michele/FPGA-Neural
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

exp: real end-to-end weight-reuse integration with neural_processor.v (EXP-0058)

Browse Source 
New tb_neural_processor_layer_reuse.v wires the real SDRAM controller,
layer_prefetch_ctrl.v and layer_weight_buffer.v into a real
neural_processor.v compute engine: one resident filter is fetched once
and reused across 8 independent jobs per layer, verified bit-exact
against an independent golden dot-product model (32/32 PASS).

Also found and fixed a real testbench-vs-DUT scheduling race present in
tb_layer_prefetch_ctrl.v (and hardened in the new file): clearing a
one-cycle control pulse on the very next clock edge lands the clear in
the same active-region pass as the edge a receiving module's own
synchronous logic reads it at, so the pulse can be silently missed
depending on implementation-defined process ordering. This had been
silently preventing tb_layer_prefetch_ctrl.v's own claimed 8192/8192
result from ever actually being observed; fixed by holding the pulse
past the edge with a real time delay before clearing, and the
8192/8192 result is now genuinely reproducible (5/5 consecutive runs).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
This commit is contained in:
michele
2026-09-16 14:26:09 +02:00
co-authored by Claude Sonnet 5
parent ca94083366
commit 80c89fa10d
3 changed files with 488 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hardware/v2/sim/tb_layer_prefetch_ctrl.v
+42 -10
View File
@@ -67,12 +67,6 @@ module tb;
reg [16*BURST_LEN-1:0] wpre_wdata;
reg pre_active;
assign ctrl_req = pre_active ? wpre_req : pf_ctrl_req;
assign ctrl_wr = pre_active ? wpre_wr : pf_ctrl_wr;
assign ctrl_addr = pre_active ? wpre_addr : pf_ctrl_addr;
assign ctrl_wdata = pre_active ? wpre_wdata : pf_ctrl_wdata;
assign ctrl_wmask = pre_active ? {(2*BURST_LEN){1'b0}} : pf_ctrl_wmask;
task automatic sdram_write_burst(input [ADDR_WIDTH-1:0] word_addr, input [16*BURST_LEN-1:0] data);
begin
@(posedge clk); while (ctrl_busy) @(posedge clk);
@@ -102,6 +96,18 @@ module tb;
wire [16*BURST_LEN-1:0] pf_ctrl_wdata;
wire [2*BURST_LEN-1:0] pf_ctrl_wmask;
// moved here (below the pf_ctrl_* declarations above) -- iverilog's
// current elaboration requires a continuous assign's RHS names to be
// declared earlier in the module than the assign itself, which the
// original position (right after wpre_*/pre_active, before pf_ctrl_*
// existed textually) violated; found while re-verifying this file for
// EXP-0058, see this file's own note by pre_active's assignment below.
assign ctrl_req = pre_active ? wpre_req : pf_ctrl_req;
assign ctrl_wr = pre_active ? wpre_wr : pf_ctrl_wr;
assign ctrl_addr = pre_active ? wpre_addr : pf_ctrl_addr;
assign ctrl_wdata = pre_active ? wpre_wdata : pf_ctrl_wdata;
assign ctrl_wmask = pre_active ? {(2*BURST_LEN){1'b0}} : pf_ctrl_wmask;
reg pf_start;
reg [ADDR_WIDTH-1:0] pf_layer_base;
wire pf_busy, pf_done;
@@ -148,6 +154,18 @@ module tb;
$display("=== preload SDRAM with %0d distinct layer patterns ===", L);
preload_sdram_layers;
@(posedge clk); // ERR-0001 workaround: sync before the first blocking
// assignment following a time-consuming task call --
// without this, pre_active's switchover was not
// reliably visible to the ctrl_* mux at the next
// clock edge, so layer_prefetch_ctrl.v's own
// ctrl_req never actually reached the real SDRAM
// controller and this whole test hung on "while
// (!pf_done)" forever instead of ever producing a
// result (found while integrating this module into
// tb_neural_processor_layer_reuse.v, EXP-0058 --
// this file's own real-hardware run had never
// actually completed before that).
pre_active = 1'b0; // hand control to layer_prefetch_ctrl.v
// NOTE: sequential (no prefetch/consume overlap) -- this test
@@ -162,9 +180,23 @@ module tb;
// information.
