mem_block_intf — cycle-by-cycle animation

Speed slow Cycle 0 / 0

Press Step or Play to begin.

Image view — 3 stripes processed in parallel

unit 0 stripe (rows 0–3) unit 1 stripe (rows 4–7) unit 2 stripe (rows 8–11) shared search window

Toy parameters: BLOCK_SIZE = 3, MAX_DISP = 5, HALF_FRAME_WIDTH = 18, FRAME_HEIGHT = 12, NUM_SAD_UNITS = 3, STRIPE_HEIGHT = 4. The DE1-SoC build uses BLOCK_SIZE = 9, MAX_DISP = 85, and FRAME_HEIGHT = 200 rows split across multiple parallel SAD units — same algorithm, just more units and bigger blocks. The synthetic stereo pair plants a different ground-truth disparity per stripe (d=1, d=2, d=3) so each unit converges to a different best_disp.

FSM & shared signals

state	IDLE
reg_phase (y in stripe)	0
reg_col_x	0
reg_disp	0
mem_req	0
mem_bank	—
mem_col (shared)	—
slide_reference	0
slide_matching	0
sad_compare_en	0

Per-unit SAD scoreboard

All units fetch the same mem_col every cycle (port B fans the byte out to all rows in parallel). Each unit only reads its own BLOCK_SIZE rows and tracks its own best.

unit 0 · rows 0–3 · planted d = 1

SAD this candidate	—
best_sad	—
best_disp	—

unit 1 · rows 4–7 · planted d = 2

SAD this candidate	—
best_sad	—
best_disp	—

unit 2 · rows 8–11 · planted d = 3

SAD this candidate	—
best_sad	—
best_disp	—

Disparity stream (serialized)

disp_valid	0
emit_unit	—
disp_out_x	—
disp_out_y (= unit · STRIPE_HEIGHT + phase)	—
disp_out_value	—

How to read this

The image is divided into NUM_SAD_UNITS = 3 horizontal stripes of STRIPE_HEIGHT = 4 rows each. Every cycle, all three units operate on the same column at the same disparity step.
Each unit owns a BLOCK_SIZE × BLOCK_SIZE reference block at (col_x, g·STRIPE_HEIGHT + phase) on the left image, and a sliding match block at the same row offset on the right.
When mem_req fires, column_prefetch_parallel reads the requested column from every row of stereo_onchip_ram in one cycle (port B). Each unit indexes only its BLOCK_SIZE rows out of the returned bus.
The result is three independent best-SAD trackers running in lock-step. They tend to converge to different best_disp values because the underlying image content differs between stripes — in this synthetic demo we plant d = 1 / 2 / 3 respectively.
When disp exits its range, the FSM goes through INCR_X. It emits the three winners serially on the disparity stream — emit_unit cycles 0 → 1 → 2, one stream word per cycle (each acknowledged by disp_ack downstream).