Skip to content

fd05_p3_override: p3 Override

Open waveform

Runs Surfer in the browser, with nothing to install. The signal setup takes a few seconds to apply.

fd05_p3_override is the FD-5 demonstration scenario. It is a demo, not a verification test. It runs a loop around two conditional branches whose outcomes follow short periodic patterns. While TAGE is learning them, the statistical corrector (SC) at p3 disagrees with the direction published at p2, the slot's successor changes, and the cluster raises a p3 redirect with the slot's pred_src set to PRED_SC. The run passes if the FTQ accepts at least one such redirect.

The scenario runs on the one-slot model (Makefile SCEN_2S, DOC-006). FD-22 (docs/fd22_sc_slot1.md) shows SC reversing the second slot of a block on the two-slot model. A recorded run:

-I: fd05_p3_override: PASS, p3 redirects 33, SC reversals among them 33,
    cyc 975

1. The program

Source: programs/fd05_p3_override.svh, task prog_fd05 (tb_fe_top's hist). top = RESET_VECTOR + 8 (0x80000008).

RESET_VECTOR+0   mixed(3)                                  8 bytes
top              c.addi, addi                              6 bytes
top+6    A       bne x6,x0,top+16     taken on odd passes
top+10           c.addi, addi         skipped when A is taken
top+16           c.addi
top+18   B       bne x6,x0,top+28     taken when pass % 3 == 2
top+22           c.addi, addi         skipped when B is taken
top+28           c.addi
top+30           bne x6,x0,top        back edge: taken on passes 0..94
top+34           mixed(3)
top+42           jal x0,0             halt (demo_go -> W.halt)
  • 96 passes. Each pass resets W.wp = top and writes the same bytes; cskip(tk, 2) (fe_tb_writer.svh:502-519) writes the skipped mixed(2) in dry mode when taken, so the code is fixed and only the expected stream changes.
  • A has period 2, B period 3. Outcomes come from the writer (bcc, fe_tb_writer.svh:412-419), not from x6.
  • Lookup blocks that hold A or B, and the position (halfwords from the lookup PC) of each: top (0x80000008), A at 3; top+10 (0x80000012, after A falls through on a redirect), B at 4; top+16 (0x80000018, A's target), B at 1; top+18 (0x8000001A), B at 0. On this model an FTB entry holds one branch. The entry for top holds A. Its end is B's PC, top+18 (the split at update, planning/arch/ftb_decisions.md section 3). B is predicted in the block that starts there.

2. Why the program produces SC reversals

  1. SC answers at p3. rtl/core/frontend/bpu/rtl/bp_cluster.sv:1296- 1297: w_sc_hit_p3[s] is the SC response for slot s (w_sc_pred_rdy_p3), with sc_enable (fe_cfg.sc_enable, fe_top.sv:486) and a branch_id matching the entry at p3.
  2. The direction. Lines 1301-1304: for a conditional slot with an SC hit, and no loop-predictor direction at p2 (~r_lp_use_p3), the p3 direction is sc_pred_tkn; otherwise the p2 direction.
  3. The redirect. Lines 1307-1325: the p3 successors are chained from the highest slot down, as at p2, and slot s redirects when it is reachable and its p3 successor differs from its p2 successor. A p3 redirect therefore fires only where SC changed a direction, or where a later slot's change moves an earlier not-taken slot's successor.
  4. The slot record. Lines 1340-1352 (p3_slot_comb): the p3 slot takes sc_pred_tkn, and pred_src becomes PRED_SC when that differs from the p2 taken.
  5. FTQ acceptance and the new request: ok_redir_p3 (rtl/core/frontend/ftq/rtl/ftq_shadow.sv:220-221); ftq_npc.sv: 338-348 (arm 3) takes the p3 PC, ahead of any p2 redirect.

From the FST, the 33 accepted p3 redirects, by the slot SC reversed (slot 0 in every one):

block        branch  pos  p2 pred_src   p3 taken   target       count
0x80000008   A       3    PRED_TAGE     0          0x8000001A      17
0x80000008   A       3    PRED_TAGE     0          0x80000028       2
0x80000008   A       3    PRED_TAGE     1          0x80000018       2
0x80000012   B       4    PRED_TAGE     1          0x80000024       2
0x80000018   B       1    PRED_TAGE     1          0x80000024       7
0x8000001A   B       0    PRED_TAGE     1          0x80000024       3

In every one the p2 direction came from TAGE (r_slot_p3[0].pred_src PRED_TAGE). A reversed to not taken redirects to the end of the block at top. Once the FTB entry has been split, that end is B's PC, 0x8000001A; the split is an update for B at position 9 from top, at cyc 261. In the two cases before it the end is top + 0x20, 0x80000028. The redirects fall between cyc 47 and 822 of the 975-cycle run.

3. What the monitor counts

The counts come from tb/fe_demo_mon.svh:424-438, taken inside the meas window.

n_redir_p3      u_ftq.w_ok_redir_p3 && any bpu_redir_p3[s].valid
                a cycle in which the FTQ accepted a p3 redirect
n_redir_p3_sc   the same, with a redirecting slot s for which
                bpu_slot_val_p3, bpu_slot_idx_p3 == bpu_redir_idx_p3
                and bpu_slot_p3[s].pred_src == PRED_SC

The proof (tb/fe_demo_proofs.svh:64-68) is n_redir_p3_sc > 0. tb_fe_demo.sv also requires the whole stream to retire with no errors before it accepts the proof.

4. Reading the numbers

  • p3 redirects 33, SC reversals among them 33: every accepted p3 redirect in this run had an SC-decided slot. The p3 direction differs from the p2 one only through SC (section 2, item 2), which is why the two figures agree here.
  • Both are counts of cycles, not slots.
  • The p2 direction that SC reversed was TAGE's in all 33 (from the FST). The monitor does not check that; it checks only pred_src == PRED_SC on the p3 slot.
  • n_tage_prov is not in this scenario's result line. DOC-002 changed it to sum the slots of a cycle; on this model a cycle has one slot.
  • Departure from planning/showcase/fe_demo_scenarios.md FD-5: its program is "a branch TAGE predicts weakly and SC corrects". As built, two history-dependent conditionals (periods 2 and 3) are run; whether TAGE's prediction was weak is not checked by the proof or recorded here. The proof is FD-5's.

5. Waveform

The saved view is waves/fd05_p3_override.sucl, which make view S=fd05_p3_override opens in Surfer. Look at:

  • group tage_p2: w_tage_pred_rdy_p2 and tage_pred_meta_p2.[0] (tage_prm_comp, tage_prm_ctr, tage_pred_tkn) for the block.
  • group bp_cluster_p2: the p2 slot with pred_src PRED_TAGE and its taken.
  • group sc_p3: w_sc_pred_rdy_p3, w_sc_hit_p3 and w_taken_p3 opposite to the p2 slot's taken.
  • group bp_cluster_p3: bpu_slot_p3.[0].pred_src PRED_SC and bpu_redir_p3.[0].valid with its target_pc.
  • group ftq: w_ok_redir_p3 accepting it.

Open waveform

Runs Surfer in the browser, with nothing to install. The signal setup takes a few seconds to apply.

Run it from examples/fe_demo/: make fd05_p3_override runs the scenario; make view S=fd05_p3_override opens its waveform in Surfer.

Files: programs/fd05_p3_override.svh, waves/fd05_p3_override.sucl.

Source of this page on GitHub