5 Sep 2023

Previous journal:	Next journal:
0142-2023-09-04.md	0144-2023-09-06.md

raybox-zero

New tests

test009: Start changing pov outputs

The first commit in the test009 branch was only run a single combo test at 4x2:5. The results were basically the same as for test005, which is to be expected since there was no functional change in the pov.v code -- just a different way to define the same logic.

test009-2: Restrict pov output bits to Q9.9

Click for details...

Code:

tt04-raybox-zero: aec8784: harden_test: Dry drun, ROUTING_CORES options
- Equivalent to: 93dd1cd: config.tcl default of 25MHz, using raybox-zero merged test005-shmul branch.
src/raybox-zero: 68dd419: (test009) Restricting pov output bits

Summary:

pov's vectors all changed from SQ12.12 to SQ9.9 range on output ports to see if it would make any difference on synth.

Options used:

  STARTED: 2023-09-04 23:34:54
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
    STOPT: 0
  OUTFILE: stats-test009-2.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test009-2: Restrict pov output bits to Q9.9
   RCORES: 7
      DRY: 0
 FINISHED: 2023-09-05 00:27:34

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	47.71	47.78	25.00	25.00
utilisation_%	51.67	25.64	51.67	25.64	0.00	0.00	43.85	21.76	43.57	21.62
wire_length_um	218,026	221,045	219,926	222,004	0	0	158,234	164,957	158,501	158,461
TotalCells	20,634	35,608	20,663	35,578	0	0	20,105	35,323	20,002	35,122
cells_pre_abc	12,041	12,041	12,041	12,041	0	0	12,041	12,041	12,041	12,041
synth_cells	7,520	7,520	7,520	7,520	0	0	6,587	6,587	6,587	6,587
logic_cells	8,005	7,976	8,113	8,050	0	0	7,109	7,117	7,081	7,081
pin_antennas	4	10	3	9	0	0	0	1	3	2
net_antennas	4	9	3	8	0	0	0	1	3	2
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	-0.96	-0.93	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	-43.33	-43.89	0.00	0.00

Findings (compared to test005):

All combos now synth.
4x2:5 util reduces from 48.86% to 43.57%
4x2:2 can now reach 50MHz (up from 47.62)
Combo 4 is now closer to 50MHz (up to ~47.7 from ~45.5)

test009-3: restrict pov output bits to typical minimum (Q7.9 and Q2.9)

Click for details...

Code:

tt04-raybox-zero: aec8784: harden_test: Dry drun, ROUTING_CORES options
- Equivalent to: 93dd1cd: config.tcl default of 25MHz, using raybox-zero merged test005-shmul branch.
src/raybox-zero: b04ac58: (test009) pov output is now Q7.9 for player, Q2.9 for facing and vplane

Summary:

Go from SQ9.9 for everything to playerX/Y using SQ7.9 and facing/vplane using SQ2.9

Options used:

  STARTED: 2023-09-04 23:46:06
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
    STOPT: 0
  OUTFILE: stats-test009-3.md
   SELECT: :[1245]
    FORCE: 1
      TAG: test009-3: restrict pov output bits to typical minimum (Q7.9 and Q2.9)
   RCORES: 14
      DRY: 0
 FINISHED: 2023-09-05 00:28:20

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	48.71	48.52	25.00	25.00
utilisation_%	48.45	24.05	48.45	24.05	0.00	0.00	41.40	20.55	41.35	20.52
wire_length_um	200,721	200,479	203,397	202,421	0	0	146,345	149,706	140,652	143,309
TotalCells	20,290	35,189	20,275	35,325	0	0	19,930	34,932	19,747	34,762
cells_pre_abc	11,838	11,838	11,838	11,838	0	0	11,838	11,838	11,838	11,838
synth_cells	7,042	7,042	7,042	7,042	0	0	6,242	6,242	6,242	6,242
logic_cells	7,469	7,472	7,588	7,578	0	0	6,721	6,728	6,674	6,687
pin_antennas	3	2	3	4	0	0	2	3	3	2
net_antennas	3	2	3	4	0	0	2	3	3	2
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	-0.53	-0.61	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	-18.35	-25.46	0.00	0.00

Findings (compared to test009-2 above):

4x2:5 util reduces from 43.57% to 41.35%
Combo 4 speed improvement (up to ~48.5MHz from ~47.7)

test009-4: pov SPI takes 74b: UQ6.9 for playerX/Y, SQ2.9 for others

Click for details...

