+	echo 'vlog "$(realpath $(ROOT_DIR))/tb/dpi/elfloader.cpp" -ccflags "-std=c++11"' >> $@
+	echo 'vopt +permissive -suppress 3053 -suppress 8885 +UVM_NO_RELNOTES $(top_level) -o $(top_level)_optimized'
 
 $(library):
 	$(QUESTA) vlib $(library)
 
-compile: $(library) $(dpi) $(dpi-library)/cl_dpi.so
+compile: $(library)
 	@test -f Bender.lock || { echo "ERROR: Bender.lock file does not exist. Did you run make checkout in bender mode?"; exit 1; }
 	@test -f scripts/compile.tcl || { echo "ERROR: scripts/compile.tcl file does not exist. Did you run make scripts in bender mode?"; exit 1; }
-	$(VSIM) -c -do 'source scripts/compile.tcl; quit'
+	$(VSIM) -c -do 'quit -code [source scripts/compile.tcl]'
 
-build: compile $(dpi)
+build: compile
 	$(VOPT) $(compile_flag) -suppress 3053 -suppress 8885 -work $(library)  $(top_level) -o $(top_level)_optimized +acc -check_synthesis
 
-
 run:
-	$(VSIM) +permissive $(questa-flags) $(questa-cmd) -suppress 3053 -suppress 8885 -lib $(library)  +MAX_CYCLES=$(max_cycles) +UVM_TESTNAME=$(test_case) +APP=$(elf-bin) +notimingchecks +nospecify  -t 1ps \
-	$(uvm-flags) $(QUESTASIM_FLAGS) -sv_lib $(dpi-library)/cl_dpi  \
+	$(VSIM) +permissive $(questa-flags) $(uvm-flags) $(QUESTASIM_FLAGS) $(questa-cmd) -suppress 3053 -suppress 8885 -lib $(library)  +MAX_CYCLES=$(max_cycles) +UVM_TESTNAME=$(test_case) +APP=$(elf-bin) +notimingchecks +nospecify  -t 1ps \
 	${top_level}_optimized +permissive-off ++$(elf-bin) ++$(target-options) ++$(cl-bin) | tee sim.log
 
 .PHONY: test-rt-par-bare

diff --git a/rtl/core_region.sv b/rtl/core_region.sv
@@ -238,7 +238,7 @@ import rapid_recovery_pkg::*;
         .PULP_CLUSTER        ( 1                           ),
         .FPU                 ( FPU                         ),
         .N_EXT_PERF_COUNTERS ( N_EXT_PERF_COUNTERS_ACTUAL  ),
-        .Zfinx               ( 0                           ),
+        .Zfinx               ( FPU                         ),
         .WAPUTYPE            ( WAPUTYPE                    ),
         .DM_HaltAddress      ( DEBUG_START_ADDR + 16'h0800 )
       ) RI5CY_CORE             (

diff --git a/rtl/pulp_cluster.sv b/rtl/pulp_cluster.sv
@@ -27,6 +27,7 @@ module pulp_cluster
   import pulp_cluster_package::*;
   import hci_package::*;
   import rapid_recovery_pkg::*;
+  import fpnew_pkg::*;
 #(
   // cluster parameters
   parameter CORE_TYPE_CL       = 1, // 0 for CV32, 1 for RI5CY, 2 for IBEX RV32IMC
@@ -69,7 +70,8 @@ module pulp_cluster
   parameter BOOT_ADDR               = 32'h78000000,
   parameter INSTR_RDATA_WIDTH       = 32,
 
-  parameter CLUST_FPU               = 0,
+  parameter bit CLUST_FPU           = 1,
+  parameter int unsigned NumFpus    = NB_CORES,
   parameter CLUST_FP_DIVSQRT        = 0,
   parameter CLUST_SHARED_FP         = 0,
   parameter CLUST_SHARED_FP_DIVSQRT = 0,
@@ -422,21 +424,6 @@ hci_core_intf #(
 // cores -> event unit ctrl
 XBAR_PERIPH_BUS s_core_euctrl_bus[NB_CORES-1:0]();
 
