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

v3: real Artix-7 compute core -- DSP48 packing verified, 2 real bugs found and fixed

Browse Source 
New hardware/v3/ (Artix-7 port, branch v3-artix7): the compute engine
that makes the 100x-vs-ESP32 target theoretically reachable on
XC7A100T's 240 DSP48E1 budget.

mac2_dsp_packed.v: packs 2 INT8 MACs sharing one resident weight into
a single DSP48-shaped 25x18 multiply, exploiting this project's own
weight-stationary reuse pattern (layer_weight_buffer.v, EXP-0057/0058)
where one weight is genuinely multiplied against many different
activations. Verified exhaustively: 16,777,216/16,777,216
(weight,x0,x1) combinations, 0 errors.

Two real bugs found and fixed during that verification (both purely
arithmetic/RTL, not toolchain-related):
1. An off-by-one in a declared wire width caused Verilog's part-select
   unsigned-by-default rule to corrupt sign extension on the upper
   product field -- ~50% of vectors failed.
2. After fixing (1), still ~50% failed: concatenating two independently
   sign-extended fields ({sext(x1,9), sext(x0,16)}) is NOT equivalent
   to the real arithmetic sum x1*2^16+x0 whenever the lower field is
   negative (its own two's-complement encoding "bleeds" an extra 2^16
   into the concatenated value). Fixed by building the packed operand
   with an explicit arithmetic shift-and-add instead of concatenation.

neural_processor_packed.v: full port of hardware/v2/rtl/
neural_processor.v's pipeline (same stage count/structure), doubled on
the accumulator/bias/activation/saturation side to process two
weight-reuse positions per weight-tile stream. Verified against TWO
real hardware/v2/rtl/neural_processor.v instances (job A / job B, same
shared weight, independent activations) -- 18/18 PASS, 0 errors,
covering the functional sweep, INT8 extremes, and back-to-back jobs.

A third real bug found in the process (in the new testbench, not the
RTL): clearing operand_valid/tile_last in the same simulation delta as
the handshake edge that should register tile_last=1 races against the
DUTs' own FSM evaluation of that same edge -- the same pulse-clearing
race class found three times already today in hardware/v2/sim (EXP-0058
and its follow-up commits). Fixed the same way: hold the pulse past the
edge with a real time delay (#1) before clearing.

