DSP - BLOCK STD ERROR OF MEAN TM
TM (time multiplexed) twin of Block Std Error of Mean: TM Factor (2..16) independent channels packed on one wide bus, each computing sem = sigma / sqrt(N) - the error bar on that lane’s block mean - of its own block of N = 2^EXP samples (EXP on a runtime pin, shared by every lane). One shared frame - one EXP, one IN_DV, one OUT_DV per block - with per-lane accumulators at II=1 and ONE serial tail serving the lanes one after the other, so the tail resources stay those of the scalar block whatever the lane count. Lane 0 sits in the LOW bits of every packed bus. Same per-lane numbers as the scalar twin, bit for bit. Optional shared BUSY / INTEGRATING / SAMPLE_COUNT status outputs.
Introduction
Per lane, over each block of N samples:
$$ \mathrm{SEM}_i = \frac{\sigma_i}{\sqrt{N}} = \sqrt{\frac{N\sum_k x_i[k]^2 - (\sum_k x_i[k])^2}{N^3}} $$
This is the time multiplexed (TM) twin of the scalar Block Std Error of Mean block: TM Factor independent channels packed on one wide bus, each computing the standard error of the mean of its own block of N consecutive samples. Nothing is shared between the channels except the frame.
As in the scalar twin, sqrt(N) is HALF A SHIFT that folds into the same radicand as the variance - the exponent enters the root as 3EXP instead of the 2EXP a plain standard deviation uses - so each lane costs exactly what a Block Std Dev TM lane costs: one serial square of that lane’s S1 and one serial square root, no divider, and not one clock more.
The TM contract
INis TM Factor lanes of input width bits each, lane 0 in the LOW bits (lane 0 is the oldest sample of the clock) - the same packing as every other TM block in the toolchain. Each lane is an independent channel: its own accumulators, its own result.- ONE shared
EXPpin, ONEIN_DV, one frame: all lanes start and end their blocks on the same accepted clocks.BUSY,INTEGRATINGandSAMPLE_COUNTtherefore stay scalar -SAMPLE_COUNTcounts per-lane samples, which are identical in every lane by construction. - Accumulation runs at II=1 on the packed bus, one accumulator per lane. TM Factor times the accumulator registers is the unavoidable cost of TM state - plus one real x*x multiplier per lane in the accumulation path.
- The serial post-processing is ONE shared engine serving the lanes one after the other. It runs once per block, so lane-multiplexing it keeps the tail resources of the SCALAR block - ONE shift-add square and ONE digit-recurrence root, not TM Factor of each - at the cost of tail latency x TM Factor, which is irrelevant against a whole block of samples.
- ONE
OUT_DVper block, after the LAST lane’s tail completes. All output lanes are staged as each lane finishes and committed together on theOUT_DVclock, so every packed output moves on that clock and no other.
N is a runtime input: the block size is $N = 2^{\mathrm{EXP}}$, EXP clamped to Max Block Exponent and latched on the first accepted sample of a block, so a change takes effect on the NEXT block - for every lane at once. Because N is a power of two, every division by N in this family is an exact shift.
When to use this instead of TM Factor scalar blocks
One TM block and TM Factor scalar blocks compute the same numbers. The TM block pays the per-lane accumulators (unavoidable either way) but shares ONE frame, ONE control FSM and ONE serial tail across all lanes - the more expensive the scalar tail, the more it saves. The price is tail latency (TM Factor times the scalar tail) and the coupling of the frame: all lanes must share the same block length and the same sample cadence. Channels that need different block sizes need scalar blocks.
Pin Description
IN_DV is high.
Properties
Number of INTEGER bits of the input sample (per lane) (the sign, when present, uses one of them).
Integer bits of ONE LANE of the input (the sign, when present, uses one of them). 1..64. Default 16.Default: 16
Options: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
Number of FRACTIONAL bits of the input sample (per lane), i.e. the bits to the right of the binary point. Total width = integer + fractional bits, and must not exceed 64.
