From 5a0f2895736846155762b454f204aa7af3c20a9a Mon Sep 17 00:00:00 2001 From: Luke Kenneth Casson Leighton Date: Sun, 3 May 2020 23:33:12 +0100 Subject: [PATCH] describe concurrent computational unit --- 3d_gpu/architecture/6600scoreboard.mdwn | 43 +++++++++++++++++++++++++ 1 file changed, 43 insertions(+) diff --git a/3d_gpu/architecture/6600scoreboard.mdwn b/3d_gpu/architecture/6600scoreboard.mdwn index 9d7263c8f..616356768 100644 --- a/3d_gpu/architecture/6600scoreboard.mdwn +++ b/3d_gpu/architecture/6600scoreboard.mdwn @@ -114,6 +114,49 @@ Source: * [Priority Pickers](https://git.libre-riscv.org/?p=nmutil.git;a=blob;f=src/nmutil/picker.py;hb=HEAD) * [ALU Comp Units](https://git.libre-riscv.org/?p=soc.git;a=blob;f=src/soc/experiment/compalu.py;h=f7b5e411a739e770777ceb71d7bd09fe4e70e8c0;hb=b08dee1c3e8cf0d635820693fe50cd0518caeed2) +# Concurrent Computational Unit + +With the original 6600 having only a 2-stage pipelined FPU (which took +many years to notice from examining the now-archaic notation from James +Thornton's book, "Design of a Computer"), there is no actual use of this +pipeline capability at the front-end Function Unit. Instead it is +treated effectively as a Finite State Machine, only one result to be +computed at a time. + +Mitch Alsup recommends, when using pipelines, to allow multiple +Function Unit "front-ends", each one having inputs that were pushed +into a particular stage of the pipeline, and, therefore, those multiple +Function Units also track and store the result as it comes out. + +The trick then is to have a method that ensures that FU front-end #1 +can get result #1 when it pops out the end of the (serial) pipeline. +Mitch recommends using timing chains, here. + +Note in this diagram that there are *multiple* ISSUE, GO\_READ and GO\_WRITE +signals. These link up to the Function Unit's ISSUE, GO\_RD and GO\_WR, +where the latches are, that will (on an available slot) feed the pipeline +with incoming data. + +[[!img concurrent_comp_unit.png size="600x"]] + +The actual design being used is slightly different, in the following +ways: + +* Due to micro-coding and thus external contention, the pipelines + have a ready/valid signalling arrangement that can result in + a stall cascading back down the pipeline. Thus a timing chain + is not appropriate. +* A decision was therefore made to pass a "context" alongside the + operands, which is the "Function Unit Index". It is *this* information + that is used to "reassociate" the result with the correct FU, when + the result is produced. +* With "Shadow cancellation" being in effect, *additional* global + context is passed (combinatorially) to every single stage of the + pipeline, as an unary bitmask. If any Function Unit's "GO_DIE" + signal is asserted, the corresponding bit in the unary mask is + asserted, terminating effective immediate the intermediary data + anywhere in the pipeline from progressing further, thus saving power. + # Multi-in cascading Priority Picker Using the Group Picker as a fundamental unit, a cascading chain is created, -- 2.30.2