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 19:48:46 +02:00
co-authored by Claude Sonnet 5
parent ff5908f25f
commit 1cbe7b85d5
4 changed files with 770 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hardware/v3/rtl/mac2_dsp_packed.v
+99
View File
@@ -0,0 +1,99 @@
`timescale 1ns/1ps
// ============================================================
// v3 (Artix-7 port) -- 2 INT8 MACs sharing one resident weight, packed
// into a single DSP48E1-shaped 25x18 multiply.
//
// Fits this project's own weight-stationary reuse architecture
// (layer_weight_buffer.v, EXP-0057/0058) exactly: one weight stays
// resident and is multiplied against MANY different activations
// (spatial reuse positions). This packs TWO of those activations
// (x0, x1) against the SAME shared weight into one multiply, instead
// of two separate DSP48 multiplies -- doubling effective MAC/DSP
// throughput for exactly this access pattern.
//
// Packing scheme (signed INT8 x0, x1, weight, all in [-128, 127]):
// packed_a = (x1 <<< 16) + sign_extend(x0, 25) (25 bits, matches
// DSP48E1 port A width)
// product = packed_a * weight (33 bits here;
// widens to 43 bits with a real 18-bit weight port on
// actual DSP48E1 silicon)
//
// packed_a is built with a real ARITHMETIC add, not bit concatenation
// -- concatenating two independently sign-extended fields ({sext(x1,9),
// sext(x0,16)}) looks equivalent on paper but is NOT: whenever x0 is
// negative, its own two's-complement encoding contributes an extra
// +2^16 into the concatenated field's value that a real sum x1*2^16+x0
// does not have (found via exhaustive verification below -- an earlier
// concatenation-based version failed exactly 8,355,840 / 16,777,216
// vectors, all sharing x0<0). The explicit shift-and-add avoids this
// class of bug entirely by construction.
//
// Because x1's field sits at bit 16 (a multiple of 2^16), the low 16
// bits of `product` always equal x0*weight exactly, taken as signed
// (modular arithmetic: (x1<<16)*weight is a multiple of 2^16, so it
// never disturbs bits [15:0] of the sum). x0*weight's magnitude is at
// most 128*128=16384, safely inside signed 16-bit range
// (-32768..32767), so no truncation.
//
// Extracting x1*weight from the upper bits needs one correction: an
// arithmetic right-shift of `product` by 16 computes
// floor(product / 2^16), which is x1*weight - 1 (not exactly
// x1*weight) whenever the low-16-bit product (x0*weight) is negative
// -- the classic "borrow" of splitting one real two's-complement sum
// into two fields after the fact (concatenating BEFORE the multiply is
// exact by construction; recovering the two products AFTER a real
// multiply-and-add requires this one correction). Fixed by adding 1
// back whenever the low product's sign bit is set.
// ============================================================
module mac2_dsp_packed #(
parameter DATA_WIDTH = 8
)(
input wire clk,
input wire rst,
input wire signed [DATA_WIDTH-1:0] weight, // shared, resident
input wire signed [DATA_WIDTH-1:0] x0,
input wire signed [DATA_WIDTH-1:0] x1,
input wire valid_in,
output reg signed [2*DATA_WIDTH-1:0] p0, // = x0 * weight, exact
output reg signed [2*DATA_WIDTH-1:0] p1, // = x1 * weight, exact
output reg valid_out
);
localparam A_WIDTH = 3*DATA_WIDTH + 1; // 25 for DATA_WIDTH=8
localparam PROD_WIDTH = A_WIDTH + DATA_WIDTH; // 43 for DATA_WIDTH=8
wire signed [A_WIDTH-1:0] x0_sext25 = {{(A_WIDTH-DATA_WIDTH){x0[DATA_WIDTH-1]}}, x0};
wire signed [A_WIDTH-1:0] x1_shifted = $signed(x1) <<< (2*DATA_WIDTH);
wire signed [A_WIDTH-1:0] packed_a = x1_shifted + x0_sext25;
wire signed [PROD_WIDTH-1:0] product = packed_a * weight;
// NOTE: a Verilog part-select (product[hi:lo]) always yields an
// UNSIGNED value regardless of the source's own `signed` keyword