Fractional bits of one lane of the input. 0..64, total lane width 2..64 bits. Default 0.Default: 0
Options: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
Select whether the input sample (per lane) is signed (two’s complement) or unsigned.
SIGNED (two’s complement) or UNSIGNED lanes. Default SIGNED. Applies to every lane.Default: SIGNED
Options: UNSIGNED SIGNED
Number of INDEPENDENT time-multiplexed channels packed on the IN bus and on every result bus. Lane 0 occupies the LOW bits (lane 0 = the oldest sample of the clock), the same packing as every other TM block. All lanes share one EXP / IN_DV / frame; each lane gets its own accumulators, but the serial post-processing is ONE engine serving the lanes one after the other, so the tail latency (and the minimum usable EXP) grows with this factor.
Number of independent channels packed on the buses, 2..16. Default 4. Multiplies the IN width, every packed output width AND the serial tail length (the shared tail serves the lanes one after the other), so it also raises the minimum usable EXP.Default: 4
Options: 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Largest block-size exponent the accumulators are sized for: the block can be up to 2^MaxBlockExponent samples long. The EXP input is clamped to this value at run time. Raising it widens the internal accumulators, and ON THE BLOCKS WHOSE SERIAL ENGINES ARE SIZED FROM THOSE ACCUMULATORS (Coefficient of Variation, SNR, Skewness, Kurtosis, Correlation, Autocorrelation, Linear Regression) it also LENGTHENS THE SERIAL TAIL – even when the runtime EXP is small. Keep it at the largest block you actually use. The default of 20 covers blocks of up to 1048576 samples.
Largest block-size exponent the per-lane accumulators are sized for; the EXP input is clamped to it at run time. Raising it widens every lane’s accumulator. 1..31, default 20.Default: 20
Options: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Number of INTEGER bits of the SEM output (per lane) (the sign, when present, uses one of them).
Integer bits of ONE LANE of the SEM output. 1..64, default 8. Each integer bit also costs one tail clock per lane (it widens the root engine).Default: 8
Options: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
Number of FRACTIONAL bits of the SEM output (per lane), i.e. the bits to the right of the binary point. Total width = integer + fractional bits, and must not exceed 64.
Fractional bits of one lane. 0..64, default 8 - and fractional bits are where a standard error actually lives: averaging exists to make this number smaller than an input LSB. Each one costs one tail clock per lane.Default: 8
Options: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
Select whether the SEM output (per lane) is signed (two’s complement) or unsigned.
SIGNED or UNSIGNED lanes. Default UNSIGNED (a standard error is non negative by construction).Default: UNSIGNED
Options: UNSIGNED SIGNED
YES: the BUSY (high from the first sample of a block until its result is out – it COVERS THE SERIAL TAIL, and its last high clock IS the OUT_DV pulse) pin is present. NO: the pin AND all of its logic are removed BEFORE synthesis, so nothing is paid for it.
YES: the shared BUSY pin exists. Default NO.Default: NO
Options: NO YES
YES: the INTEGRATING (high only while the block is ACCUMULATING; it drops as soon as the N-th sample has been taken and the tail starts, so BUSY-and-not-INTEGRATING means ‘computing’) pin is present. NO: the pin AND all of its logic are removed BEFORE synthesis, so nothing is paid for it.
YES: the shared INTEGRATING pin exists. Default NO.Default: NO
Options: NO YES
YES: the SAMPLE_COUNT (32 bit, how many samples have been accumulated so far in the current block: 1 after the first, N after the N-th. It is NOT cleared at the block end – it holds N until the NEXT block’s first accepted sample takes it back to 1. On a CONTINUOUS stream that happens DURING the serial tail, so at OUT_DV it reads how far into the next block the input has already got, NOT N. To capture the length of the block being presented, latch SAMPLE_COUNT on the clock INTEGRATING falls – that one always reads N) pin is present. NO: the pin AND all of its logic are removed BEFORE synthesis, so nothing is paid for it.