Code:

tt04-raybox-zero: aec8784: harden_test: Dry drun, ROUTING_CORES options
- Equivalent to: 93dd1cd: config.tcl default of 25MHz, using raybox-zero merged test005-shmul branch.
src/raybox-zero: f45301a: (test009): pov SPI takes 74b: UQ6.9 for playerX/Y, SQ2.9 for others. Sim updated to suit

Summary:

playerX/Y changed from SQ7.9 to UQ6.9
SPI input stream size changed from 144 bits to 74 bits, and internal pov registers/buffers updated accordingly

Options used:

  STARTED: 2023-09-05 12:14:17
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
    STOPT: 0
  OUTFILE: stats-test009-4.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test009-4: pov SPI takes 74b: UQ6.9 for playerX/Y, SQ2.9 for others
   RCORES: 7
      DRY: 0
 FINISHED: 2023-09-05 13:07:31

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	46.99	46.69	25.00	25.00
utilisation_%	49.23	24.43	49.23	24.43	0.00	0.00	39.40	19.56	39.33	19.52
wire_length_um	225,379	225,137	227,330	227,253	0	0	140,968	146,688	135,373	137,823
TotalCells	20,499	35,418	20,376	35,510	0	0	19,609	34,615	19,472	34,516
cells_pre_abc	11,774	11,774	11,774	11,774	0	0	11,774	11,774	11,774	11,774
synth_cells	7,263	7,263	7,263	7,263	0	0	5,912	5,912	5,912	5,912
logic_cells	7,639	7,610	7,716	7,736	0	0	6,279	6,289	6,240	6,243
pin_antennas	5	3	0	0	0	0	2	3	2	0
net_antennas	5	3	0	0	0	0	2	3	2	0
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	-1.28	-1.42	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	-65.45	-75.88	0.00	0.00

Findings (compared to test009-3 above):

4x2:5 util reduces from 41.35% to 39.33%
Combo 4 speed is worse (down to ~46.7MHz from ~48.5)

test009-4 TARGETING 2x2

I ran this command:

time ./harden_test.sh -t "test009-4 with 2x2 target" -o stats-test009-4-2x2:5.md :[1245] -p 7 -u 0.82 -s 2x2 -c 5

...to target a 2x2 tile size with a density of 0.82 (over the default of 0.65) and to my surprise it fit.

NOTE: I did also try combos 1..4 with a 2x2 target, but they all failed given they don't override target density, so I didn't bother including their results.

Click for details...

This run was otherwise the same as test009-4.

Options used:

  STARTED: 2023-09-05 14:22:20
    SIZES: 2x2 (sorted: 2x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
  DENSITY: 0.82
    STOPT: 0
  OUTFILE: stats-test009-4-2x2.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test009-4 with 2x2 target
   RCORES: 7
      DRY: 0
 FINISHED: 2023-09-05 14:28:57

	2x2:5
suggested_mhz	25.00
utilisation_%	79.84
wire_length_um	137,415
TotalCells	11,416
cells_pre_abc	11,774
synth_cells	5,912
logic_cells	6,252
pin_antennas	0
net_antennas	0
spef_wns	0.00
spef_tns	0.00

Findings:

2x2:5 SUCCEEDS at 25MHz with ~80% utilisation

NOTE:

I'm not clear on whether 80% utilisation is OK for TT04. Here Uri thought 88% was OK for 1x1? Matt just confirmed that 80% for a 2x2 is acceptable for TT04.