$display("=== real-RTL prefetch + reuse, %0d layers, sequential (correctness only) ===", L);
t0 = cyc;
pf_layer_base = 0; pf_start = 1'b1; @(posedge clk); pf_start = 1'b0;
pf_layer_base = 0; pf_start = 1'b1; @(posedge clk); #1; pf_start = 1'b0;
while (!pf_done) @(posedge clk);
consume_done = 1'b1; @(posedge clk); consume_done = 1'b0; // initial swap
// Clock-edge-adjacent pulse idiom, hardened: setting a pulse then
// clearing it on the VERY NEXT @(posedge clk) puts the clear in the
// SAME active-region pass as the edge where a receiving module's own
// synchronous always block reads it -- their relative execution
// order at that shared edge is implementation-defined (Icarus does
// not guarantee testbench-thread-vs-DUT-always-block ordering), so
// the clear can occasionally run before the DUT's read and the pulse
// is silently missed (found via direct $strobe tracing while
// integrating this module for EXP-0058 -- layer_weight_buffer.v's
// own consume_done_latched stayed 0 even though this exact sequence
// visibly drove consume_done=1 for a full clock period). Fixed by
// holding the pulse past the edge with a real time delay (#1) before
// clearing, so the clear unambiguously lands in a later time step
// than every process that reacted to the edge.
consume_done = 1'b1; @(posedge clk); #1; consume_done = 1'b0; // initial swap
@(posedge clk); #1;
for (li_i = 0; li_i < L; li_i = li_i + 1) begin
@@ -182,11 +214,11 @@ module tb;
@(posedge clk);
end
end
consume_done = 1'b1; @(posedge clk); consume_done = 1'b0;
consume_done = 1'b1; @(posedge clk); #1; consume_done = 1'b0; // same pulse-hardening as the initial swap above
if (li_i+1 < L) begin
pf_layer_base = (li_i+1)*WORDS_PER_LAYER;
pf_start = 1'b1; @(posedge clk); pf_start = 1'b0;
pf_start = 1'b1; @(posedge clk); #1; pf_start = 1'b0;
while (!pf_done) @(posedge clk);
end
@(posedge clk); #1; // let the swap settle before the next iteration reads
hardware/v2/sim/tb_neural_processor_layer_reuse.v
+358
View File
@@ -0,0 +1,358 @@
`timescale 1ns/1ps
// ============================================================
// EXP-0058 -- first real end-to-end integration of the EXP-0057
// weight-reuse building blocks with the real M1 compute engine
// (neural_processor.v). Wires together, ALL real RTL except the
// per-tile weight-byte gather (see note below):
//
// sdram_controller_openrow.v + sdram_model.v (real DDR-less SDR SDRAM path)
// -> layer_prefetch_ctrl.v (real RTL, EXP-0057b)
// -> layer_weight_buffer.v (real RTL, double-buffered, EXP-0057)
// -> [testbench byte-gather, see note]
// -> neural_processor.v (real RTL, M1 compute engine)
//
// One "layer" = one resident filter (N_INPUTS=128 taps, 16 P_IN=8
// tiles) fetched ONCE from SDRAM into layer_weight_buffer.v, then
// REUSED across M independent "positions" (jobs) -- exactly modeling
// a real convolution filter held stationary while it slides across M
// different input windows, which is the whole point of EXP-0057's
// architecture. Input data for each position is synthetic (formula-
// generated, not fetched from SDRAM -- representing the activation/
// sliding-window path, which is a separate, already-existing memory
// path not the subject of this test) but deterministic and combined
// with an INDEPENDENT golden dot-product+bias+ReLU model (same
// "third oracle" style as tb_neural_processor.v's own expect_relu,
// duplicated here rather than shared, per that file's own stated
// convention).
