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

Update doc

Browse Source
This commit is contained in:
Hansung Kim
2024-10-21 22:37:20 -07:00
parent 4dcbc31a88
commit 0f06afc3ef
3 changed files with 2 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
hw/rtl/core/VX_commit.sv
View File

@@ -177,7 +177,7 @@ module VX_commit import VX_gpu_pkg::*; #(
// probably want to change this at some point // probably want to change this at some point
// (i.e. pass a "don't count this towards pending instructions" signal down the pipeline) // (i.e. pass a "don't count this towards pending instructions" signal down the pipeline)
wire [`ISSUE_WIDTH-1:0] final_hmma; wire [`ISSUE_WIDTH-1:0] final_hmma;
// if this is a "ghost" commit generated from the tensor core, don't count // if this is a "ghost" commit generated at the tensor core, don't count
// toward committed // toward committed
wire [`ISSUE_WIDTH-1:0] tensor_ghost; wire [`ISSUE_WIDTH-1:0] tensor_ghost;
`ifdef EXT_T_ENABLE `ifdef EXT_T_ENABLE

3
hw/rtl/core/VX_tensor_hopper_core.sv
View File

@@ -158,9 +158,6 @@ module VX_tensor_hopper_core_block import VX_gpu_pkg::*; #(
// mark as "ghost" commit. This will prevent this commit from // mark as "ghost" commit. This will prevent this commit from
// decrementing from pending_instr buffer // decrementing from pending_instr buffer
commit_if.data.tensor = 1'b1; commit_if.data.tensor = 1'b1;
// eop is deliberately set so that we don't underflow the pending_instr
// buffer in VX_schedule. An instruction is considered committed only
// when the eop bit is set to one (see VX_commit).
// 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;

2
hw/rtl/fpu/VX_tensor_dpu.sv
View File

@@ -290,7 +290,7 @@ module VX_tensor_threadgroup #(
// d_col_sel: 1, 3, 1, 3 // d_col_sel: 1, 3, 1, 3
// //
// substep 0: // substep 0:
// [ 0 x 2 x ] // [ 0 x 2 x ] (0~3 is 'i', the dpu id)
// [ 1 x 3 x ] // [ 1 x 3 x ]
// substep 1: // substep 1:
// [ x 0 x 2 ] // [ x 0 x 2 ]