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

tensor: Document configuring queue depths

Browse Source
This commit is contained in:
Hansung Kim
2024-05-30 18:32:27 -07:00
parent 83f9f6d84f
commit 574cc0e5f0
2 changed files with 9 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
hw/rtl/core/VX_tensor_core.sv
View File

@@ -318,8 +318,6 @@ module VX_tensor_octet #(
output result_valid, output result_valid,
input result_ready input result_ready
); );
localparam ISSUE_QUEUE_DEPTH = 4;
// 512 bits/octet * 4 octets per warp // 512 bits/octet * 4 octets per warp
logic [`NUM_WARPS-1:0][3:0][31:0] A_buffer, A_buffer_n; logic [`NUM_WARPS-1:0][3:0][31:0] A_buffer, A_buffer_n;
logic [`NUM_WARPS-1:0][3:0][31:0] B_buffer, B_buffer_n; logic [`NUM_WARPS-1:0][3:0][31:0] B_buffer, B_buffer_n;
@@ -471,7 +469,7 @@ module VX_tensor_octet #(
VX_tensor_dpu #( VX_tensor_dpu #(
.ISW(ISW), .ISW(ISW),
.OCTET(OCTET), .OCTET(OCTET),
.ISSUE_QUEUE_DEPTH(4) .ISSUE_QUEUE_DEPTH(4 /*@perf: arbtirary*/)
) dpu ( ) dpu (
.clk(clk), .clk(clk),
.reset(reset), .reset(reset),
@@ -503,10 +501,9 @@ module VX_tensor_octet #(
// regular, every-2-cycle commit traffic to ensure the commit pipeline is // regular, every-2-cycle commit traffic to ensure the commit pipeline is
// used more efficiently. // used more efficiently.
// FIXME: unnecessary? // FIXME: unnecessary?
// TODO: This is probably oversized.
VX_fifo_queue #( VX_fifo_queue #(
.DATAW ($bits(D_wid) + $bits(D_out)), .DATAW ($bits(D_wid) + $bits(D_out)),
.DEPTH (2 /*`LATENCY_HMMA*/) .DEPTH (2 /* arbitrary */)
) output_buffer ( ) output_buffer (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),

9
hw/rtl/fpu/VX_tensor_dpu.sv
View File

@@ -4,6 +4,9 @@
module VX_tensor_dpu #( module VX_tensor_dpu #(
parameter ISW, parameter ISW,
parameter OCTET, parameter OCTET,
// @perf: has big impact on throughput. A rule of thumb is to set it to
// the pipeline length of FEDPs in order to make sure there are enough
// entries to fully saturate the pipeline, but this is still rough
parameter ISSUE_QUEUE_DEPTH = `LATENCY_HMMA parameter ISSUE_QUEUE_DEPTH = `LATENCY_HMMA
) ( ) (
input clk, input clk,
@@ -105,7 +108,9 @@ module VX_tensor_dpu #(
// need to pass along warp id's to do multithreading // need to pass along warp id's to do multithreading
VX_fifo_queue #( VX_fifo_queue #(
.DATAW ($bits(wid)), .DATAW ($bits(wid)),
.DEPTH (ISSUE_QUEUE_DEPTH + ISSUE_QUEUE_DEPTH) // @perf: seems to require deeper depth than the FEDP issue queues to
// not cause stalls.
.DEPTH (2 * ISSUE_QUEUE_DEPTH)
) wid_queue ( ) wid_queue (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -167,7 +172,7 @@ module VX_tensor_threadgroup #(
// threadgroups, so we need only 1 queue per octet for B // threadgroups, so we need only 1 queue per octet for B
VX_fifo_queue #( VX_fifo_queue #(
.DATAW ($bits(A_frag) + $bits(B_frag) + $bits(C_frag)), .DATAW ($bits(A_frag) + $bits(B_frag) + $bits(C_frag)),
.DEPTH (ISSUE_QUEUE_DEPTH) .DEPTH (ISSUE_QUEUE_DEPTH)
) input_buffer ( ) input_buffer (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),