//
// NOTE on the byte-gather step: neural_processor.v consumes one
// P_IN=8-wide (64-bit) weight_data tile per handshake cycle, but
// layer_weight_buffer.v is byte-wide (one address = one byte, already
// verified in isolation, EXP-0057). Assembling 8 sequential
// byte-wide reads into one 64-bit tile bus is done here by the
// testbench driver task. This gather step is NOT synthesizable RTL
// yet -- a real "tile gather adapter" (8:1 byte-to-tile packer) would
// be the natural next M4 Memory Manager deliverable if this
// architecture is adopted, deliberately out of scope here: this
// test's purpose is to verify DATA correctness of the weight-reuse
// path feeding the real compute engine, not to deliver the final
// gather RTL.
//
// Scope: correctness only (sequential, no prefetch/consume overlap
// across layers -- the double-buffered PERFORMANCE benefit was
// already measured in isolation, 7.16x, tb_layer_reuse_vs_zero_reuse.v,
// EXP-0057, not re-derived here).
// ============================================================
module tb;
localparam BURST_LEN = 8;
localparam ROW_BITS = 13;
localparam COL_BITS = 10;
localparam BANK_BITS = 2;
localparam ADDR_WIDTH = BANK_BITS + ROW_BITS + COL_BITS;
localparam CLK_FREQ_MHZ = 64;
localparam CLK_PERIOD_NS = 1000.0/CLK_FREQ_MHZ;
localparam DATA_WIDTH = 8;
localparam P_IN = 8;
localparam ACC_WIDTH = 32;
localparam N_INPUTS = 128; // one resident filter = 128 taps
localparam N_TILES = N_INPUTS/P_IN; // 16
localparam LAYER_BYTES = N_INPUTS; // 1 byte/tap, DATA_WIDTH=8
localparam L = 4; // layers (resident filters)
localparam M = 8; // reuse positions per layer
localparam WORDS_PER_LAYER = LAYER_BYTES/2;
localparam ACT_RELU = 2'd1;
reg clk = 0;
always #(CLK_PERIOD_NS/2.0) clk = ~clk;
reg rst;
integer cyc;
always @(posedge clk) if (!rst) cyc <= cyc + 1;
// ---- real SDRAM controller + model ----
wire ctrl_req, ctrl_wr;
wire [ADDR_WIDTH-1:0] ctrl_addr;
wire [16*BURST_LEN-1:0] ctrl_wdata;
wire [2*BURST_LEN-1:0] ctrl_wmask;
wire [16*BURST_LEN-1:0] ctrl_rdata;
wire ctrl_ready, ctrl_busy;
wire cke, cs_n, ras_n, cas_n, we_n;
wire [BANK_BITS-1:0] ba;
wire [ROW_BITS-1:0] a;
wire [15:0] dq;
wire [1:0] dqm;
sdram_controller_openrow #(
.CLK_FREQ_MHZ(CLK_FREQ_MHZ), .BURST_LEN(BURST_LEN),
.ROW_BITS(ROW_BITS), .COL_BITS(COL_BITS), .BANK_BITS(BANK_BITS)
) u_ctrl (
.clk(clk), .rst(rst),
.req(ctrl_req), .wr(ctrl_wr), .addr(ctrl_addr), .wdata(ctrl_wdata), .wmask(ctrl_wmask),
.rdata(ctrl_rdata), .ready(ctrl_ready), .busy(ctrl_busy),
.sdram_cke(cke), .sdram_cs_n(cs_n), .sdram_ras_n(ras_n), .sdram_cas_n(cas_n), .sdram_we_n(we_n),
.sdram_ba(ba), .sdram_a(a), .sdram_dq(dq), .sdram_dqm(dqm)
);
sdram_model #(
.CLK_FREQ_MHZ(CLK_FREQ_MHZ), .ROW_BITS(ROW_BITS), .COL_BITS(COL_BITS), .BANK_BITS(BANK_BITS)
) u_mem (
.clk(clk), .cke(cke), .cs_n(cs_n), .ras_n(ras_n), .cas_n(cas_n), .we_n(we_n),
.ba(ba), .a(a), .dq(dq), .dqm(dqm)
