fix rtl gpr zero

driver/rtlsim/Makefile

@@ -3,8 +3,8 @@ CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -Wfatal-errors
 
 CFLAGS += -I../../include -I../../../hw/simulate -I../../../runtime
 
-MULTICORE += -DNUM_CLUSTERS=2 -DNUM_CORES=2
-#DEBUG = 1
+#MULTICORE += -DNUM_CLUSTERS=2 -DNUM_CORES=2
+DEBUG = 1
 
 CFLAGS += -fPIC
 

driver/tests/demo/Makefile

@@ -6,12 +6,12 @@ VX_CXX = $(RISCV_TOOL_PATH)/bin/riscv32-unknown-elf-g++
 VX_DMP = $(RISCV_TOOL_PATH)/bin/riscv32-unknown-elf-objdump
 VX_CPY = $(RISCV_TOOL_PATH)/bin/riscv32-unknown-elf-objcopy
 
-VX_NEWLIB = $(VX_RT_PATH)/newlib/newlib.c
+#VX_NEWLIB = $(VX_RT_PATH)/newlib/newlib.c
 VX_STR = $(VX_RT_PATH)/startup/vx_start.S
 VX_INT = $(VX_RT_PATH)/intrinsics/vx_intrinsics.S
-VX_IO  = $(VX_RT_PATH)/io/vx_io.S $(VX_RT_PATH)/io/vx_io.c
+#VX_IO  = $(VX_RT_PATH)/io/vx_io.S $(VX_RT_PATH)/io/vx_io.c
 VX_API = $(VX_RT_PATH)/vx_api/vx_api.c
-VX_FIO = $(VX_RT_PATH)/fileio/fileio.S
+#VX_FIO = $(VX_RT_PATH)/fileio/fileio.S
 
 VX_CFLAGS = -march=rv32im -mabi=ilp32 -O3 -Wl,-Bstatic,-T,$(VX_RT_PATH)/startup/vx_link.ld -ffreestanding -nostartfiles -Wl,--gc-sections
 

driver/tests/demo/common.h

@@ -1,7 +1,7 @@
 #ifndef _COMMON_H_
 #define _COMMON_H_
 
-#define KERNEL_ARG_DEV_MEM_ADDR 0x7fffff00
+#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
 
 struct kernel_arg_t {
   uint32_t num_warps;

driver/tests/demo/demo.cpp

@@ -224,12 +224,6 @@ int main(int argc, char *argv[]) {
     cleanup();
     return ret;
   }
-  
-  ret = run_test(device, buffer, kernel_arg, buf_size, num_points);
-  if (ret != 0) {
-    cleanup();
-    return ret;
-  }
 
   // cleanup
   std::cout << "cleanup" << std::endl;  

hw/rtl/VX_gpr.v

@@ -7,13 +7,18 @@ module VX_gpr (
     VX_gpr_read_if  gpr_read_if,
     VX_wb_if        writeback_if,
 
-    output reg[`NUM_THREADS-1:0][`NUM_GPRS-1:0] a_reg_data,
-    output reg[`NUM_THREADS-1:0][`NUM_GPRS-1:0] b_reg_data
+    output wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] a_reg_data,
+    output wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] b_reg_data
 );
-    wire write_enable;
+    wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] a_reg_data_uqual;
+    wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] b_reg_data_uqual;
+
+    assign a_reg_data = (gpr_read_if.rs1 != 0) ? a_reg_data_uqual : 0;
+    assign b_reg_data = (gpr_read_if.rs2 != 0) ? b_reg_data_uqual : 0;
+
+    wire write_enable = valid_write_request && ((writeback_if.wb != 0));
     
     `ifndef ASIC
-        assign write_enable = valid_write_request && ((writeback_if.wb != 0)) && (writeback_if.rd != 0);
 
         VX_gpr_ram gpr_ram (
             .we    (write_enable),
@@ -24,16 +29,17 @@ module VX_gpr (
             .raddr2(gpr_read_if.rs2),
             .be    (writeback_if.valid),
             .wdata (writeback_if.data),
-            .q1    (a_reg_data),
-            .q2    (b_reg_data)
+            .q1    (a_reg_data_uqual),
+            .q2    (b_reg_data_uqual)
         );
+
     `else 
-        assign write_enable = valid_write_request && ((writeback_if.wb != 0));
+
         wire going_to_write = write_enable & (| writeback_if.wb_valid);
         wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] write_bit_mask;
 
         genvar i;
-        for (i = 0; i < `NUM_THREADS; i=i+1) begin
+        for (i = 0; i < `NUM_THREADS; i = i + 1) begin
             wire local_write = write_enable & writeback_if.wb_valid[i];
             assign write_bit_mask[i] = {`NUM_GPRS{~local_write}};
         end
@@ -52,17 +58,17 @@ module VX_gpr (
     `ifndef SYN
         genvar j;
         for (i = 0; i < `NUM_THREADS; i = i + 1) begin
-            for (j = 0; j < `NUM_GPRS; j=j+1) begin
-                assign a_reg_data[i][j] = ((temp_a[i][j] === 1'dx) || cena_1 )? 1'b0 : temp_a[i][j];
-                assign b_reg_data[i][j] = ((temp_b[i][j] === 1'dx) || cena_2) ? 1'b0 : temp_b[i][j];
+            for (j = 0; j < `NUM_GPRS; j = j + 1) begin
+                assign a_reg_data_uqual[i][j] = ((temp_a[i][j] === 1'dx) || cena_1 )? 1'b0 : temp_a[i][j];
+                assign b_reg_data_uqual[i][j] = ((temp_b[i][j] === 1'dx) || cena_2) ? 1'b0 : temp_b[i][j];
             end
         end
     `else
-        assign a_reg_data = temp_a;
-        assign b_reg_data = temp_b;
+        assign a_reg_data_uqual = temp_a;
+        assign b_reg_data_uqual = temp_b;
     `endif
 