// (LRM rule -- part-selects are never signed) -- explicit $signed()
// casts below are therefore load-bearing, not decorative: without
// them the arithmetic right shift used to recover p1_raw would
// truncate/zero-extend instead of sign-extending, corrupting every
// case where x1*weight is negative (found via exhaustive
// verification, tb_mac2_dsp_packed.v -- an earlier version without
// these casts, and with an off-by-one in p1_raw's declared width,
// failed ~50% of all 16,777,216 (weight,x0,x1) vectors).
wire signed [2*DATA_WIDTH-1:0] p0_comb = product[2*DATA_WIDTH-1:0];
wire signed [A_WIDTH+DATA_WIDTH-2*DATA_WIDTH-1:0] p1_raw = $signed(product) >>> (2*DATA_WIDTH);
wire signed [2*DATA_WIDTH-1:0] p1_comb = p1_raw[2*DATA_WIDTH-1:0] + (p0_comb[2*DATA_WIDTH-1] ? 1'b1 : 1'b0);
always @(posedge clk) begin
if (rst) begin
p0 <= {2*DATA_WIDTH{1'b0}};
p1 <= {2*DATA_WIDTH{1'b0}};
valid_out <= 1'b0;
end else begin
p0 <= p0_comb;
p1 <= p1_comb;
valid_out <= valid_in;
end
end
endmodule
hardware/v3/rtl/neural_processor_packed.v
+370
View File
@@ -0,0 +1,370 @@
// ============================================================
// FPGA-Neural V3 (Artix-7 port) -- Neural Processor, DSP48-packed.
//
// Direct port of hardware/v2/rtl/neural_processor.v (M1), restructured
// for the weight-stationary reuse pattern (layer_weight_buffer.v,
// EXP-0057/0058): ONE resident weight tile is shared by TWO reuse
// positions (job A, job B) processed in lockstep, each tap-lane packing
// its two x*w multiplies into a single DSP48-shaped multiply instead of
// two separate ones (see hardware/v3/rtl/mac2_dsp_packed.v, verified
// exhaustively 16,777,216/16,777,216 bit-exact -- the packing math
// here is the SAME formula, inlined per-lane rather than instantiated,
// to keep this module's own pipeline depth/stage count identical to
// the V2 original for a direct structural comparison).
//
// Pipeline stages match V2's neural_processor.v exactly, just doubled
// on the accumulator side (one accumulate/bias/activation/saturation
// path per job, A and B, sharing the SAME multiply/adder-tree stages
// since they consume the SAME weight stream):
// Stage 0 input alignment (x0_a, x0_b, w0 -- ONE shared weight)
// Stage 1 P_IN packed-MAC lanes: p0[i]=x0_a[i]*w0[i], p1[i]=x0_b[i]*w0[i]
// Stage 2..(1+TREE_LEVELS) TWO balanced adder trees (A and B)
// Stage (2+TREE_LEVELS) TWO accumulators
// Stage (3+TREE_LEVELS) bias add (shared bias/activation -- same
// neuron/filter, different spatial position)
// + activation, per job
// Stage (4+TREE_LEVELS) INT8 saturation / output register, per job
//
// job_bias/job_activation are SHARED between A and B (same resident
// neuron), matching this project's own weight-reuse semantics (a
// neuron/filter's bias and activation type don't vary by spatial
// position -- only its accumulated dot product does). node_id differs
// per job (A and B are different output positions).
// ============================================================
module neural_processor_packed #(
parameter DATA_WIDTH = 8,
parameter P_IN = 8,
parameter ACC_WIDTH = 32
)(
input clk,
input rst,
// ---- job descriptor (NP_LOAD_JOB) ----
input job_valid,
output job_ready,
input [15:0] job_node_id_a,
input [15:0] job_node_id_b,
input signed [DATA_WIDTH-1:0] job_bias, // shared (same neuron)
input [1:0] job_activation, // shared (same neuron)
// ---- operand stream: ONE shared weight stream, TWO activation streams ----
input operand_valid,
output operand_ready,
input signed [DATA_WIDTH*P_IN-1:0] input_data_a,
input signed [DATA_WIDTH*P_IN-1:0] input_data_b,
input signed [DATA_WIDTH*P_IN-1:0] weight_data,
input tile_last,
// ---- result stream: two results per job pair, same-cycle ----
output reg result_valid,
input result_ready,
