profiling optimizations

minor updates
This commit is contained in:
Blaise Tine 2023-12-05 05:12:13 -08:00
parent e5b41bcd66
commit 6c7ac35054
5 changed files with 65 additions and 42 deletions

View file

@ -436,7 +436,7 @@
// Number of Banks
`ifndef DCACHE_NUM_BANKS
`define DCACHE_NUM_BANKS (`NUM_LSU_LANES)
`define DCACHE_NUM_BANKS `MIN(`NUM_LSU_LANES, 4)
`endif
// Core Response Queue Size

View file

@ -174,30 +174,38 @@ module VX_dispatch import VX_gpu_pkg::*; #(
|| (sfu_operands_if[i].ready && (operands_if[i].data.ex_type == `EX_SFU));
end
`ifdef PERF_ENABLE
reg [`NUM_EX_UNITS-1:0][`PERF_CTR_BITS-1:0] perf_stalls_n, perf_stalls_r;
wire [`ISSUE_WIDTH-1:0] operands_stall;
wire [`ISSUE_WIDTH-1:0][`EX_BITS-1:0] operands_ex_type;
`ifdef PERF_ENABLE
wire [`NUM_EX_UNITS-1:0] perf_unit_stalls_per_cycle, perf_unit_stalls_per_cycle_r;
reg [`ISSUE_WIDTH-1:0][`NUM_EX_UNITS-1:0] perf_issue_unit_stalls_per_cycle;
reg [`NUM_EX_UNITS-1:0][`PERF_CTR_BITS-1:0] perf_stalls_r;
for (genvar i=0; i < `ISSUE_WIDTH; ++i) begin
assign operands_stall[i] = operands_if[i].valid && ~operands_if[i].ready;
assign operands_ex_type[i] = operands_if[i].data.ex_type;
end
always @(*) begin
perf_stalls_n = perf_stalls_r;
for (integer i=0; i < `ISSUE_WIDTH; ++i) begin
if (operands_stall[i]) begin
perf_stalls_n[operands_ex_type[i]] += `PERF_CTR_BITS'(1);
always @(*) begin
perf_issue_unit_stalls_per_cycle[i] = '0;
if (operands_if[i].valid && ~operands_if[i].ready) begin
perf_issue_unit_stalls_per_cycle[i][operands_if[i].data.ex_type] = 1;
end
end
end
always @(posedge clk) begin
if (reset) begin
perf_stalls_r <= '0;
end else begin
perf_stalls_r <= perf_stalls_n;
VX_reduce #(
.DATAW_IN (`NUM_EX_UNITS),
.N (`ISSUE_WIDTH),
.OP ("|")
) reduce (
.data_in (perf_issue_unit_stalls_per_cycle),
.data_out (perf_unit_stalls_per_cycle)
);
`BUFFER(perf_unit_stalls_per_cycle_r, perf_unit_stalls_per_cycle);
for (genvar i = 0; i < `NUM_EX_UNITS; ++i) begin
always @(posedge clk) begin
if (reset) begin
perf_stalls_r[i] <= '0;
end else begin
perf_stalls_r[i] <= perf_stalls_r[i] + `PERF_CTR_BITS'(perf_unit_stalls_per_cycle_r[i]);
end
end
end

