wu-arch/vortex
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tensor: Test many-commit per execute with an FSM

Browse Source 
Trick is to set commit_if.data.eop to 0, since the commit module only
signals instruction completion to VX_schedule if the eop bit is 1.
Otherwise it underflows the pending_instr buffer.

The same eop trick works for VX_scoreboard, which works around the
invalid rd writeback error.
This commit is contained in:
Hansung Kim
2024-10-07 21:29:44 -07:00
parent e8ca4677df
commit 7ab14445f0
1 changed files with 49 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
hw/rtl/core/VX_tensor_hopper_core.sv
View File

@@ -11,7 +11,7 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
VX_execute_if.slave execute_if,
VX_commit_if.master commit_if
);
localparam METADATA_QUEUE_DEPTH = 2; // FIXME: arbitrary
localparam METADATA_QUEUE_DEPTH = 16; // FIXME: arbitrary
/* commit_if.data_t parts that we need to keep around:
- uuid
@@ -37,6 +37,7 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
};
wire [`NUM_WARPS-1:0][DATAW-1:0] execute_if_data_deq;
logic [DATAW-1:0] execute_if_data_new_rd;
wire [`NUM_WARPS-1:0] metadata_queue_fulls;
wire [`NUM_WARPS-1:0] metadata_queue_emptys;
@@ -47,6 +48,8 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
`RUNTIME_ASSERT((!execute_if.valid || execute_if.data.wid == `NW_WIDTH'(0)),
("runtime error: WGMMA execute not supported for warps other than 0!"))
logic metadata_deq;
for (genvar i = 0; i < `NUM_WARPS; i++) begin
// Metadata queue for commit_if. This simply copies execute_if's
// metadata and pops them in conjunction with commit fire.
@@ -58,7 +61,7 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
wire enq = operand_enq_fire && (execute_if.data.wid == `NW_WIDTH'(i));
// FIXME: commit only warp 0
wire deq = commit_if_fire && (`NW_WIDTH'(i) == `NW_WIDTH'(0));
wire deq = metadata_deq && commit_if.ready && (`NW_WIDTH'(i) == `NW_WIDTH'(0));
VX_fifo_queue #(
.DATAW(DATAW),
@@ -82,18 +85,58 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
// the commit stage
`RUNTIME_ASSERT(!(!reset && metadata_queue_full), ("tensor core uop queue is full!"))
// FIXME: only checks warp 0 for commit!
assign commit_if.valid = ~metadata_queue_emptys[0/*FIXME*/];
// dummy FSM that generates commits
logic [1:0] state, state_n;
localparam STATE_IDLE = 4'd0;
always @(*) begin
state_n = state;
metadata_deq = 1'b0;
// when incremented to 1, count up until wrap-around to 0
if (state != STATE_IDLE) begin
state_n = state + 1'd1;
end else begin
// kick-off from idle when execute valid
// FIXME: only checks warp 0 for commit!
if (~metadata_queue_emptys[0/*FIXME*/]) begin
state_n = 4'd1;
end
end
// dequeue metadata when wrapping around
if ((state != STATE_IDLE) && (state_n == STATE_IDLE)) begin
metadata_deq = 1'b1;
end
// change rd of the commit data according to state
execute_if_data_new_rd =
{execute_if_data_deq[0/*FIXME*/][DATAW-1:`NR_BITS],
(`NR_BITS'(`NUM_IREGS) + `NR_BITS'(state))};
end
always @(posedge clk) begin
if (reset) begin
state <= '0;
end else begin
state <= state_n;
end
end
// assign commit_if.valid = metadata_deq;
assign commit_if.valid = (state != STATE_IDLE);
wire [`NUM_THREADS-1:0][`XLEN-1:0] wb_data = '0;
localparam COMMIT_DATAW = `UUID_WIDTH + `NW_WIDTH + `NUM_THREADS + `XLEN + 1 + `NR_BITS + (`NUM_THREADS * `XLEN) + 1 + 1 + 1;
wire [COMMIT_DATAW-1:0] commit_if_data = {
execute_if_data_deq[0/*FIXME*/], /* uuid ~ rd */
// write-back to the correct rd only when eop
((state == 2'b11) ? execute_if_data_deq[0/*FIXME*/] : execute_if_data_new_rd), /* uuid ~ rd */
wb_data, /* data */
1'b0, /* pid */
1'b1, /* sop */
1'b1 /* eop */
(state == 2'b11) /* eop */
// 1'b1 /* eop */
};
assign commit_if.data = commit_if_data;