How much power will this consume in normal operation?

metrics.csv suggests 0.236µW for typical switching, where it is written as uW:

 power_typical_internal_uW :    0.172
power_typical_switching_uW :    0.236
  power_typical_leakage_uW : 2.84e-08

...but it's confusing because it says these figures are in watts in runs/wokwi/reports/signoff/31-rcx_sta.power.rpt:

===========================================================================
report_power
============================================================================
======================= Typical Corner ===================================

Group                  Internal  Switching    Leakage      Total
                          Power      Power      Power      Power (Watts)
----------------------------------------------------------------
Sequential             6.50e-04   1.62e-04   3.72e-09   8.12e-04   0.2%
Combinational          1.71e-01   2.35e-01   2.47e-08   4.07e-01  99.8%
Macro                  0.00e+00   0.00e+00   0.00e+00   0.00e+00   0.0%
Pad                    0.00e+00   0.00e+00   0.00e+00   0.00e+00   0.0%
----------------------------------------------------------------
Total                  1.72e-01   2.36e-01   2.84e-08   4.08e-01 100.0%
                          42.2%      57.8%       0.0%

So which is right? 0.236µW or 0.236W, because the latter could be bad, especially if the total is ~408mW in such a small area (~339x226um)

Power in normal use, however, might be nowhere near this bad... the bulk of switching might stop at the end of each line trace, which could be as little as 10% in... and in fact the heavy switching for calculating multiplication and reciprocals might be more like 1% of the time.

test010: Fixed point internal precision changed to Q9.10 (down from Q12.12)

Click for details...

Code:

tt04-raybox-zero: 48b07c3: harden_test: Fix for overriding PL_TARGET_DENSITY
- Equivalent to: 9ffc209: Update submodule for pov UQ6.9/SQ2.9 changes, and try fit in 2x2 with 82% target density
src/raybox-zero: 55753a4: test010: Fixed point internal precision changed to Q9.10 from Q12.12

Summary:

Change fixed_point_params.v to use Q9.10 instead of usual Q12.12
Manually updated lzc.v accordingly.
Updated other code to make it automatically support a change in these parameters.

Options used:

  STARTED: 2023-09-05 18:03:22
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
  DENSITY: 0.65
    STOPT: 0
  OUTFILE: stats-test010.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test010: Q9.10 overall internally
   RCORES: 7
      DRY: 0
 FINISHED: 2023-09-05 18:43:15

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	50.00	50.00	25.00	25.00
utilisation_%	37.90	18.81	37.90	18.81	0.00	0.00	31.66	15.71	31.66	15.71
wire_length_um	153,767	151,102	161,197	158,307	0	0	117,756	119,241	113,366	114,396
TotalCells	19,113	34,105	19,118	34,149	0	0	18,611	33,649	18,580	33,546
cells_pre_abc	8,441	8,441	8,441	8,441	0	0	8,441	8,441	8,441	8,441
synth_cells	5,492	5,492	5,492	5,492	0	0	4,688	4,688	4,688	4,688
logic_cells	5,806	5,823	5,924	5,958	0	0	5,070	5,087	5,003	4,993
pin_antennas	2	3	2	1	0	0	1	6	2	1
net_antennas	2	3	2	1	0	0	1	6	2	1
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Findings (compared to test009-4 above):

All combos now synth and meet 50MHz timing
4x2:5 util down from 39.33% to 31.66%
Combos 1 and 2 reduced substantially.
NOTE: There is a small amount of precision lost that is visually noticeable: Wall edges get a little more jagged, but not too bad (and possibly worth the util savings). If we can get textures working soon too, then this might be less or more noticeable.

test010-2: Try Q8.10 instead of Q9.10

Click for details...

Code:

tt04-raybox-zero: 48b07c3: harden_test: Fix for overriding PL_TARGET_DENSITY
- Equivalent to: 9ffc209: Update submodule for pov UQ6.9/SQ2.9 changes, and try fit in 2x2 with 82% target density
src/raybox-zero: 7ec3bf8: (test010) Try Q8.10 instead of Q9.10

Summary:

Small change from Q9.10 to Q8.10, since I don't think we need that extra range at all.

Options used:

  STARTED: 2023-09-05 18:04:20
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
  DENSITY: 0.65
    STOPT: 0
  OUTFILE: stats-test009-2.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test010-2: Try Q8.10 instead of Q9.10
   RCORES: 14
      DRY: 0
 FINISHED: 2023-09-05 18:40:52