-// apu-interconnect
-// handshake signals
-logic [NB_CORES-1:0]                           s_apu_master_req;
-logic [NB_CORES-1:0]                           s_apu_master_gnt;
-// request channel
-logic [NB_CORES-1:0][APU_NARGS_CPU-1:0][31:0]  s_apu_master_operands;
-logic [NB_CORES-1:0][APU_WOP_CPU-1:0]          s_apu_master_op;
-logic [NB_CORES-1:0][WAPUTYPE-1:0]             s_apu_master_type;
-logic [NB_CORES-1:0][APU_NDSFLAGS_CPU-1:0]     s_apu_master_flags;
-// response channel
-logic [NB_CORES-1:0]                           s_apu_master_rready;
-logic [NB_CORES-1:0]                           s_apu_master_rvalid;
-logic [NB_CORES-1:0][31:0]                     s_apu_master_rdata;
-logic [NB_CORES-1:0][APU_NUSFLAGS_CPU-1:0]     s_apu_master_rflags;
-
 //----------------------------------------------------------------------//
 // Interfaces between ICache - L0 - Icache_Interco and Icache_ctrl_unit //
 //                                                                      //
@@ -954,16 +941,16 @@ generate
       .pc_backup_o         ( backup_bus[i].pc_backup      ),
       .csr_backup_o        ( backup_bus[i].csr_backup     ),
       //apu interface
-      .apu_master_req_o      ( s_apu_master_req     [i] ),
-      .apu_master_gnt_i      ( s_apu_master_gnt     [i] ),
-      .apu_master_type_o     ( s_apu_master_type    [i] ),
-      .apu_master_operands_o ( s_apu_master_operands[i] ),
-      .apu_master_op_o       ( s_apu_master_op      [i] ),
-      .apu_master_flags_o    ( s_apu_master_flags   [i] ),
-      .apu_master_valid_i    ( s_apu_master_rvalid  [i] ),
-      .apu_master_ready_o    ( s_apu_master_rready  [i] ),
-      .apu_master_result_i   ( s_apu_master_rdata   [i] ),
-      .apu_master_flags_i    ( s_apu_master_rflags  [i] )
+      .apu_master_req_o      (    ),
+      .apu_master_gnt_i      ( '0 ),
+      .apu_master_type_o     (    ),
+      .apu_master_operands_o (    ),
+      .apu_master_op_o       (    ),
+      .apu_master_flags_o    (    ),
+      .apu_master_valid_i    ( '0 ),
+      .apu_master_ready_o    (    ),
+      .apu_master_result_i   ( '0 ),
+      .apu_master_flags_i    ( '0 )
     );
 
     assign dbg_core_halted[i] = core2hmr[i].debug_halted;
@@ -1111,84 +1098,6 @@ hmr_unit #(
   .core_bus_outputs_i     ( '0           )
 );
 
