STEP 4 - ROM-W' + MATRIX-VECTOR MULTIPLIER   (patch_embedding -> linear)
==========================================================================
(Step 3 = Input ROM + S2P wiring only; S2P of Step 2 already gives the 8 x 32 patches)

Guide architecture (pages 1 + 2)
--------------------------------
   ROM-W = 32 ROMs, each 1 x 128        ROM1 -> W1 ... ROM32 -> W32
   S2P1 (16) + S2P2 (16) + ROM-W (32) --> Matrix-vector multiplier

   Every clock j = 0..127:
      32 ROMs give W'(k, j) at address j      (k = 1..32)
      32 multiplies with the patch x1..x32     (x1..x16 = I from S2P1, x17..x32 = Q from S2P2)
      sum -> ONE value  e[j] = sum_k x_k * W'(k, j)
   128 clocks per patch, 8 patches -> 8 x 128 = 1024 values, one per clock.
   A hold register keeps the current patch for its 128 clocks.

ROM-W' = folded weights (mean/std removed from hardware)
   ROM 1..16  = W / std_I ,  ROM 17..32 = W / std_Q
   bias b' is NOT added here (it goes into ROM-P' in Step 6):
      this block gives E' = W' x ;   E' + b' = Patch_embedding_calculated.csv

Number formats
   input  x  : ap_fixed<25,1>  raw samples (from Step 2)
   weight W' : ap_fixed<18,7>  |W'| < 50   -> 25 x 18 bit = ONE DSP48E1 per multiply
   acc       : ap_fixed<40,8>  no rounding inside the sum of 32 products
   output e  : ap_fixed<25,8>  |e| < 3

Folder (keep it its own folder next to S1_input_rom_raw_src and S2_s2p)
   S4_patch_embed_src\
      hls_config.cfg                 top = patch_embed
      src\patch_embed.h / .cpp       ROM-W' (32 ROMs) + hold register + 32 MACs
      tb\patch_embed_tb.cpp          feeds 8 raw patches, checks 1024 outputs
      data\rom_w.dat                 32 ROMs x 128 folded weights W'
      data\patches_in.dat            8 x 32 raw patches (= Step 2 output)
      data\embed_ref.dat             8 x 128 expected E' = W' x
      data\patch_embedding_ref.dat   8 x 128 Patch_embedding_calculated.csv (model)
      data\bias_fold.dat             128 values b'

Checks
   A) HW vs E'                                  expect ~1.7e-05 (tol 1e-04)
   B) HW + b' vs Patch_embedding_calculated.csv expect ~1.7e-05 (tol 1e-04)

How to run
   1. Extract to  C:\Users\<you>\Documents\MRT_Blocks\MRT_Blocks\S4_patch_embed_src\
      (src, tb, data, hls_config.cfg directly inside; no double folder)
   2. Vitis: New Component -> HLS
        Name: s4_patch_embed       Location: ...\MRT_Blocks\MRT_Blocks  (no workspace switch)
        Config: Existing -> S4_patch_embed_src\hls_config.cfg
        Top: patch_embed           Part: xc7s50csga324-1  -> Finish
   3. C SIMULATION        expect "*** STEP 4 PATCH EMBEDDING TEST PASSED ***", 1024 outputs
   4. C SYNTHESIS         expect loop MVM: II = 1, trip count 1024, ~1030-1040 clocks,
                          DSP about 32, BRAM_18K about 32 (the 32 ROMs)
   5. C/RTL COSIMULATION  expect PASS

Waveform (xsim --gui in ...\s4_patch_embed\patch_embed\hls\sim\verilog, Tcl console):
   set g [add_wave_group "S4_per_clock"]; set p /apatb_patch_embed_top/AESL_inst_patch_embed; add_wave -into $g $p/ap_clk; add_wave -into $g $p/ap_start; add_wave -into $g $p/patch_in_read; add_wave -into $g $p/emb_out_write; add_wave -into $g -radix hex $p/emb_out_din; add_wave -into $g $p/ap_done
   expect: patch_in_read = 1 once every 128 clocks (8 times), emb_out_write = 1 every clock.