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	50.00	50.00	25.00	25.00
utilisation_%	37.07	18.39	37.07	18.39	0.00	0.00	30.60	15.19	30.60	15.19
wire_length_um	152,686	153,424	154,568	154,460	0	0	107,493	114,422	103,466	104,715
TotalCells	18,971	34,087	19,044	34,029	0	0	18,618	33,660	18,485	33,522
cells_pre_abc	7,910	7,910	7,910	7,910	0	0	7,910	7,910	7,910	7,910
synth_cells	5,393	5,393	5,393	5,393	0	0	4,590	4,590	4,590	4,590
logic_cells	5,742	5,754	5,813	5,800	0	0	4,993	4,996	4,907	4,895
pin_antennas	2	2	3	3	0	0	2	2	1	2
net_antennas	2	2	3	3	0	0	2	2	1	2
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Findings (compared to test010 immediately above):

Slightly smaller util: 30.6% at 4x2:5, down from 31.66%
Visual appearance is no worse.

test010-3: Try Q8.9

Click for details...

Code:

tt04-raybox-zero: 48b07c3: harden_test: Fix for overriding PL_TARGET_DENSITY
- Equivalent to: 9ffc209: Update submodule for pov UQ6.9/SQ2.9 changes, and try fit in 2x2 with 82% target density
src/raybox-zero: 7ec3bf8: as above, but...
- Modified to be Q8.9 instead. Not committed.

Summary:

Tried out Q8.9 instead of Q8.10

Options used:

  STARTED: 2023-09-05 19:22:31
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
  DENSITY: 0.65
    STOPT: 0
  OUTFILE: stats-test010-3.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test010-3: Try Q8.9
   RCORES: 14
      DRY: 0
 FINISHED: 2023-09-05 19:54:00

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	50.00	50.00	25.00	25.00
utilisation_%	34.06	16.90	34.06	16.90	0.00	0.00	28.32	14.05	28.32	14.05
wire_length_um	136,016	130,809	136,913	133,245	0	0	101,304	100,780	102,260	96,453
TotalCells	18,567	33,617	18,650	33,721	0	0	18,132	33,256	18,152	33,131
cells_pre_abc	7,274	7,274	7,274	7,274	0	0	7,274	7,274	7,274	7,274
synth_cells	4,903	4,903	4,903	4,903	0	0	4,178	4,178	4,178	4,178
logic_cells	5,226	5,235	5,319	5,360	0	0	4,553	4,549	4,525	4,497
pin_antennas	2	4	1	3	0	0	1	3	5	0
net_antennas	2	4	1	3	0	0	1	3	5	0
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Findings:

Better again, but the walls have ragged edges now. It also surfaced a visual glitch that exists in the other test010 instances:
I've noticed this glitch now in Q9.10, though less severe, and usually only when close to a visual axis.
It might have something to do with overflows and perhaps using the saturation signal might be a way to stop this.

test010-4: Try Q10.10

Click for details...

Code:

tt04-raybox-zero: 48b07c3: harden_test: Fix for overriding PL_TARGET_DENSITY
- Equivalent to: 9ffc209: Update submodule for pov UQ6.9/SQ2.9 changes, and try fit in 2x2 with 82% target density
src/raybox-zero: af9d694: (test010): Try Q10.10

Summary:

Because of visual glitches seen above, I tried Q10.10 instead.

Options used:

  STARTED: 2023-09-05 20:43:02
    SIZES: 4x2 8x2 (sorted: 4x2 8x2)
   COMBOS: 5 4 1 2 3 (sorted: 1 2 3 4 5)
  DENSITY: 0.65
    STOPT: 0
  OUTFILE: stats-test010-4.md
   SELECT: :[1245]
    FORCE: 0
      TAG: test010: Try Q10.10
   RCORES: 7
      DRY: 0
 FINISHED: 2023-09-05 21:22:20

	4x2:1	8x2:1	4x2:2	8x2:2	4x2:3	8x2:3	4x2:4	8x2:4	4x2:5	8x2:5
suggested_mhz	25.00	25.00	50.00	50.00	0.00	0.00	50.00	50.00	25.00	25.00
utilisation_%	38.63	19.17	38.63	19.17	0.00	0.00	32.86	16.31	32.86	16.31
wire_length_um	157,046	151,410	152,844	158,081	0	0	117,648	122,223	118,080	113,917
TotalCells	19,188	34,209	19,236	34,252	0	0	18,855	33,844	18,728	33,723
cells_pre_abc	9,019	9,019	9,019	9,019	0	0	9,019	9,019	9,019	9,019
synth_cells	5,673	5,673	5,673	5,673	0	0	4,890	4,890	4,890	4,890
logic_cells	6,033	6,036	6,062	6,103	0	0	5,298	5,280	5,225	5,217
pin_antennas	2	1	3	1	0	0	0	1	0	1
net_antennas	2	1	3	1	0	0	0	1	0	1
spef_wns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
spef_tns	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Findings:

Given Q12.12 was getting 39.33% util, I'm happy with Q10.10 being 32.86%. Still a good improvement.
All combos still hit their target clock speeds, so I'm happy with that.
There are still minor visual glitches with Q10.10, especially near axes, but I think it's acceptable.

