Rename sched_barrier_stalls -> perf_sched_barrier_idles

Sched stall by barrier is really idle because it causes !scheduler_if.valid,
which is counted as part of sched_idle.
This commit is contained in:
Hansung Kim
2024-03-28 22:45:12 -07:00
parent dd90736382
commit 50263a5f7d
3 changed files with 21 additions and 18 deletions

View File

@@ -415,13 +415,13 @@ module VX_core import VX_gpu_pkg::*; #(
$itor(instrs) / $itor(cycles)); $itor(instrs) / $itor(cycles));
$display("scheduler idle: %d cycles (%.2f%%)", pipeline_perf_if.sched_idles, $display("scheduler idle: %d cycles (%.2f%%)", pipeline_perf_if.sched_idles,
$itor(pipeline_perf_if.sched_idles) / $itor(cycles) * 100.0); $itor(pipeline_perf_if.sched_idles) / $itor(cycles) * 100.0);
$display("scheduler barrier idle: %d count across NUM_WARPS=%d",
pipeline_perf_if.sched_barrier_idles, `NUM_WARPS);
// sched_stalls can happen when the later issue stage stalls,
// causing the ibuffer to clog.
$display("scheduler stalls: %d cycles (%.2f%%)", pipeline_perf_if.sched_stalls, $display("scheduler stalls: %d cycles (%.2f%%)", pipeline_perf_if.sched_stalls,
$itor(pipeline_perf_if.sched_stalls) / $itor(cycles) * 100.0); $itor(pipeline_perf_if.sched_stalls) / $itor(cycles) * 100.0);
$display("scheduler barrier stalls: %d count across NUM_WARPS=%d (%.2f%%)", $display("decode stalls (ibuffer not ready): %d cycles (%.2f%%)",pipeline_perf_if.ibf_stalls,
pipeline_perf_if.sched_barrier_stalls,
`NUM_WARPS,
$itor(pipeline_perf_if.sched_barrier_stalls) / $itor(cycles) * 100.0);
$display("decode stalls: %d cycles (%.2f%%)",pipeline_perf_if.ibf_stalls,
$itor(pipeline_perf_if.ibf_stalls) / $itor(cycles) * 100.0); $itor(pipeline_perf_if.ibf_stalls) / $itor(cycles) * 100.0);
// see VX_scoreboard.sv // see VX_scoreboard.sv
// scb_stalls: valid & ~ready (ready = stg_ready_in && operands_ready) // scb_stalls: valid & ~ready (ready = stg_ready_in && operands_ready)
@@ -472,12 +472,12 @@ module VX_core import VX_gpu_pkg::*; #(
pipeline_perf_if.dispatch_any_fire_cycles, pipeline_perf_if.dispatch_any_fire_cycles,
$itor(pipeline_perf_if.dispatch_any_fire_cycles) / $itor(cycles) * 100.0); $itor(pipeline_perf_if.dispatch_any_fire_cycles) / $itor(cycles) * 100.0);
$display("ifetches: %d", perf_ifetches); $display("ifetches: %d", perf_ifetches);
$display("ifetch latency: %f Cycles", $display("ifetch latency: %f cycles",
$itor(icache_lat) / $itor(ifetches)); $itor(icache_lat) / $itor(ifetches));
$display("loads: %d", perf_loads); $display("dcache loads: %d", perf_loads);
$display("load latency: %f Cycles", $display("dcache load latency: %f cycles",
$itor(dcache_lat) / $itor(loads)); $itor(dcache_lat) / $itor(loads));
$display("stores: %d", perf_stores); $display("dcache stores: %d", perf_stores);
end end
end end