YES: the shared 32 bit SAMPLE_COUNT pin exists. Default NO.Default: NO
Options: NO YES
ROUND: round to nearest when a result has to be requantised into a coarser output format. TRUNCATE: drop the bits (cheaper, adds a negative bias).
ROUND: round to nearest at each lane’s final requantisation. TRUNCATE: drop the bits. Default ROUND.Default: ROUND
Options: TRUNCATE ROUND
YES: clip to the largest representable value of each output format (symmetric for signed formats). NO: wrap around.
YES: clip each lane to its output format (symmetric for signed). NO: wrap. Default YES.Default: YES
Options: NO YES
Accuracy, per lane
Each lane carries the scalar twin’s bound: the variance numerator N*S2 - S1^2 is an exact integer, the root truncates below SQG = 4 guard bits and the final requantisation rounds, so each lane’s SEM is within 1 LSB of the exactly-rounded value - and each lane is bit-identical to the scalar twin run on the same lane stream. Like the scalar twin, GIVE IT FRACTIONAL BITS: a standard error is small by construction (the default lane format is Q8.8).
Timing: the TM latency contract
OUT_DV pulses ONCE per block, L clocks after the clock on which the
N-th sample was accepted, where
$$ L = 1 + \mathrm{TM},(\mathrm{IN_SW} + \mathrm{EXP} + \mathrm{RTW} + 3) $$
(BSET_TAIL in the core, with RTW = SEM lane width + 5). ONE shared
shift-add engine squares each lane’s |S1| and ONE shared digit-recurrence
root serves the lanes in turn, so the tail DEPENDS ON THE RUNTIME EXPONENT,
like the scalar twin. This is the family’s lane-multiplexed rule
$L_{tm}(e) = \mathrm{TM},(L_{scalar}(e)-1)+1$. At the defaults (16 bit
signed input, Q8.8 SEM, TM 4, EXP 10) that is 1 + 4*50 = 201 clocks, and
the minimum usable EXP is 8.
The family drop rule applies with the TM tail: the tail of one block must
finish before the NEXT block completes, $2^{{\mathrm{{EXP}}}} \ge L$, or the
completing block’s result is silently DROPPED (no OUT_DV, accumulators
unaffected, no error pin). The TM tail makes the minimum usable EXP larger
than the scalar twin’s - the compiler prints both the worst-case tail and
the minimum EXP in the compilation log, and the property window refuses a
configuration whose minimum exceeds Max Block Exponent.
Knowing where the block is: BUSY, INTEGRATING and SAMPLE_COUNT
Identical to the scalar family, and SHARED by all lanes: INTEGRATING is
high exactly while the block is accumulating (it dips one clock per block
boundary on a continuous stream), BUSY also covers the (TM-long) tail and
its last high clock IS the OUT_DV pulse, SAMPLE_COUNT reads 1 after the
first accepted sample and N after the N-th.
SAMPLE_COUNTis NOT cleared at the block end - it holds N until the NEXT block’s first accepted sample takes it back to 1. On a CONTINUOUS stream that happens DURING the serial tail (which on a TM block is TM Factor times longer), so atOUT_DVit reads how far into the next block the input has got, NOT N. The clock that always reads N is the oneINTEGRATINGfalls on - latch it there.
Verification
The core is regression tested by a host-side csim harness
(tb/block-ops-tm/run_tb_tm.ps1) with per-lane goldens computed by
gen_golden_tm.py in exact rational arithmetic ON EACH LANE’S STREAM ALONE
(lanes deliberately carry different signals - the generator refuses
identical lanes), plus the strongest available lane-independence check:
after every run, the SCALAR twin is replayed on each lane’s stream by
itself and lane k of every TM result must match it BIT FOR BIT. The
status waveform is checked clock by clock, every packed output is checked
to move only on OUT_DV, and the TM-specific mutant classes (shared
accumulator, lane swaps, wrong-lane tail reads, early commit, drop rule)
are killed. What no host harness can prove - that Vitis accepts and
schedules the core at II=1 - is stated in AGENT/block_ops.log, not
silently implied.