View file

@ -32,19 +32,20 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
localparam DATAW = `UUID_WIDTH + ISSUE_WIS_W + `NUM_THREADS + `XLEN + `EX_BITS + `INST_OP_BITS + `INST_MOD_BITS + 1 + 1 + `XLEN + (`NR_BITS * 4) + 1;
`ifdef PERF_ENABLE
wire [`NUM_EX_UNITS-1:0] scoreboard_uses_per_cycle;
wire [`CLOG2(`ISSUE_WIDTH+1)-1:0] scoreboard_stalls_per_cycle;
reg [`ISSUE_WIDTH-1:0][`NUM_EX_UNITS-1:0] scoreboard_uses;
wire [`ISSUE_WIDTH-1:0] scoreboard_stalls;
wire [`NUM_EX_UNITS-1:0] perf_uses_per_cycle;
wire [`CLOG2(`ISSUE_WIDTH+1)-1:0] perf_stalls_per_cycle;
reg [`ISSUE_WIDTH-1:0][`NUM_EX_UNITS-1:0] perf_issue_uses_per_cycle;
wire [`ISSUE_WIDTH-1:0] perf_issue_stalls_per_cycle;
`POP_COUNT(perf_stalls_per_cycle, perf_issue_stalls_per_cycle);
`POP_COUNT(scoreboard_stalls_per_cycle, scoreboard_stalls);
VX_reduce #(
.DATAW_IN (`NUM_EX_UNITS),
.N (`ISSUE_WIDTH),
.OP ("|")
) reduce (
.data_in (scoreboard_uses),
.data_out (scoreboard_uses_per_cycle)
.data_in (perf_issue_uses_per_cycle),
.data_out (perf_uses_per_cycle)
);
`endif
@ -62,23 +63,23 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
`ifdef PERF_ENABLE
reg [`UP(ISSUE_RATIO)-1:0][`NUM_REGS-1:0][`EX_BITS-1:0] inuse_units;
always @(*) begin
scoreboard_uses[i] = '0;
perf_issue_uses_per_cycle[i] = '0;
if (ibuffer_if[i].valid) begin
if (inuse_rd) begin
scoreboard_uses[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rd]] = 1;
perf_issue_uses_per_cycle[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rd]] = 1;
end
if (inuse_rs1) begin
scoreboard_uses[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs1]] = 1;
perf_issue_uses_per_cycle[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs1]] = 1;
end
if (inuse_rs2) begin
scoreboard_uses[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs2]] = 1;
perf_issue_uses_per_cycle[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs2]] = 1;
end
if (inuse_rs3) begin
scoreboard_uses[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs3]] = 1;
perf_issue_uses_per_cycle[i][inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rs3]] = 1;
end
end
end
assign scoreboard_stalls[i] = ibuffer_if[i].valid && ~ibuffer_if[i].ready;
assign perf_issue_stalls_per_cycle[i] = ibuffer_if[i].valid && ~ibuffer_if[i].ready;
`endif
reg [DATAW-1:0] data_out_r;
@ -164,19 +165,26 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
end
`ifdef PERF_ENABLE
wire [`CLOG2(`ISSUE_WIDTH+1)-1:0] perf_stalls_per_cycle_r;
wire [`NUM_EX_UNITS-1:0] perf_uses_per_cycle_r;
`BUFFER(perf_stalls_per_cycle_r, perf_stalls_per_cycle);
`BUFFER(perf_uses_per_cycle_r, perf_uses_per_cycle);
always @(posedge clk) begin
if (reset) begin
perf_scb_stalls <= '0;
perf_scb_stalls <= '0;
end else begin
perf_scb_stalls <= perf_scb_stalls + `PERF_CTR_BITS'(scoreboard_stalls_per_cycle);
perf_scb_stalls <= perf_scb_stalls + `PERF_CTR_BITS'(perf_stalls_per_cycle_r);
end
end
for (genvar i = 0; i < `NUM_EX_UNITS; ++i) begin
always @(posedge clk) begin
if (reset) begin
perf_scb_uses[i] <= '0;
end else begin
perf_scb_uses[i] <= perf_scb_uses[i] + `PERF_CTR_BITS'(scoreboard_uses_per_cycle[i]);
perf_scb_uses[i] <= perf_scb_uses[i] + `PERF_CTR_BITS'(perf_uses_per_cycle_r[i]);
end
end
end

View file

@ -183,21 +183,23 @@ module VX_stream_xbar #(
per_cycle_collision = 0;
for (integer i = 0; i < NUM_INPUTS; ++i) begin
for (integer j = 1; j < (NUM_INPUTS-i); ++j) begin
if (valid_in[i] && valid_in[j+i] && sel_in[i] == sel_in[j+i]) begin
per_cycle_collision[i] |= ready_in[i] | ready_in[j+i];
end
per_cycle_collision[i] |= valid_in[i]
&& valid_in[j+i]
&& (sel_in[i] == sel_in[j+i])
&& (ready_in[i] | ready_in[j+i]);
end
end
end
wire [`CLOG2(NUM_INPUTS+1)-1:0] collision_count;
wire [`CLOG2(NUM_INPUTS+1)-1:0] collision_count, collision_count_r;
`POP_COUNT(collision_count, per_cycle_collision);
`BUFFER(collision_count_r, collision_count);
always @(posedge clk) begin
if (reset) begin
collisions_r <= '0;
end else begin
collisions_r <= collisions_r + PERF_CTR_BITS'(collision_count);
collisions_r <= collisions_r + PERF_CTR_BITS'(collision_count_r);
end
end

View file

@ -245,14 +245,19 @@ module VX_shared_mem import VX_gpu_pkg::*; #(
reg [`PERF_CTR_BITS-1:0] perf_writes;
reg [`PERF_CTR_BITS-1:0] perf_crsp_stalls;
wire [`CLOG2(NUM_REQS+1)-1:0] perf_reads_per_cycle_r;
wire [`CLOG2(NUM_REQS+1)-1:0] perf_writes_per_cycle_r;
`BUFFER(perf_reads_per_cycle_r, perf_reads_per_cycle);
`BUFFER(perf_writes_per_cycle_r, perf_writes_per_cycle);
always @(posedge clk) begin
if (reset) begin
perf_reads <= '0;
perf_writes <= '0;
perf_crsp_stalls <= '0;
end else begin
perf_reads <= perf_reads + `PERF_CTR_BITS'(perf_reads_per_cycle);
perf_writes <= perf_writes + `PERF_CTR_BITS'(perf_writes_per_cycle);
perf_reads <= perf_reads + `PERF_CTR_BITS'(perf_reads_per_cycle_r);
perf_writes <= perf_writes + `PERF_CTR_BITS'(perf_writes_per_cycle_r);
perf_crsp_stalls <= perf_crsp_stalls + `PERF_CTR_BITS'(perf_crsp_stall_per_cycle);
end
end