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

tensor: Fix commit/metadata logic for HGMMA

Browse Source 
Block HGMMA commit until previous ones are all done; always commit
HGMMA_WAIT after it passes the scoreboard.
This commit is contained in:
Hansung Kim
2024-10-22 20:01:37 -07:00
parent 3abaaff16f
commit 47dff74d3a
1 changed files with 42 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
hw/rtl/core/VX_tensor_hopper_core.sv
View File

@@ -90,9 +90,10 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
logic writeback_ready; logic writeback_ready;
wire metadata_valid = ~metadata_queue_emptys[0/*FIXME*/]; wire metadata_valid = ~metadata_queue_emptys[0/*FIXME*/];
wire not_wait = metadata_valid && (execute_if_data_op_type[0] != `INST_TENSOR_HGMMA_WAIT); wire hmma_wait = metadata_valid &&
(execute_if_data_op_type[0] == `INST_TENSOR_HGMMA_WAIT);
// skip HGMMA_WAIT for kickoff // skip HGMMA_WAIT for kickoff
wire initiate_valid = metadata_valid && not_wait; wire initiate_valid = metadata_valid && !hmma_wait;
// we're recycling execute_if.op_type as operands_if.op_type which might // we're recycling execute_if.op_type as operands_if.op_type which might
// have a different width; let's be safe // have a different width; let's be safe
@@ -156,40 +157,40 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
// ); // );
wire [`NUM_THREADS-1:0][`XLEN-1:0] wb_data = '0; wire [`NUM_THREADS-1:0][`XLEN-1:0] wb_data = '0;
logic commit_select_tensor;
always @(*) begin always @(*) begin
metadata_deq = 1'b0; metadata_deq = 1'b0;
// if there's something in the meta queue, give it priority for commit, // 1'b0: commit from metadata queue
// since every HGMMA instructions are asynchronous and should not // 1'b1: commit from tensor core writeback output
// block commit_select_tensor = 1'b0;
writeback_ready = commit_if.ready;
// if there's something in the meta queue, give it priority for commit
// to keep asynchrony of HGMMA instructions. note HGMMA's should be
// stalled if the tensor core is already busy.
if (metadata_valid) begin if (metadata_valid) begin
// block tensor core writeback if (hmma_wait) begin
writeback_ready = 1'b0; // block tensor core writeback
writeback_ready = 1'b0;
commit_if.valid = metadata_valid; // commit HGMMA_WAIT regardless of tensor core busy
commit_if.data.uuid = execute_if_data_uuid[0]; commit_select_tensor = 1'b0;
commit_if.data.wid = execute_if_data_wid[0]; metadata_deq = metadata_valid && commit_if.ready;
commit_if.data.tmask = execute_if_data_tmask[0]; end else begin
commit_if.data.PC = execute_if_data_PC[0]; // hold commit and meta dequeue until tensor core is ready.
commit_if.data.wb = execute_if_data_wb[0]; // This will stall newer HGMMAs when tensor core is already
commit_if.data.rd = execute_if_data_rd[0]; // busy with an older one.
commit_if.data.data = wb_data; // FIXME ? commit_select_tensor = !initiate_ready;
commit_if.data.tensor = 1'b0; metadata_deq = metadata_valid && commit_if.ready && initiate_ready;
commit_if.data.pid = 1'b0; end
commit_if.data.sop = 1'b1;
commit_if.data.eop = 1'b1;
// block meta queue until tensor core is ready. This will
// effectively stall further issue of async HGMMA when tensor core
// is busy with too many outstanding requests (depth of meta queue).
// be careful to not miss the commit backpressure.
metadata_deq = metadata_valid && commit_if.ready && initiate_ready;
end else begin end else begin
// allow tensor core writeback, provided there's no commit commit_select_tensor = 1'b1;
// backpressure end
writeback_ready = commit_if.ready;
if (commit_select_tensor) begin
commit_if.valid = writeback_valid; commit_if.valid = writeback_valid;
commit_if.data.uuid = '0; commit_if.data.uuid = '0;
commit_if.data.wid = '0; // FIXME commit_if.data.wid = '0; // FIXME
@@ -204,6 +205,19 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
// only the last ghost commit has eop set, which will trigger // only the last ghost commit has eop set, which will trigger
// scoreboard to clear out the busy bit. // scoreboard to clear out the busy bit.
commit_if.data.eop = writeback_last; commit_if.data.eop = writeback_last;
end else begin
commit_if.valid = metadata_valid;
commit_if.data.uuid = execute_if_data_uuid[0];
commit_if.data.wid = execute_if_data_wid[0];
commit_if.data.tmask = execute_if_data_tmask[0];
commit_if.data.PC = execute_if_data_PC[0];
commit_if.data.wb = execute_if_data_wb[0];
commit_if.data.rd = execute_if_data_rd[0];
commit_if.data.data = wb_data; // FIXME ?
commit_if.data.tensor = 1'b0;
commit_if.data.pid = 1'b0;
commit_if.data.sop = 1'b1;
commit_if.data.eop = 1'b1;
end end
end end