Xilinx
TM
HLS
Block Preview

Introduction

Per lane, over each block of N samples:

Per lane: a Schmitt trigger with upper level = THRESHOLD and lower level = THRESHOLD - |HYSTERESIS| counts ONE peak per excursion, on its rising edge; the lane’s signal must fall back through the lower level before another can be counted. PEAK_RATE = peaks / N is an exact shift. Comparisons are STRICT.

This is the time multiplexed (TM) twin of the scalar Block Peak Count block: TM Factor independent channels packed on one wide bus, each counting the level excursions of its own block of N consecutive samples. Nothing is shared between the channels except the frame and the two level pins.

The TM contract

  • IN is 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 Schmitt state, its own counter, its own result.
  • THRESHOLD and HYSTERESIS are ONE SCALAR BROADCAST PIN EACH (in the INPUT lane format, latched per block like EXP): one reference pair for every lane, so a block’s counts are never a mixture of two thresholds - in any lane. Lanes needing different thresholds need scalar blocks.
  • THE SCHMITT STATE IS PER LANE: a lane’s detector arms and re-arms on ITS OWN samples only, and - exactly like the scalar twin - it CARRIES ACROSS BLOCK BOUNDARIES per lane, so a pulse that straddles a boundary is counted once, not twice; only RESET clears it.
  • ONE shared EXP pin, ONE IN_DV, one frame: all lanes start and end their blocks on the same accepted clocks. BUSY, INTEGRATING and SAMPLE_COUNT therefore stay scalar - SAMPLE_COUNT counts per-lane samples, which are identical in every lane by construction.
  • Accumulation runs at II=1 on the packed bus - per lane it is two comparisons and a counter, no multiplier anywhere. The serial post-processing is ONE shared engine serving the lanes one after the other; ONE OUT_DV per block, after the LAST lane finishes, with all output lanes staged and committed together on that clock, so every packed output moves on the OUT_DV 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, PEAK_RATE = peaks / N is an exact shift in every lane.

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 detector state (unavoidable either way) but shares ONE frame, ONE control FSM and ONE serial tail across all lanes. The price is tail latency (TM Factor times the scalar tail - trivial here) and the coupling of the frame AND of the levels: all lanes share the same block length, the same sample cadence and the same THRESHOLD / HYSTERESIS pair. Channels that need different block sizes or different thresholds need scalar blocks.

Pin Description

IN Input (IN_BitsInt + IN_BitsFract) * TMFactor bit TM
Packed TM sample bus: TM Factor lanes of input width bits, lane 0 in the LOW bits. Each lane is an independent channel, compared and counted only on the clocks where IN_DV is high.
Default: Must be connected
IN_DV Input 1 bit BIT
Per-sample qualifier for ALL lanes at once, active high, the only qualifier this block has (no CE pin, by family rule - gate this instead). Unconnected defaults to ‘1’.
EXP Input 6 bit BIT VECTOR
Block size exponent, shared by every lane: each lane’s block is 2^EXP samples of THAT lane. 6 bits, clamped to Max Block Exponent, latched on the first accepted sample of a block. Unconnected defaults to 10 (N = 1024).
THRESHOLD Input IN_BitsInt + IN_BitsFract bit BIT VECTOR
ONE broadcast upper level for every lane, INPUT lane format, latched per block. A lane counts a peak when its sample goes strictly above this. Unconnected ties to 0.
HYSTERESIS Input IN_BitsInt + IN_BitsFract bit BIT VECTOR
ONE broadcast hysteresis for every lane, INPUT lane format, latched per block; its MAGNITUDE is used. The lower (re-arm) level is THRESHOLD - |HYSTERESIS|: a lane must fall strictly below it before another peak can be counted. 0 degrades to a bare comparator with strict above / below. Unconnected ties to 0.
PEAK_COUNT Output (MaxBlockExponent + 1) * TMFactor bit TM
Packed per-lane peak counts, lane 0 in the LOW bits: excursions counted in each lane’s block. Integer, MaxBlockExponent + 1 bits per lane, cannot overflow. Present when Enable PEAK_COUNT = YES.
PEAK_RATE Output (PEAK_RATE_BitsInt + PEAK_RATE_BitsFract) * TMFactor bit TM
Packed per-lane peak rates, lane 0 in the LOW bits: each lane’s count divided by N (an exact shift). Default Q1.15: a peak on every sample would read 1.0. Present when Enable PEAK_RATE = YES.
OUT_DV Output 1 bit BIT
ONE one-clock pulse per block, L clocks after the N-th sample was accepted - after the LAST lane’s tail completed. Every packed output updates on this clock and on no other.
CLK 1 bit
Clock.
RESET 1 bit
Synchronous reset: clears every lane (Schmitt state included), the shared frame and the tail.
BUSY 1 bit
Shared by all lanes: high from the first sample of a block until the LAST lane’s result is out (it covers the whole lane-multiplexed tail); its last high clock is the OUT_DV pulse. Present only when Enable BUSY = YES.
INTEGRATING 1 bit
Shared by all lanes: high only while the block is accumulating. The clock it falls on is the one where SAMPLE_COUNT always reads N. Present only when Enable INTEGRATING = YES.
SAMPLE_COUNT 1 bit
Shared by all lanes (they advance in lockstep): samples accumulated so far in the current block, 1 after the first, N after the N-th. NOT cleared at the block end; on a continuous stream it has already restarted by OUT_DV, so latch it when INTEGRATING falls. 32 bits. Present only when Enable SAMPLE_COUNT = YES.

