DSP - BLOCK SLOPE STATS TM
TM (time multiplexed) twin of Block Slope Stats: TM Factor (2..16) independent channels packed on one wide bus, each computing the MEAN and the STANDARD DEVIATION of the FIRST DIFFERENCE d[i] = x[i] - x[i-1] - average ramp rate and sample-to-sample roughness - 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 differencers and 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:
$$ d_i[k] = x_i[k] - x_i[k-1], \qquad \mathrm{SLOPE_MEAN}_i = \frac{1}{N}\sum_k d_i[k], \qquad \mathrm{SLOPE_SIGMA}_i = \sqrt{\overline{d_i^2} - \overline{d_i}^2} $$
This is the time multiplexed (TM) twin of the scalar Block Slope Stats block: TM Factor independent channels packed on one wide bus, each computing the mean and standard deviation of the first difference of its own block of N consecutive samples. Nothing is shared between the channels except the frame.
Per lane the first difference is a one-tap high pass: SLOPE_MEAN is that lane’s average RAMP RATE (drift, a leaking baseline, a detector warming up) and SLOPE_SIGMA is that lane’s sample-to-sample ROUGHNESS, blind to slow wander - the better noise estimator on a drifting signal (for white noise of standard deviation s it converges to s*sqrt(2)).
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.- THE PREVIOUS-SAMPLE MEMORY IS PER LANE: a lane’s first difference is
against the last ACCEPTED sample of that SAME lane, and - exactly like
the scalar twin - it crosses block boundaries deliberately, so there are
exactly N differences per lane per block; only
RESETclears it. The scalar caveat applies TO EVERY LANE: afterRESETthe held sample is 0, so the FIRST block is contaminated by each lane’s x[0] - discard the first block. - 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 differencer and one
accumulator pair per lane - plus one real d*d multiplier per lane when
SLOPE_SIGMAis enabled. - The serial post-processing is ONE shared engine serving the lanes one
after the other (the |Sd| square and the digit-recurrence root, shared,
not TM Factor of each); ONE
OUT_DVper block, after the LAST lane’s tail completes, with all output lanes staged and committed together on that clock, so every packed output moves on theOUT_DVclock 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; the first difference is signed either way.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
YES: the the SLOPE_MEAN output (per lane, packed) pin is present. NO: the pin AND all of its logic are removed BEFORE synthesis, so nothing is paid for it.
YES: the packed SLOPE_MEAN bus exists. Default YES.Default: YES
Options: NO YES
YES: the the SLOPE_SIGMA output (per lane, packed) pin is present. NO: the pin AND all of its logic are removed BEFORE synthesis, so nothing is paid for it.
YES: the packed SLOPE_SIGMA bus exists - and with it the per-lane d*d multipliers and the exponent-dependent serial tail. NO: the tail collapses to 1 + TM Factor clocks. Default YES.Default: YES
Options: NO YES
Number of INTEGER bits of the SLOPE_MEAN output (per lane) (the sign, when present, uses one of them).
Integer bits of ONE LANE of SLOPE_MEAN. 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 SLOPE_MEAN 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 0 - and 0 is usually the wrong choice here: a drift of less than one input LSB per sample is exactly what this output exists to measure. Give it fractional bits.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 SLOPE_MEAN output (per lane) is signed (two’s complement) or unsigned.
SIGNED or UNSIGNED lanes. Default SIGNED - a ramp goes both ways.Default: SIGNED
Options: UNSIGNED SIGNED
Number of INTEGER bits of the SLOPE_SIGMA output (per lane) (the sign, when present, uses one of them).
Integer bits of ONE LANE of SLOPE_SIGMA. 1..64, default 16. Together with the fractional bits it sizes the shared root engine, and therefore the per-lane tail.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 SLOPE_SIGMA 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 0. Each bit of lane width costs one root clock per lane.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 SLOPE_SIGMA output (per lane) is signed (two’s complement) or unsigned.
SIGNED or UNSIGNED lanes. Default UNSIGNED (a standard deviation 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
Like the scalar twin, per lane: SLOPE_MEAN is BIT EXACT (an exact sum of differences, divided by N with a shift; the only error is the single final requantisation, tolerance 0 in the harness) and SLOPE_SIGMA is within 1 LSB (the truncated digit-recurrence root is the only inexact step). Each lane is bit-identical to the scalar twin run on the same lane stream.
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
$$ L = 1 + \mathrm{TM} \quad (\mathrm{SLOPE_MEAN\ only}), \qquad L = 1 + \mathrm{TM},(1 + \mathrm{DW} + \mathrm{EXP} + \mathrm{RTW} + 2) \quad (\mathrm{with\ SLOPE_SIGMA}) $$
(BSLT_TAIL in the core, with DW = IN_SW + 1 and RTW = SLOPE_SIGMA lane
width + 5). The mean alone needs no serial arithmetic - one requantise
clock per lane; the sigma adds a serial square of the difference sum and a
serial root, per lane, out of ONE shared engine pair, and then the tail
DEPENDS ON THE RUNTIME EXPONENT. 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, Q16.0 sigma, TM 4, EXP 10, both outputs on) that is
1 + 4*51 = 205 clocks, and the minimum usable EXP is 8 (3 with the sigma
disabled).
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
previous-sample register, 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.