View File

@@ -413,28 +413,28 @@ module VX_schedule import VX_gpu_pkg::*; #(
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
reg [`PERF_CTR_BITS-1:0] perf_sched_idles; reg [`PERF_CTR_BITS-1:0] perf_sched_idles;
reg [`PERF_CTR_BITS-1:0] perf_sched_stalls; reg [`PERF_CTR_BITS-1:0] perf_sched_stalls;
reg [`PERF_CTR_BITS-1:0] perf_sched_barrier_stalls; reg [`PERF_CTR_BITS-1:0] perf_sched_barrier_idles;
wire schedule_idle = ~schedule_valid; wire schedule_idle = ~schedule_valid;
wire schedule_stall = schedule_if.valid && ~schedule_if.ready; wire schedule_stall = schedule_if.valid && ~schedule_if.ready;
wire [`CLOG2(`NUM_WARPS+1)-1:0] schedule_barrier_stall; wire [`CLOG2(`NUM_WARPS+1)-1:0] schedule_barrier_idle;
`POP_COUNT(schedule_barrier_stall, barrier_stalls); `POP_COUNT(schedule_barrier_idle, barrier_stalls);
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
perf_sched_idles <= '0; perf_sched_idles <= '0;
perf_sched_barrier_idles <= '0;
perf_sched_stalls <= '0; perf_sched_stalls <= '0;
perf_sched_barrier_stalls <= '0;
end else begin end else begin
perf_sched_idles <= perf_sched_idles + `PERF_CTR_BITS'(schedule_idle); perf_sched_idles <= perf_sched_idles + `PERF_CTR_BITS'(schedule_idle);
perf_sched_barrier_idles <= perf_sched_barrier_idles + `PERF_CTR_BITS'(schedule_barrier_idle);
perf_sched_stalls <= perf_sched_stalls + `PERF_CTR_BITS'(schedule_stall); perf_sched_stalls <= perf_sched_stalls + `PERF_CTR_BITS'(schedule_stall);
perf_sched_barrier_stalls <= perf_sched_barrier_stalls + `PERF_CTR_BITS'(schedule_barrier_stall);
end end
end end
assign perf_schedule_if.sched_idles = perf_sched_idles; assign perf_schedule_if.sched_idles = perf_sched_idles;
assign perf_schedule_if.sched_barrier_idles = perf_sched_barrier_idles;
assign perf_schedule_if.sched_stalls = perf_sched_stalls; assign perf_schedule_if.sched_stalls = perf_sched_stalls;
assign perf_schedule_if.sched_barrier_stalls = perf_sched_barrier_stalls;
`endif `endif
endmodule endmodule

View File

@@ -16,9 +16,10 @@
interface VX_pipeline_perf_if (); interface VX_pipeline_perf_if ();
wire [`PERF_CTR_BITS-1:0] sched_idles; wire [`PERF_CTR_BITS-1:0] sched_idles;
wire [`PERF_CTR_BITS-1:0] sched_stalls; wire [`PERF_CTR_BITS-1:0] sched_stalls;
wire [`PERF_CTR_BITS-1:0] sched_barrier_stalls; wire [`PERF_CTR_BITS-1:0] sched_barrier_idles;
wire [`PERF_CTR_BITS-1:0] ibf_stalls; wire [`PERF_CTR_BITS-1:0] ibf_stalls;
wire [`PERF_CTR_BITS-1:0] scb_stalls; wire [`PERF_CTR_BITS-1:0] scb_stalls;
wire [`PERF_CTR_BITS-1:0] scb_fires;
wire [`PERF_CTR_BITS-1:0] units_uses [`NUM_EX_UNITS]; wire [`PERF_CTR_BITS-1:0] units_uses [`NUM_EX_UNITS];
wire [`PERF_CTR_BITS-1:0] sfu_uses [`NUM_SFU_UNITS]; wire [`PERF_CTR_BITS-1:0] sfu_uses [`NUM_SFU_UNITS];
wire [`PERF_CTR_BITS-1:0] dispatch_stalls [`NUM_EX_UNITS]; wire [`PERF_CTR_BITS-1:0] dispatch_stalls [`NUM_EX_UNITS];
@@ -34,13 +35,14 @@ interface VX_pipeline_perf_if ();
modport schedule ( modport schedule (
output sched_idles, output sched_idles,
output sched_barrier_stalls, output sched_barrier_idles,
output sched_stalls output sched_stalls
); );
modport issue ( modport issue (
output ibf_stalls, output ibf_stalls,
output scb_stalls, output scb_stalls,
output scb_fires,
output units_uses, output units_uses,
output sfu_uses, output sfu_uses,
output dispatch_stalls, output dispatch_stalls,
@@ -51,10 +53,11 @@ interface VX_pipeline_perf_if ();
modport slave ( modport slave (
input sched_idles, input sched_idles,
input sched_barrier_stalls, input sched_barrier_idles,
input sched_stalls, input sched_stalls,
input ibf_stalls, input ibf_stalls,
input scb_stalls, input scb_stalls,
input scb_fires,
input units_uses, input units_uses,
input sfu_uses, input sfu_uses,
input dispatch_stalls, input dispatch_stalls,