-//****************************************************
-//**** Shared FPU cluster - Shared execution units ***
-//****************************************************
-// request channel
-logic [NB_CORES-1:0][2:0][31:0]                s_apu__operands;
-logic [NB_CORES-1:0][5:0]                      s_apu__op;
-logic [NB_CORES-1:0][2:0]                      s_apu__type;
-logic [NB_CORES-1:0][14:0]                     s_apu__flags;
-// response channel
-logic [NB_CORES-1:0][4:0]                      s_apu__rflags;
-
-genvar k;
-for(k=0;k<NB_CORES;k++)
-begin
-  assign s_apu__operands[k][2:0] = s_apu_master_operands[k][2:0];
-  assign s_apu__op[k][5:0]       = s_apu_master_op[k][5:0];
-  assign s_apu__type[k][2:0]     = s_apu_master_type[k][2:0];
-  assign s_apu__flags[k][14:0]   = s_apu_master_flags[k][14:0];
-  assign s_apu_master_rflags[k][4:0] = s_apu__rflags[k][4:0];
-end
-
-generate
-  if (CLUST_FPU) begin
-    shared_fpu_cluster #(
-      .NB_CORES         ( NB_CORES          ),
-      .NB_APUS          ( 1                 ),
-      .NB_FPNEW         ( 4                 ),
-      .FP_TYPE_WIDTH    ( 3                 ),
-
-      .NB_CORE_ARGS      ( 3                ),
-      .CORE_DATA_WIDTH   ( 32               ),
-      .CORE_OPCODE_WIDTH ( 6                ),
-      .CORE_DSFLAGS_CPU  ( 15               ),
-      .CORE_USFLAGS_CPU  ( 5                ),
-
-      .NB_APU_ARGS      ( 2                 ),
-      .APU_OPCODE_WIDTH ( 6                 ),
-      .APU_DSFLAGS_CPU  ( 15                ),
-      .APU_USFLAGS_CPU  ( 5                 ),
-
-      .NB_FPNEW_ARGS        ( 3             ), //= 3,
-      .FPNEW_OPCODE_WIDTH   ( 6             ), //= 6,
-      .FPNEW_DSFLAGS_CPU    ( 15            ), //= 15,
-      .FPNEW_USFLAGS_CPU    ( 5             ), //= 5,
-
-      .APUTYPE_ID       ( 1                 ),
-      .FPNEWTYPE_ID     ( 0                 ),
-
-      .C_FPNEW_FMTBITS     (fpnew_pkg::FP_FORMAT_BITS  ),
-      .C_FPNEW_IFMTBITS    (fpnew_pkg::INT_FORMAT_BITS ),
-      .C_ROUND_BITS        (3                          ),
-      .C_FPNEW_OPBITS      (fpnew_pkg::OP_BITS         ),
-      .USE_FPU_OPT_ALLOC   ("FALSE"),
-      .USE_FPNEW_OPT_ALLOC ("TRUE"),
-      .FPNEW_INTECO_TYPE   ("SINGLE_INTERCO")
-    ) i_shared_fpu_cluster (
-      .clk                   ( clk_i                   ),
-      .rst_n                 ( rst_ni                  ),
-      .test_mode_i           ( test_mode_i             ),
-      .core_slave_req_i      ( s_apu_master_req        ),
-      .core_slave_gnt_o      ( s_apu_master_gnt        ),
-      .core_slave_type_i     ( s_apu__type             ),
-      .core_slave_operands_i ( s_apu__operands         ),
-      .core_slave_op_i       ( s_apu__op               ),
-      .core_slave_flags_i    ( s_apu__flags            ),
-      .core_slave_rready_i   ( s_apu_master_rready     ),
-      .core_slave_rvalid_o   ( s_apu_master_rvalid     ),
-      .core_slave_rdata_o    ( s_apu_master_rdata      ),
-      .core_slave_rflags_o   ( s_apu__rflags           )
-    );
-  end else begin
-    assign s_apu_master_gnt    = '0;
-    assign s_apu_master_rvalid = '0;
-    assign s_apu_master_rdata  = '0;
-    assign s_apu__rflags       = '0;
-  end
-endgenerate
-
 //**************************************************************
 //**** HW Processing Engines / Cluster-Coupled Accelerators ****
 //**************************************************************

diff --git a/scripts/start.tcl b/scripts/start.tcl
@@ -3,7 +3,7 @@ if {![info exists VSIM_PATH ]} {
     set VSIM_PATH ""
 }
 
-vsim +permissive -suppress 3053 -suppress 8885 -lib $VSIM_PATH/work +APP=./build/test/test +notimingchecks +nospecify  -t 1ps -sv_lib $VSIM_PATH/work-dpi/cl_dpi  pulp_cluster_tb_optimized +permissive-off ++./build/test/test
+vsim +permissive -suppress 3053 -suppress 8885 +UVM_NO_RELNOTES -lib $VSIM_PATH/work +APP=./build/test/test +notimingchecks +nospecify  -t 1ps pulp_cluster_tb_optimized +permissive-off ++./build/test/test
 
 add log -r /*
 run -all