software updaet for new thread mask design

hw/rtl/VX_gpu_unit.v

@@ -48,20 +48,20 @@ module VX_gpu_unit #(
 
     // split
 
-    wire [`NUM_THREADS-1:0] split_then_mask;
-    wire [`NUM_THREADS-1:0] split_else_mask;
+    wire [`NUM_THREADS-1:0] split_then_tmask;
+    wire [`NUM_THREADS-1:0] split_else_tmask;
 
     for (genvar i = 0; i < `NUM_THREADS; i++) begin
-        wire taken = gpu_req_if.rs1_data[i][gpu_req_if.tid];
-        assign split_then_mask[i] = gpu_req_if.tmask[i] & taken;
-        assign split_else_mask[i] = gpu_req_if.tmask[i] & ~taken;
+        wire taken = gpu_req_if.rs1_data[i][0];
+        assign split_then_tmask[i] = gpu_req_if.tmask[i] & taken;
+        assign split_else_tmask[i] = gpu_req_if.tmask[i] & ~taken;
     end
 
-    assign split.valid     = is_split;
-    assign split.diverged  = (| split_then_mask) && (| split_else_mask);
-    assign split.then_mask = split_then_mask;
-    assign split.else_mask = split_else_mask;
-    assign split.pc        = gpu_req_if.next_PC;
+    assign split.valid      = is_split;
+    assign split.diverged   = (| split_then_tmask) && (| split_else_tmask);
+    assign split.then_tmask = split_then_tmask;
+    assign split.else_tmask = split_else_tmask;
+    assign split.pc         = gpu_req_if.next_PC;
 
     // barrier
     

hw/rtl/VX_types.vh

@@ -41,8 +41,8 @@ typedef struct packed {
 typedef struct packed {
     logic                   valid;
     logic                   diverged;
-    logic [`NUM_THREADS-1:0] then_mask;
-    logic [`NUM_THREADS-1:0] else_mask;
+    logic [`NUM_THREADS-1:0] then_tmask;
+    logic [`NUM_THREADS-1:0] else_tmask;
     logic [31:0]            pc;
 } gpu_split_t;
 

hw/rtl/VX_warp_sched.v

@@ -47,7 +47,9 @@ module VX_warp_sched #(
     reg [`NW_BITS-1:0]      scheduled_warp;
     wire                    warp_scheduled;
 
-    wire ifetch_rsp_fire = ifetch_rsp_if.valid && ifetch_rsp_if.ready;    
+    wire ifetch_rsp_fire = ifetch_rsp_if.valid && ifetch_rsp_if.ready; 
+
+    wire tmc_active = (warp_ctl_if.tmc.tmask != 0);   
 
     always @(*) begin
         active_warps_n = active_warps;
@@ -55,14 +57,14 @@ module VX_warp_sched #(
             active_warps_n = warp_ctl_if.wspawn.wmask;
         end
         if (warp_ctl_if.valid && warp_ctl_if.tmc.valid) begin
-            active_warps_n[warp_ctl_if.wid] = (warp_ctl_if.tmc.tmask != 0);
+            active_warps_n[warp_ctl_if.wid] = tmc_active;
         end        
     end
 
     always @(*) begin
         schedule_table_n = schedule_table;
         if (warp_ctl_if.valid && warp_ctl_if.tmc.valid) begin
-            schedule_table_n[warp_ctl_if.wid] = (warp_ctl_if.tmc.tmask != 0);
+            schedule_table_n[warp_ctl_if.wid] = tmc_active;
         end        
         if (warp_scheduled) begin // remove scheduled warp (round-robin)
             schedule_table_n[scheduled_warp] = 0;
@@ -104,12 +106,12 @@ module VX_warp_sched #(
                     barrier_stall_mask[warp_ctl_if.barrier.id][warp_ctl_if.wid] <= 1;
                 end
             end else if (warp_ctl_if.valid && warp_ctl_if.tmc.valid) begin
-                thread_masks[warp_ctl_if.wid] <= warp_ctl_if.tmc.tmask;
+                thread_masks[warp_ctl_if.wid]  <= warp_ctl_if.tmc.tmask;
                 stalled_warps[warp_ctl_if.wid] <= 0;
             end else if (warp_ctl_if.valid && warp_ctl_if.split.valid) begin
                 stalled_warps[warp_ctl_if.wid] <= 0;
                 if (warp_ctl_if.split.diverged) begin
-                    thread_masks[warp_ctl_if.wid] <= warp_ctl_if.split.then_mask;
+                    thread_masks[warp_ctl_if.wid] <= warp_ctl_if.split.then_tmask;
                 end
             end          
 
@@ -178,6 +180,8 @@ module VX_warp_sched #(
     // split/join stack management    
 
     wire [(1+32+`NUM_THREADS-1):0] ipdom [`NUM_WARPS-1:0];
+
+    wire [`NUM_THREADS-1:0] curr_tmask = thread_masks[warp_ctl_if.wid];
     
     for (genvar i = 0; i < `NUM_WARPS; i++) begin
         wire push = warp_ctl_if.valid 
@@ -186,9 +190,9 @@ module VX_warp_sched #(
 
         wire pop = join_if.valid && (i == join_if.wid);
 
