timing optimizations

vortexgpgpu · Jan 11, 2025 · 083cf04 · 083cf04
1 parent 84b1c8a
commit 083cf04
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 22 deletions.
diff --git a/hw/rtl/core/VX_schedule.sv b/hw/rtl/core/VX_schedule.sv
@@ -68,8 +68,6 @@ module VX_schedule import VX_gpu_pkg::*; #(
 
     reg [`PERF_CTR_BITS-1:0] cycles;
 
-    reg [`NUM_WARPS-1:0][`UUID_WIDTH-1:0] issued_instrs;
-
     wire schedule_fire = schedule_valid && schedule_ready;
     wire schedule_if_fire = schedule_if.valid && schedule_if.ready;
 
@@ -113,6 +111,16 @@ module VX_schedule import VX_gpu_pkg::*; #(
         barrier_stalls_n= barrier_stalls;
         warp_pcs_n      = warp_pcs;
 
+        // decode unlock
+        if (decode_sched_if.valid && decode_sched_if.unlock) begin
+            stalled_warps_n[decode_sched_if.wid] = 0;
+        end
+
+        // CSR unlock
+        if (sched_csr_if.unlock_warp) begin
+            stalled_warps_n[sched_csr_if.unlock_wid] = 0;
+        end
+
         // wspawn handling
         if (wspawn.valid && is_single_warp) begin
             active_warps_n |= wspawn.wmask;
@@ -170,6 +178,7 @@ module VX_schedule import VX_gpu_pkg::*; #(
                 stalled_warps_n[warp_ctl_if.wid] = 0; // unlock warp
             end
         end
+
     `ifdef GBAR_ENABLE
         if (gbar_bus_if.rsp_valid && (gbar_req_id == gbar_bus_if.rsp_data.id)) begin
             barrier_ctrs_n[warp_ctl_if.barrier.id] = '0; // reset barrier counter
@@ -188,16 +197,6 @@ module VX_schedule import VX_gpu_pkg::*; #(
             end
         end
 
-        // decode unlock
-        if (decode_sched_if.valid && decode_sched_if.unlock) begin
-            stalled_warps_n[decode_sched_if.wid] = 0;
-        end
-
-        // CSR unlock
-        if (sched_csr_if.unlock_warp) begin
-            stalled_warps_n[sched_csr_if.unlock_wid] = 0;
-        end
-
         // stall the warp until decode stage
         if (schedule_fire) begin
             stalled_warps_n[schedule_wid] = 1;
@@ -223,7 +222,6 @@ module VX_schedule import VX_gpu_pkg::*; #(
             active_warps    <= '0;
             thread_masks    <= '0;
             barrier_stalls  <= '0;
-            issued_instrs   <= '0;
             cycles          <= '0;
             wspawn.valid    <=  0;
 
@@ -268,10 +266,6 @@ module VX_schedule import VX_gpu_pkg::*; #(
             end
         `endif
 
-            if (schedule_if_fire) begin
-                issued_instrs[schedule_if.data.wid] <= issued_instrs[schedule_if.data.wid] + `UUID_WIDTH'(1);
-            end
-
             if (busy) begin
                 cycles <= cycles + 1;
             end

diff --git a/hw/rtl/core/VX_scoreboard.sv b/hw/rtl/core/VX_scoreboard.sv
@@ -151,11 +151,14 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
         end
     `endif
 
-        always @(*) begin
-            for (integer i = 0; i < NUM_OPDS; ++i) begin
+        for (genvar i = 0; i < NUM_OPDS; ++i) begin : g_operands_busy_n
+            always @(*) begin
                 operands_busy_n[i] = operands_busy[i];
                 if (ibuffer_fire) begin
                     operands_busy_n[i] = inuse_regs[ibuf_opds[i]];
+                    if (staging_fire && staging_if[w].data.wb && staging_if[w].data.rd == ibuf_opds[i]) begin
+                        operands_busy_n[i] = 1;
+                    end
                 end
                 if (writeback_fire) begin
                     if (ibuffer_fire) begin
@@ -168,9 +171,6 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
                        end
                     end
                 end
-                if (staging_fire && staging_if[w].data.wb && staging_if[w].data.rd == ibuf_opds[i]) begin
-                    operands_busy_n[i] = 1;
-                end
             end
         end
 
@@ -185,8 +185,10 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
                     inuse_regs[staging_if[w].data.rd] <= 1;
                 end
             end
+
             operands_busy <= operands_busy_n;
             operands_ready[w] <= ~(| operands_busy_n);
+
         `ifdef PERF_ENABLE
             if (staging_fire && staging_if[w].data.wb) begin
                 inuse_units[staging_if[w].data.rd] <= staging_if[w].data.ex_type;