vecadd bug fixing + ebreak for termination

emulator/core.cpp

@@ -203,7 +203,7 @@ void Warp::step() {
     D(3, "Register state:");
     for (unsigned i = 0; i < reg[0].size(); ++i) {
       D_RAW("  %r" << setfill(' ') << setw(2) << dec << i << ':');
-      for (unsigned j = 0; j < reg.size(); ++j) 
+      for (unsigned j = 0; j < (this->activeThreads); ++j) 
         D_RAW(' ' << setfill('0') << setw(8) << hex << reg[j][i] << setfill(' ') << ' ');
       D_RAW('(' << shadowReg[i] << ')' << endl);
     }

rtl/VX_decode.v

@@ -8,7 +8,9 @@ module VX_decode(
 	// Outputs
 	VX_frE_to_bckE_req_inter VX_frE_to_bckE_req,
 	VX_wstall_inter          VX_wstall,
-	VX_join_inter            VX_join
+	VX_join_inter            VX_join,
+
+	output wire              terminate_sim
 
 );
 
@@ -224,7 +226,7 @@ module VX_decode(
 		wire ebreak = (curr_opcode == `SYS_INST) && (jal_sys_jal && (|in_valid));
 		assign VX_frE_to_bckE_req.ebreak = ebreak;
 		wire out_ebreak = ebreak;
-
+		assign terminate_sim = is_e_inst;
 
 
 		// CSR

rtl/VX_front_end.v

@@ -31,6 +31,11 @@ wire total_freeze = schedule_delay;
 // wire real_fetch_ebreak;
 /* verilator lint_on UNUSED */
 
+wire vortex_ebreak;
+wire terminate_sim;
+
+assign fetch_ebreak = vortex_ebreak || terminate_sim;
+
 
 VX_wstall_inter          VX_wstall();
 VX_join_inter            VX_join();
@@ -47,7 +52,7 @@ VX_fetch vx_fetch(
 
 		.icache_request     (icache_request_fe),
 		.VX_branch_rsp      (VX_branch_rsp),
-		.out_ebreak         (fetch_ebreak), // fetch_ebreak
+		.out_ebreak         (vortex_ebreak), // fetch_ebreak
 		.fe_inst_meta_fd    (fe_inst_meta_fd)
 	);
 
@@ -64,7 +69,8 @@ VX_decode vx_decode(
 		.fd_inst_meta_de   (fd_inst_meta_de),
 		.VX_frE_to_bckE_req(VX_frE_to_bckE_req),
 		.VX_wstall         (VX_wstall),
-		.VX_join           (VX_join)
+		.VX_join           (VX_join),
+		.terminate_sim     (terminate_sim)
 	);
 
 wire no_br_stall = 0;

rtl/modelsim/vortex_tb.v

@@ -88,6 +88,7 @@ reg         out_ebreak;
 	initial begin
 		// $fdumpfile("vortex1.vcd");
 		load_file("../../runtime/mains/simple/vx_simple_main.hex");
+		// load_file("../../emulator/riscv_tests/rv32ui-p-add.hex");
 		//load_file("../../kernel/vortex_test.hex");
 		$dumpvars(0, vortex_tb);
 		reset = 1;

runtime/mains/vecadd/Makefile

@@ -14,7 +14,7 @@ VX_IO   = ../../io/vx_io.s ../../io/vx_io.c
 VX_API  = ../../vx_api/vx_api.c
 VX_TEST = ../../tests/tests.c
 VX_FIO  = ../../fileio/fileio.s
-LIBS    = ./libs/libvecadd.a ./libs/libOpenCL.a ~/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    = ./libs/libvecadd.a /home/fares/dev/pocl/drops_riscv_rt/lib/static/libOpenCL.a ~/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
 
 VX_MAIN = vx_pocl_main
 

runtime/mains/vecadd/vx_pocl_main.c

@@ -55,8 +55,17 @@ static int g_num_kernels = 0;
 static kernel_info_t g_kernels [MAX_KERNELS];
 
 int _pocl_register_kernel(const char* name, const void* pfn, uint32_t num_args, uint32_t num_locals, const uint8_t* arg_types, const uint32_t* local_sizes) {
+  printf("******** _pocl_register_kernel\n");
+  printf("Name to register: %s\n", name);
+  printf("PTR of name: %x\n", name);
   if (g_num_kernels == MAX_KERNELS)
+  {
+  	printf("ERROR: REACHED MAX KERNELS\n");
     return -1;	
+  }
+
+  printf("Going to register at index: %d\n", g_num_kernels);
+
   kernel_info_t* kernel = g_kernels + g_num_kernels++;
   kernel->name = name;
   kernel->pfn = pfn;
@@ -64,14 +73,23 @@ int _pocl_register_kernel(const char* name, const void* pfn, uint32_t num_args,
   kernel->num_locals = num_locals;
   kernel->arg_types = arg_types;
   kernel->local_sizes = local_sizes;
+  printf("New kernel name: %s\n", kernel->name);
   return 0;
 }
 
 int _pocl_query_kernel(const char* name, const void** p_pfn, uint32_t* p_num_args, uint32_t* p_num_locals, const uint8_t** p_arg_types, const uint32_t** p_local_sizes) {
+  printf("********* Inside _pocl_query_kernel\n");
+  printf("name: %s\n", name);
+  printf("g_num_kernels: %d\n", g_num_kernels);
   for (int i = 0; i < g_num_kernels; ++i) {
+  	printf("Currently quering index %d\n", i);
     kernel_info_t* kernel = g_kernels + i;
     if (strcmp(kernel->name, name) != 0)
+    {
+      printf("STR CMP failed! kernel->name = %s \t name: %s\n", kernel->name, name);
       continue;
+    }
+    printf("!!!!!!!!!STR CMP PASSED\n");
     if (p_pfn) *p_pfn = kernel->pfn;
     if (p_num_args) *p_num_args = kernel->num_args;
     if (p_num_locals) *p_num_locals = kernel->num_locals;