Properties

Property window

IN Integer Bits IN_BitsInt

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. THRESHOLD and HYSTERESIS share this format.

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

IN Fractional Bits IN_BitsFract

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

IN Sign IN_Sign

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 and to both level pins.

Default: SIGNED

Options: UNSIGNED SIGNED

TM Factor TMFactor

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

Max Block Exponent MaxBlockExponent

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 counters are sized for; the EXP input is clamped to it at run time. It sets each PEAK_COUNT lane’s width to MaxBlockExponent + 1 bits. 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

Enable PEAK_COUNT EnablePeakCount

YES: the the PEAK_COUNT 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 PEAK_COUNT bus exists. Default YES.

Default: YES

Options: NO YES

Enable PEAK_RATE EnablePeakRate

YES: the the PEAK_RATE 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 PEAK_RATE bus exists. Default YES.

Default: YES

Options: NO YES

PEAK_RATE Integer Bits PEAK_RATE_BitsInt

Number of INTEGER bits of the PEAK_RATE output (per lane) (the sign, when present, uses one of them).

Integer bits of ONE LANE of PEAK_RATE. Default 1.

Default: 1

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

PEAK_RATE Fractional Bits PEAK_RATE_BitsFract

Number of FRACTIONAL bits of the PEAK_RATE 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. Default 15.

Default: 15

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

PEAK_RATE Sign PEAK_RATE_Sign

Select whether the PEAK_RATE output (per lane) is signed (two’s complement) or unsigned.

Default UNSIGNED (a rate is non negative by construction).

Default: UNSIGNED

Options: UNSIGNED SIGNED

Enable BUSY EnableBusy

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

Enable INTEGRATING EnableIntegrating

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

Enable SAMPLE_COUNT EnableSampleCount

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

Rounding Rounding

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 PEAK_RATE requantisation (PEAK_COUNT is never requantised). TRUNCATE: drop the bits. Default ROUND.

Default: ROUND

Options: TRUNCATE ROUND

Saturation EnableSaturation

YES: clip to the largest representable value of each output format (symmetric for signed formats). NO: wrap around.

YES: clip each PEAK_RATE lane to its format (PEAK_COUNT cannot overflow and ignores this). NO: wrap. Default YES.

Default: YES

Options: NO YES

Accuracy, per lane

Each lane is BIT EXACT, like the scalar twin: PEAK_COUNT is a plain per-lane counter, Max Block Exponent + 1 bits per lane, which CANNOT overflow (a lane’s count can never exceed N) and is never requantised - Rounding and Saturation do not touch it. PEAK_RATE is exact up to the single final requantisation per lane: PEAK_COUNT * 2^-EXP produced by a shift of an exact integer. The harness demands tolerance 0 per lane.

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

$$ L = 1 + \mathrm{TM} $$

(BPCT_TAIL in the core). Everything is evaluated inside the window, so the shared tail is one requantise clock per lane plus the registered OUT_DV - the family’s lane-multiplexed rule $L_{tm}(e) = \mathrm{TM},(L_{scalar}(e)-1)+1$ with $L_{scalar} = 2$. It is a CONSTANT: 5 clocks at the default TM Factor of 4, 17 at TM 16.

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 (EXP >= 3 at TM 4, EXP >= 5 at TM 16, against the scalar’s 1) - 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_COUNT 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 (which on a TM block is TM Factor times longer), so at OUT_DV it reads how far into the next block the input has got, NOT N. The clock that always reads N is the one INTEGRATING falls 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 detector state, 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.