-        wire [`NUM_THREADS-1:0] else_mask = warp_ctl_if.split.diverged ? warp_ctl_if.split.else_mask : thread_masks[warp_ctl_if.wid];
-        wire [(1+32+`NUM_THREADS-1):0] q_end  = {1'b0, 32'b0,                thread_masks[warp_ctl_if.wid]};
-        wire [(1+32+`NUM_THREADS-1):0] q_else = {1'b1, warp_ctl_if.split.pc, else_mask};
+        wire [`NUM_THREADS-1:0] else_tmask = warp_ctl_if.split.diverged ? warp_ctl_if.split.else_tmask : curr_tmask;
+        wire [(1+32+`NUM_THREADS-1):0] q_end  = {1'b0, 32'b0,                curr_tmask};
+        wire [(1+32+`NUM_THREADS-1):0] q_else = {1'b1, warp_ctl_if.split.pc, else_tmask};
 
         VX_ipdom_stack #(
             .WIDTH (1+32+`NUM_THREADS), 

runtime/src/vx_spawn.c

@@ -43,7 +43,8 @@ inline int fast_log2(int x) {
 }
 
 static void spawn_tasks_callback() {  
-  vx_tmc(vx_num_threads());
+  // activate all threads
+  vx_tmc(-1);
 
   int core_id = vx_core_id();
   int wid     = vx_warp_id();
@@ -60,11 +61,13 @@ static void spawn_tasks_callback() {
     (p_wspawn_args->callback)(task_id, p_wspawn_args->arg);
   }
 
+  // set warp0 to single-threaded and stop other warps
   vx_tmc(0 == wid);
 }
 
-void spawn_remaining_tasks_callback(int nthreads) {    
-  vx_tmc(nthreads);
+void spawn_remaining_tasks_callback(int thread_mask) {  
+  // activate threads  
+  vx_tmc(thread_mask);
 
   int core_id = vx_core_id(); 
   int tid = vx_thread_gid();
@@ -74,6 +77,7 @@ void spawn_remaining_tasks_callback(int nthreads) {
   int task_id = p_wspawn_args->offset + tid;
   (p_wspawn_args->callback)(task_id, p_wspawn_args->arg);
 
+  // back to single-threaded
   vx_tmc(1);
 }
 
@@ -132,7 +136,8 @@ void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
 ///////////////////////////////////////////////////////////////////////////////
 
 static void spawn_kernel_callback() {  
-  vx_tmc(vx_num_threads());
+  // activate all threads
+  vx_tmc(-1);
 
   int core_id = vx_core_id();
   int wid     = vx_warp_id();
@@ -162,11 +167,13 @@ static void spawn_kernel_callback() {
     (p_wspawn_args->callback)(p_wspawn_args->arg, p_wspawn_args->ctx, gid0, gid1, gid2);
   }
 
+  // set warp0 to single-threaded and stop other warps
   vx_tmc(0 == wid);
 }
 
-static void spawn_kernel_remaining_callback(int nthreads) {    
-  vx_tmc(nthreads);
+static void spawn_kernel_remaining_callback(int thread_mask) {    
+  // activate threads
+  vx_tmc(thread_mask);
 
   int core_id = vx_core_id(); 
   int tid = vx_thread_gid();
@@ -190,6 +197,7 @@ static void spawn_kernel_remaining_callback(int nthreads) {
 
   (p_wspawn_args->callback)(p_wspawn_args->arg, p_wspawn_args->ctx, gid0, gid1, gid2);
 
+  // back to single-threaded
   vx_tmc(1);
 }
 

runtime/src/vx_start.S

@@ -59,7 +59,7 @@ label_exit_next:
 .global vx_set_sp
 vx_set_sp:
   # activate all threads
-  csrr a0, CSR_NT             # get num threads
+  li a0, -1
   .insn s 0x6b, 0, x0, 0(a0)  # tmc a0
   
   # set global pointer register

simX/core.cpp

@@ -267,6 +267,9 @@ Word Core::get_csr(Addr addr, int tid, int wid) {
   } else if (addr == CSR_GCID) {
     // Processor coreID
     return id_;
+  } else if (addr == CSR_TMASK) {
+    // Processor coreID
+    return warps_.at(wid)->getTmask();
   } else if (addr == CSR_NT) {
     // Number of threads per warp
     return arch_.num_threads();

simX/execute.cpp

@@ -817,10 +817,9 @@ void Warp::execute(const Instr &instr, Pipeline *pipeline) {
       switch (func3) {
       case 0: {
         // TMC
-        int active_threads = std::min<int>(rsdata[0], num_threads);          
         tmask_.reset();
-        for (int i = 0; i < active_threads; ++i) {
-          tmask_[i] = true;
+        for (size_t i = 0; i < tmask_.size(); ++i) {
+          tmask_[i] = rsdata[0] & (1 << i);
         }
         active_ = tmask_.any();
         pipeline->stall_warp = true;

simX/warp.h

@@ -74,6 +74,12 @@ public:
     active_ = tmask_.any();
   }
 
+  Word getTmask() const {
+    if (active_)
+      return tmask_.to_ulong();
+    return 0;
+  }
+
   Word getIRegValue(int reg) const {
     return iRegFile_[0][reg];
   }