);
// separate write-capable path to preload SDRAM with the L filters'
// weight bytes (write and prefetch never run concurrently here)
reg wpre_req, wpre_wr;
reg [ADDR_WIDTH-1:0] wpre_addr;
reg [16*BURST_LEN-1:0] wpre_wdata;
reg pre_active;
wire pf_ctrl_req, pf_ctrl_wr;
wire [ADDR_WIDTH-1:0] pf_ctrl_addr;
wire [16*BURST_LEN-1:0] pf_ctrl_wdata;
wire [2*BURST_LEN-1:0] pf_ctrl_wmask;
assign ctrl_req = pre_active ? wpre_req : pf_ctrl_req;
assign ctrl_wr = pre_active ? wpre_wr : pf_ctrl_wr;
assign ctrl_addr = pre_active ? wpre_addr : pf_ctrl_addr;
assign ctrl_wdata = pre_active ? wpre_wdata : pf_ctrl_wdata;
assign ctrl_wmask = pre_active ? {(2*BURST_LEN){1'b0}} : pf_ctrl_wmask;
// ---- deterministic weight/input formulas (shared between SDRAM
// preload, the golden model, and -- for weights -- indirectly
// verified via the real buffer read-back) ----
function automatic signed [7:0] weight_byte(input integer li, input integer t);
weight_byte = $signed(8'((li*17 + t*29 + 13) & 8'hFF));
endfunction
function automatic signed [7:0] input_byte(input integer li, input integer pos, input integer t);
input_byte = $signed(8'((li*11 + pos*41 + t*7 + 3) & 8'hFF));
endfunction
task automatic sdram_write_burst(input [ADDR_WIDTH-1:0] word_addr, input [16*BURST_LEN-1:0] data);
begin
@(posedge clk); while (ctrl_busy) @(posedge clk);
wpre_req = 1'b1; wpre_wr = 1'b1; wpre_addr = word_addr; wpre_wdata = data;
@(posedge clk); wpre_req = 1'b0;
while (!ctrl_ready) @(posedge clk);
end
endtask
task automatic preload_sdram_layers;
integer li, bi, wb, tt;
reg [16*BURST_LEN-1:0] burst_data;
begin
for (li = 0; li < L; li = li + 1) begin
for (bi = 0; bi < (LAYER_BYTES/(2*BURST_LEN)); bi = bi + 1) begin
for (wb = 0; wb < BURST_LEN; wb = wb + 1) begin
// one burst = BURST_LEN words = 2*BURST_LEN bytes/taps
// (1 byte/tap); word wb holds taps [bi*2*BURST_LEN+wb*2]
// (low byte) and [...+wb*2+1] (high byte), matching
// layer_prefetch_ctrl.v's own drain_cnt byte order
// (drain_cnt = wb*2 -> low byte, wb*2+1 -> high byte).
tt = bi*(2*BURST_LEN) + wb*2;
burst_data[wb*16 +: 16] = {weight_byte(li, tt+1), weight_byte(li, tt)};
end
sdram_write_burst((li*WORDS_PER_LAYER + bi*BURST_LEN), burst_data);
end
end
end
endtask
// ---- layer_prefetch_ctrl.v (real RTL) ----
reg pf_start;
reg [ADDR_WIDTH-1:0] pf_layer_base;
wire pf_busy, pf_done;
wire pf_fill_we;
wire [$clog2(LAYER_BYTES)-1:0] pf_fill_addr;
wire [7:0] pf_fill_data;
layer_prefetch_ctrl #(
.DATA_WIDTH(8), .LAYER_BYTES(LAYER_BYTES), .BURST_LEN(BURST_LEN), .ADDR_WIDTH(ADDR_WIDTH)
) u_pf (
.clk(clk), .rst(rst),
.start(pf_start), .layer_base(pf_layer_base), .busy(pf_busy), .done(pf_done),
.fill_we(pf_fill_we), .fill_addr(pf_fill_addr), .fill_data(pf_fill_data),
.ctrl_req(pf_ctrl_req), .ctrl_wr(pf_ctrl_wr), .ctrl_addr(pf_ctrl_addr),
.ctrl_wdata(pf_ctrl_wdata), .ctrl_wmask(pf_ctrl_wmask),
