Files
FPGA-Neural/hardware/v2/nms/sim/tb_nms_activation_candidates.v
T
micheleandClaude Sonnet 5 8e014d8d49 V2.0.0 hardware freeze - single SDRAM
FASE #1 hardware freeze for FPGA-Neural V2, N4/P8, single external
SDRAM (Alliance Memory AS4C4M16SA-6TIN) serving weights, activations,
and results through one physical sdram_controller.v instance. Removes
the PSRAM dependency (hardware/v1/rtl/psram_controller.v +
memory_interface.v) from the V2 physical path entirely -- V1 itself
remains fully unmodified, the golden reference.

New RTL: sdram_unified_backend.v (2-way W/AR arbitration over one
SDRAM controller, real per-byte DQM write masking added to
sdram_controller.v for correct single-byte result writes with no
read-modify-write), nms_neural_multiprocessor_sdram_unified.v (the
frozen top-level). Two real bugs found and fixed via full-system
testing before being accepted (ERR-0023): a deadlock and an off-by-one
data-shift bug in the new arbitration logic.

Real results: N=4 and N=2 D-Stress bit-exact (256/256 neurons), 40
real AUTO REFRESH events interleaved with zero corruption, real
Yosys+nextpnr-ecp5 synthesis/P&R for LFE5U-45F-8CABGA381 (149/245
TRELLIS_IO, a real 45-pin reduction from the prior dual-memory
design). Timing is MARGINAL (1/8 P&R seeds >=80MHz), reported honestly
rather than masked by the best seed.

Real, sourced ball-level pinout for the SDRAM bus + clk/rst (39/149
signals, P&R-verified) using the official Lattice ECP5U-45 pinout CSV
found on disk during this step's own pre-commit review -- corrects an
earlier draft that wrongly assumed no real pinout data was available.

Chip readiness: NO. Real, disclosed blockers remain (no physical host
interface exists yet -- the RTL's own reg_* ports are a 110-pin raw
test-harness bus; clock source/PLL decision; power/configuration
component selection) -- see hardware/v2/docs/{HARDWARE_FREEZE,
CHIP_READINESS,OPEN_ITEMS}.md for the complete, itemized status.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
2026-09-06 13:39:55 +02:00