Accomplishments

Reduced the range/resolution of outputs from pov. The module still outputs 24-bit values, but synth optimisation can benefit from the following:
- SPI vector update only needs 74 bits now instead of 144, so the buffers/registers are smaller.
- playerX/Y uses UQ6.9 (15b/ea) to cover [0,63] range with 1/512 resolution (which 'feels' like 3.5~5mm minimum step size).
- This reduces
- Sim updated to suit.
In sim, CTRL+up/down will increase/decrease 'facing' vector scaling, in the range (0.01,1.99). This narrows/widens FOV.
In sim, CTRL+right/left increases/decreases 'vplane' vector scaling, in the range (0.01,1.99). This widens/narrows apparent aspect ratio.
Merged test009 experiment into main line via PR#1
P key in sim can now toggle rotation of the display by 90 degrees.
Tried Q9.10, Q8.10, and even Q8.9 and decided they were OK but not always good enough.
I've settled on Q10.10 as probably OK for all internal fixed-point stuff.
- There's still occasionally a visual glitch (of the kind seen in the image above) but it's pretty rare and not overly noticeable.
- The savings it offers in area is important to me right now, and if I find that other features added later still leave enough room, I can try pusing up to Q11.11 or whatever.
- I decided to merge test010 with this Q10.10 config into main.

Next steps

See if there is some cheap effect we can do with wall/ceiling.
Remove o_sat and i_abs (make it always absolute) from reciprocal and see if it makes any difference to area.
Try running on FPGA at a slower (or faster) clock.

Notes

When modifying `Qm and `Qn in fixed_point_params.v, it is necessary to update lzc.v to match. Ensure:
- The number of casez options matches the number of `Qmn bits, plus one (i.e. 0..n).
- The pattern is extended (or contracted) as required on the right-hand side of the bit constants.

Planning TT04 submission/completion

Doco
Look for @@@, SMELLs, TODOs, etc.
Test on FPGA
Configurability (e.g. VGA timing alts)
Check antennas and other reports e.g. fanout.
Check possible saturation check error in reciprocal
Create libre textures that vary colour.
tt04-solo-squash with SPI
Registered outputs. Could that just be in the TT04 wrapper?

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

0143-2023-09-05.md

0143-2023-09-05.md

5 Sep 2023

raybox-zero

New tests

test009: Start changing pov outputs

test009-2: Restrict pov output bits to Q9.9

test009-3: restrict pov output bits to typical minimum (Q7.9 and Q2.9)

test009-4: pov SPI takes 74b: UQ6.9 for playerX/Y, SQ2.9 for others

test009-4 TARGETING 2x2

test010: Fixed point internal precision changed to Q9.10 (down from Q12.12)

test010-2: Try Q8.10 instead of Q9.10

test010-3: Try Q8.9

test010-4: Try Q10.10

Accomplishments

Next steps

Notes

Planning TT04 submission/completion

Files

0143-2023-09-05.md

Latest commit

History

0143-2023-09-05.md

File metadata and controls

5 Sep 2023

raybox-zero

New tests

test009: Start changing pov outputs

test009-2: Restrict pov output bits to Q9.9

test009-3: restrict pov output bits to typical minimum (Q7.9 and Q2.9)

test009-4: pov SPI takes 74b: UQ6.9 for playerX/Y, SQ2.9 for others

test009-4 TARGETING 2x2

test010: Fixed point internal precision changed to Q9.10 (down from Q12.12)

test010-2: Try Q8.10 instead of Q9.10

test010-3: Try Q8.9

test010-4: Try Q10.10

Accomplishments

Next steps

Notes

Planning TT04 submission/completion