.ctrl_rdata(ctrl_rdata), .ctrl_ready(ctrl_ready), .ctrl_busy(ctrl_busy)
);
// ---- layer_weight_buffer.v (real RTL) ----
reg [$clog2(LAYER_BYTES)-1:0] rd_addr;
wire [7:0] rd_data;
reg consume_done;
wire active_sel, swapped;
layer_weight_buffer #(.DATA_WIDTH(8), .LAYER_DEPTH(LAYER_BYTES)) u_lwb (
.clk(clk), .rst(rst),
.fill_we(pf_fill_we), .fill_addr(pf_fill_addr), .fill_data(pf_fill_data), .fill_done(pf_done),
.rd_addr(rd_addr), .rd_data(rd_data), .consume_done(consume_done),
.active_sel(active_sel), .swapped(swapped)
);
// ---- neural_processor.v (real RTL, M1 compute engine under test) ----
reg job_valid;
wire job_ready;
reg [15:0] job_node_id;
reg signed [DATA_WIDTH-1:0] job_bias;
reg [1:0] job_activation;
reg operand_valid;
wire operand_ready;
reg signed [DATA_WIDTH*P_IN-1:0] input_data, weight_data;
reg tile_last;
wire result_valid;
reg result_ready;
wire signed [DATA_WIDTH-1:0] result_data;
wire [15:0] result_node_id;
wire [3:0] np_state;
wire np_error;
neural_processor #(
.DATA_WIDTH(DATA_WIDTH), .P_IN(P_IN), .ACC_WIDTH(ACC_WIDTH)
) u_np (
.clk(clk), .rst(rst),
.job_valid(job_valid), .job_ready(job_ready),
.job_node_id(job_node_id), .job_bias(job_bias), .job_activation(job_activation),
.operand_valid(operand_valid), .operand_ready(operand_ready),
.input_data(input_data), .weight_data(weight_data), .tile_last(tile_last),
.result_valid(result_valid), .result_ready(result_ready),
.result_data(result_data), .result_node_id(result_node_id),
.np_state(np_state), .np_error(np_error)
);
integer errors, tests;
integer li_i, pos_i, t, k, tt;
reg [7:0] wbyte [0:P_IN-1];
integer acc_calc, s_calc;
reg signed [DATA_WIDTH-1:0] expected;
integer t0, total_cycles;
task automatic run_one_position(input integer li, input integer pos);
begin
@(posedge clk);
tests = tests + 1;
job_node_id = li[15:8]*8'(M) + pos[15:0];
job_bias = {DATA_WIDTH{1'b0}};
job_activation = ACT_RELU;
job_valid = 1;
while (!job_ready) @(posedge clk);
@(posedge clk); #1; // handshake edge, then hold one extra delta before
// clearing -- see the pulse-hardening note by
// consume_done below (same class of same-edge
// testbench-vs-DUT scheduling race)
job_valid = 0;
acc_calc = 0;
for (t = 0; t < N_TILES; t = t + 1) begin
// gather this tile's P_IN weight bytes from the real,
// resident (already-swapped-in) layer_weight_buffer.v
for (k = 0; k < P_IN; k = k + 1) begin
rd_addr = (t*P_IN + k);
#1;
wbyte[k] = rd_data;
end
input_data = {DATA_WIDTH*P_IN{1'b0}};
weight_data = {DATA_WIDTH*P_IN{1'b0}};
for (k = 0; k < P_IN; k = k + 1) begin
tt = t*P_IN + k;
input_data[k*DATA_WIDTH +: DATA_WIDTH] = input_byte(li, pos, tt);
weight_data[k*DATA_WIDTH +: DATA_WIDTH] = wbyte[k];
// real, resident buffer readback must match the
// formula used to preload SDRAM -- checked directly
// (not just indirectly via the final dot product),
// so a wrong buffer byte is caught even if the dot
// product would coincidentally still match.