149 lines
6.1 KiB
Verilog

// ============================================================
// NMS STEP4/5 -- correctness check for BOTH candidate activation
// memories (nms_activation_replicated.v, nms_activation_banked.v)
// before trusting any synthesis number from either. Fills a small
// known vector, then drives concurrent requests covering: pure
// broadcast (all slots want the identical tile), no-contention
// parallel access (all slots want different tiles mapping to
// different banks), and forced same-bank contention (two slots want
// different tiles that alias to the same bank) -- checks bit-exact
// data on every ack and that every request eventually gets acked
// (no starvation).
// ============================================================
`timescale 1ns/1ps
module tb_nms_activation_candidates;
parameter DATA_WIDTH = 8;
parameter P_IN = 8;
parameter N_SLOTS = 4;
parameter N_BANKS = 4;
parameter MAX_TILES = 8;
parameter TIW = $clog2(MAX_TILES);
reg clk = 0;
always #5 clk = ~clk;
reg rst;
reg fill_we;
reg [TIW-1:0] fill_tile_idx;
reg [DATA_WIDTH*P_IN-1:0] fill_data;
reg [N_SLOTS-1:0] req_valid;
reg [N_SLOTS*TIW-1:0] req_tile_idx_flat;
wire [N_SLOTS-1:0] ack_rep;
wire [N_SLOTS*DATA_WIDTH*P_IN-1:0] data_rep;
wire [N_SLOTS-1:0] ack_bank;
wire [N_SLOTS*DATA_WIDTH*P_IN-1:0] data_bank;
nms_activation_replicated #(
.DATA_WIDTH(DATA_WIDTH), .P_IN(P_IN), .N_SLOTS(N_SLOTS), .MAX_TILES(MAX_TILES)
) u_rep (
.clk(clk), .rst(rst),
.fill_we(fill_we), .fill_addr(fill_tile_idx), .fill_data(fill_data),
.rd_en(req_valid), .rd_addr_flat(req_tile_idx_flat), .rd_data_flat(data_rep)
);
// replicated candidate: 1-cycle read latency, no ack signal of its
// own -- model "ack" as rd_en registered (always accepted, private
// port), matching its own zero-contention design.
reg [N_SLOTS-1:0] ack_rep_reg;
always @(posedge clk) ack_rep_reg <= rst ? {N_SLOTS{1'b0}} : req_valid;
assign ack_rep = ack_rep_reg;
nms_activation_banked #(
.DATA_WIDTH(DATA_WIDTH), .P_IN(P_IN), .N_SLOTS(N_SLOTS), .N_BANKS(N_BANKS), .MAX_TILES(MAX_TILES)
) u_bank (
.clk(clk), .rst(rst),
.fill_we(fill_we), .fill_tile_idx(fill_tile_idx), .fill_data(fill_data),
.req_valid(req_valid), .req_tile_idx_flat(req_tile_idx_flat),
.ack(ack_bank), .rd_data_flat(data_bank)
);
integer errors;
integer t;
task fill_pattern;
integer k;
begin
for (k = 0; k < MAX_TILES; k = k + 1) begin
@(posedge clk);
fill_we = 1'b1;
fill_tile_idx = k[TIW-1:0];
fill_data = {P_IN{k[DATA_WIDTH-1:0]}};
end
@(posedge clk);
fill_we = 1'b0;
end
endtask
// drive one slot's request for `tidx` until it is acked (retrying
// each cycle if not), checking bit-exact data on both candidates
// when it lands. Runs slots in parallel via fork/join in the
// caller; this task itself only manages ONE slot's own state.
reg [N_SLOTS-1:0] want_valid;
reg [TIW-1:0] want_tile [0:N_SLOTS-1];
reg [N_SLOTS-1:0] got_rep, got_bank;
task run_case;
input [TIW-1:0] t0, t1, t2, t3;
integer cyc, s;
begin
want_tile[0] = t0; want_tile[1] = t1; want_tile[2] = t2; want_tile[3] = t3;
got_rep = 0; got_bank = 0;
for (s = 0; s < N_SLOTS; s = s + 1)
req_tile_idx_flat[s*TIW +: TIW] = want_tile[s];
req_valid = {N_SLOTS{1'b1}};
for (cyc = 0; cyc < 40 && (got_rep != {N_SLOTS{1'b1}} || got_bank != {N_SLOTS{1'b1}}); cyc = cyc + 1) begin
@(posedge clk);
#1;
for (s = 0; s < N_SLOTS; s = s + 1) begin
if (ack_rep[s] && !got_rep[s]) begin
got_rep[s] = 1'b1;
if (data_rep[s*DATA_WIDTH*P_IN +: DATA_WIDTH] !== want_tile[s][DATA_WIDTH-1:0]) begin
$display("FAIL replicated slot=%0d expected=%0d got=%0d", s, want_tile[s], data_rep[s*DATA_WIDTH*P_IN +: DATA_WIDTH]);
errors = errors + 1;
end
end
if (ack_bank[s] && !got_bank[s]) begin
got_bank[s] = 1'b1;
if (data_bank[s*DATA_WIDTH*P_IN +: DATA_WIDTH] !== want_tile[s][DATA_WIDTH-1:0]) begin
$display("FAIL banked slot=%0d expected=%0d got=%0d", s, want_tile[s], data_bank[s*DATA_WIDTH*P_IN +: DATA_WIDTH]);
errors = errors + 1;
end
end
// stop re-requesting a slot once it's been served by BOTH
if (got_rep[s] && got_bank[s]) req_valid[s] = 1'b0;
end
end
if (got_rep != {N_SLOTS{1'b1}} || got_bank != {N_SLOTS{1'b1}}) begin
$display("FAIL starvation: got_rep=%b got_bank=%b", got_rep, got_bank);
errors = errors + 1;
end
req_valid = {N_SLOTS{1'b0}};
@(posedge clk);
end
endtask
initial begin
errors = 0;
rst = 1'b1; fill_we = 1'b0; req_valid = 0; req_tile_idx_flat = 0;
repeat (3) @(posedge clk);
rst = 1'b0;
fill_pattern();
// Test 1: pure broadcast (all 4 slots want tile 0)
run_case(0, 0, 0, 0);
// Test 2: no contention (different tiles, different banks: 0,1,2,3)
run_case(0, 1, 2, 3);
// Test 3: forced same-bank contention (tile 0 and tile 4 both %4==0,
// tile 1 and tile 5 both %4==1)
run_case(0, 4, 1, 5);
// Test 4: mixed broadcast + contention (slots 0,1 want tile 2 (broadcast),
// slots 2,3 want DIFFERENT tiles 6 and 3, tile 6%4==2 collides with... )
run_case(2, 2, 6, 3);
if (errors == 0) $display("ALL TESTS PASSED (nms_activation_replicated + nms_activation_banked, N_SLOTS=%0d N_BANKS=%0d)", N_SLOTS, N_BANKS);
else $display("%0d FAILURES", errors);
$finish;
end
endmodule