-        wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] to_write = (writeback_if.rd != 0) ? writeback_if.write_data : 0;
+        wire[`NUM_THREADS-1:0][`NUM_GPRS-1:0] to_write = writeback_if.write_data;
 
         for (i = 0; i < 'NT; i=i+4)
         begin

hw/simulate/simulator.cpp

@@ -111,11 +111,12 @@ void Simulator::io_driver() {
   vortex_->io_req_ready = true;
 }
 
-void Simulator::reset() {  
-  time_stamp = 0;
+void Simulator::reset() {      
   vortex_->reset = 1;
-  this->step();
+  this->step();  
   vortex_->reset = 0;
+
+  dram_rsp_vec_.clear();
 }
 
 void Simulator::step() {
@@ -150,6 +151,7 @@ bool Simulator::is_busy() {
 void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {  
   // send snoop requests to the caches
   printf("[sim] total cycles: %ld\n", time_stamp/2);
+  
   // align address to LLC block boundaries
   auto aligned_addr_start = mem_addr / GLOBAL_BLOCK_SIZE;
   auto aligned_addr_end = (mem_addr + size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE;

runtime/newlib/newlib.c

@@ -274,14 +274,6 @@ void * _sbrk (int nbytes)
  //    }
 }       /* _sbrk () */
 
-
-void _exit(int val)
-{
-	// vx_print_str("Hello from exit\n");
-	vx_tmc(0);
-}
-
-
 int _open(const char *name, int flags, int mode)
 {
 	// char * write_buffer = (char *) FILE_IO_WRITE;

runtime/startup/vx_start.S

@@ -4,27 +4,35 @@
 .global _start
 .type   _start, @function
 _start:
-    la a1, vx_set_sp
-    li a0, NUM_WARPS    # activate all warps
-    .word 0x00b5106b # wspawn a0(numWarps), a1(PC SPAWN)
-    jal vx_set_sp
-    li a0, 1
-    .word 0x0005006b    # back to single thread
-    # Initialize global pointerp
-    # call __cxx_global_var_init
-    # Clear the bss segment
-    la      a0, _edata
-    la      a2, _end
-    sub     a2, a2, a0
-    li      a1, 0
-    call    memset
-    la      a0, __libc_fini_array   # Register global termination functions
-    call    atexit                  # to be called upon exit
-    call    __libc_init_array       # Run global initialization functions
-    call    main
-    tail    exit
+  la a1, vx_set_sp
+  li a0, NUM_WARPS    # activate all warps
+  .word 0x00b5106b # wspawn a0(numWarps), a1(PC SPAWN)
+  jal vx_set_sp
+  li a0, 1
+  .word 0x0005006b    # back to single thread
+  # Initialize global pointerp
+  # call __cxx_global_var_init
+  # Clear the bss segment
+  la      a0, _edata
+  la      a2, _end
+  sub     a2, a2, a0
+  li      a1, 0
+  call    memset
+  la      a0, __libc_fini_array   # Register global termination functions
+  call    atexit                  # to be called upon exit
+  call    __libc_init_array       # Run global initialization functions
+  call    main
+  tail    exit
 .size  _start, .-_start
 
+.section .text
+.type _exit, @function
+.global _exit
+_exit:
+  li a0, 0
+  .word 0x0005006b    # disable all threads
+
+
 .section .text
 .type vx_set_sp, @function
 .global vx_set_sp

runtime/tests/simple/Makefile

@@ -1,8 +1,8 @@
 
 COMP     = ~/dev/riscv-gnu-toolchain/drops/bin/riscv32-unknown-elf-g++
-#COMP     = /opt/riscv-new/drops/bin/riscv32-unknown-elf-g++
 
-CC_FLAGS = -march=rv32im -mabi=ilp32 -O3 -Wl,-Bstatic,-T,../../startup/vx_link.ld -ffreestanding -nostdlib
+CC_FLAGS = -march=rv32im -mabi=ilp32 -O3 -Wl,-Bstatic,-T,../../startup/vx_link.ld 
+CC_FLAGS += -nostartfiles -ffreestanding -fno-rtti -fno-exceptions -Wl,--gc-sections 
 
 DMP  = ~/dev/riscv-gnu-toolchain/drops/bin/riscv32-unknown-elf-objdump
 CPY  = ~/dev/riscv-gnu-toolchain/drops/bin/riscv32-unknown-elf-objcopy
@@ -17,8 +17,9 @@ VX_FIO  = ../../fileio/fileio.S
 
 VX_MAIN = vx_simple_main
 
-LIBS    = ~/dev/riscv-gnu-toolchain/drops/riscv32-unknown-elf/lib/libc.a ~/dev/riscv-gnu-toolchain/drops/riscv32-unknown-elf/lib/libstdc++.a -static-libgcc -lgcc
-#LIBS    = /opt/riscv-new/drops/riscv32-unknown-elf/lib/libc.a /opt/riscv-new/drops/riscv32-unknown-elf/lib/libstdc++.a -static-libgcc -lgcc
+#LIBS  += ~/dev/riscv-gnu-toolchain/drops/riscv32-unknown-elf/lib/libc.a 
+#LIBS  += ~/dev/riscv-gnu-toolchain/drops/riscv32-unknown-elf/lib/libstdc++.a 
+#LIBS  += -static-libgcc -lgcc
 
 VX_SRCS = vx_simple_main.c tests.c