if (wbyte[k] !== weight_byte(li, tt)) begin
$display("FAIL li=%0d pos=%0d t=%0d k=%0d: buffer weight byte %0d expected %0d",
li, pos, t, k, $signed(wbyte[k]), weight_byte(li, tt));
errors = errors + 1;
end
acc_calc = acc_calc + (input_byte(li, pos, tt) * weight_byte(li, tt));
end
tile_last = (t == N_TILES - 1);
operand_valid = 1;
while (!operand_ready) @(posedge clk);
@(posedge clk); #1;
end
operand_valid = 0;
tile_last = 0;
result_ready = 1;
while (!result_valid) @(posedge clk);
s_calc = acc_calc + 0; // job_bias == 0
if (s_calc <= 0) expected = {DATA_WIDTH{1'b0}};
else if (s_calc > 127) expected = 8'sd127;
else expected = s_calc[DATA_WIDTH-1:0];
if (result_data !== expected) begin
$display("FAIL li=%0d pos=%0d: result=%0d expected=%0d (acc=%0d)",
li, pos, $signed(result_data), $signed(expected), acc_calc);
errors = errors + 1;
end else begin
$display("PASS li=%0d pos=%0d: result=%0d (acc=%0d, weight-reuse path, real RTL)",
li, pos, $signed(result_data), acc_calc);
end
@(posedge clk);
while (!job_ready || np_state !== 4'd0) @(posedge clk);
end
endtask
initial begin
errors = 0; tests = 0; cyc = 0;
rst = 1; pre_active = 1'b1;
wpre_req = 0; wpre_wr = 0; wpre_addr = 0; wpre_wdata = 0;
pf_start = 0; pf_layer_base = 0; rd_addr = 0; consume_done = 0;
job_valid = 0; job_node_id = 0; job_bias = 0; job_activation = ACT_RELU;
operand_valid = 0; input_data = 0; weight_data = 0; tile_last = 0;
result_ready = 0;
repeat(5) @(posedge clk);
rst = 0;
@(posedge clk); while (ctrl_busy) @(posedge clk);
$display("=== preload SDRAM with %0d resident-filter weight sets (%0d taps each) ===", L, N_INPUTS);
preload_sdram_layers;
@(posedge clk); // ERR-0001 workaround: sync before the first blocking
// assignment following a time-consuming task call,
// otherwise it can be invisible to other modules at
// the next clock edge (see tb_neural_processor.v header)
pre_active = 1'b0; // hand control to layer_prefetch_ctrl.v
$display("=== EXP-0058: real RTL weight-reuse path -> real neural_processor.v, %0d layers x %0d reuse positions ===", L, M);
t0 = cyc;
for (li_i = 0; li_i < L; li_i = li_i + 1) begin
pf_layer_base = li_i * WORDS_PER_LAYER;
pf_start = 1'b1; @(posedge clk); #1; pf_start = 1'b0;
while (!pf_done) @(posedge clk);
#1;
// pulse-hardening: hold consume_done past its edge with a real
// time delay before clearing, rather than clearing on the very
// next @(posedge clk) -- otherwise the clear can land in the
// SAME active-region pass as the edge where layer_weight_
// buffer.v's own always block reads it, and their relative
// order is implementation-defined, so the pulse can be silently
// missed (found via direct $strobe tracing while debugging this
// exact sequence in tb_layer_prefetch_ctrl.v, EXP-0058 -- see
// that file's own longer note on this).
consume_done = 1'b1; @(posedge clk); #1; consume_done = 1'b0; // swap into active
@(posedge clk); #1;
for (pos_i = 0; pos_i < M; pos_i = pos_i + 1) begin
run_one_position(li_i, pos_i);
end
end
total_cycles = cyc - t0;
$display("=== RESULT: %0d/%0d PASS, %0d errors, %0d total cycles for %0d layers x %0d positions (EXP-0058 integration) ===",
tests-errors, tests, errors, total_cycles, L, M);
if (errors == 0) $display("ALL TESTS PASSED (tb_neural_processor_layer_reuse)");
$finish;
end
endmodule