output reg signed [DATA_WIDTH-1:0] result_data_a,
output reg signed [DATA_WIDTH-1:0] result_data_b,
output reg [15:0] result_node_id_a,
output reg [15:0] result_node_id_b,
output reg [3:0] np_state,
output reg np_error
);
localparam ACT_NONE = 2'd0;
localparam ACT_RELU = 2'd1;
localparam NP_IDLE = 4'd0;
localparam NP_LOAD_JOB = 4'd1;
localparam NP_WAIT_OPERANDS = 4'd2;
localparam NP_FINISH = 4'd3;
localparam NP_WRITE_RESULT = 4'd4;
localparam NP_DONE = 4'd5;
localparam NP_ERROR = 4'd6;
localparam TREE_LEVELS = $clog2(P_IN);
localparam PROD_WIDTH = 2 * DATA_WIDTH;
reg signed [DATA_WIDTH-1:0] bias_reg;
reg [1:0] activation_reg;
reg [15:0] node_id_a_reg, node_id_b_reg;
assign operand_ready = (np_state == NP_WAIT_OPERANDS);
// ============================================================
// STAGE 0 -- input alignment
// ============================================================
reg valid0, last0;
reg signed [DATA_WIDTH-1:0] xa0 [0:P_IN-1];
reg signed [DATA_WIDTH-1:0] xb0 [0:P_IN-1];
reg signed [DATA_WIDTH-1:0] w0 [0:P_IN-1];
integer gi;
always @(posedge clk) begin
if (rst) begin
valid0 <= 1'b0;
last0 <= 1'b0;
end else begin
valid0 <= operand_valid && operand_ready;
last0 <= (operand_valid && operand_ready) ? tile_last : 1'b0;
if (operand_valid && operand_ready) begin
for (gi = 0; gi < P_IN; gi = gi + 1) begin
xa0[gi] <= input_data_a[gi*DATA_WIDTH +: DATA_WIDTH];
xb0[gi] <= input_data_b[gi*DATA_WIDTH +: DATA_WIDTH];
w0[gi] <= weight_data[gi*DATA_WIDTH +: DATA_WIDTH];
end
end
end
end
// ============================================================
// STAGE 1 -- P_IN packed-MAC lanes (mac2_dsp_packed.v's own
// verified combinational formula, inlined per lane)
// ============================================================
reg valid1, last1;
reg signed [ACC_WIDTH-1:0] proda1 [0:P_IN-1];
reg signed [ACC_WIDTH-1:0] prodb1 [0:P_IN-1];
localparam A_WIDTH = 3*DATA_WIDTH + 1;
wire signed [PROD_WIDTH-1:0] pa_comb [0:P_IN-1];
wire signed [PROD_WIDTH-1:0] pb_comb [0:P_IN-1];
genvar gm;
generate
for (gm = 0; gm < P_IN; gm = gm + 1) begin : GEN_MAC_PACKED
wire signed [A_WIDTH-1:0] x0_sext25 = {{(A_WIDTH-DATA_WIDTH){xa0[gm][DATA_WIDTH-1]}}, xa0[gm]};
wire signed [A_WIDTH-1:0] x1_shifted = $signed(xb0[gm]) <<< (2*DATA_WIDTH);
wire signed [A_WIDTH-1:0] packed_a = x1_shifted + x0_sext25;
wire signed [A_WIDTH+DATA_WIDTH-1:0] product = packed_a * w0[gm];
assign pa_comb[gm] = product[PROD_WIDTH-1:0];
wire signed [A_WIDTH+DATA_WIDTH-2*DATA_WIDTH-1:0] pb_raw =
$signed(product) >>> (2*DATA_WIDTH);
assign pb_comb[gm] = pb_raw[PROD_WIDTH-1:0] + (pa_comb[gm][PROD_WIDTH-1] ? 1'b1 : 1'b0);
end
endgenerate
always @(posedge clk) begin
if (rst) begin
valid1 <= 1'b0;
last1 <= 1'b0;
end else begin
valid1 <= valid0;
last1 <= last0;
for (gi = 0; gi < P_IN; gi = gi + 1) begin
proda1[gi] <= {{(ACC_WIDTH-PROD_WIDTH){pa_comb[gi][PROD_WIDTH-1]}}, pa_comb[gi]};
prodb1[gi] <= {{(ACC_WIDTH-PROD_WIDTH){pb_comb[gi][PROD_WIDTH-1]}}, pb_comb[gi]};
end
end
end
// ============================================================
// STAGES 2..(1+TREE_LEVELS) -- TWO balanced adder trees (A, B)
// ============================================================
wire signed [ACC_WIDTH-1:0] level0a [0:P_IN-1];
wire signed [ACC_WIDTH-1:0] level0b [0:P_IN-1];
genvar gz;
generate
for (gz = 0; gz < P_IN; gz = gz + 1) begin : GEN_TREE_L0
assign level0a[gz] = proda1[gz];
assign level0b[gz] = prodb1[gz];
end
endgenerate
reg [TREE_LEVELS-1:0] valid_tree;
reg [TREE_LEVELS-1:0] last_tree;
reg signed [ACC_WIDTH-1:0] treea [1:TREE_LEVELS][0:P_IN-1];
reg signed [ACC_WIDTH-1:0] treeb [1:TREE_LEVELS][0:P_IN-1];
genvar gl, gn;
generate
for (gl = 0; gl < TREE_LEVELS; gl = gl + 1) begin : GEN_TREE_LEVEL
always @(posedge clk) begin
if (rst) begin
valid_tree[gl] <= 1'b0;
last_tree[gl] <= 1'b0;
end else begin
valid_tree[gl] <= (gl == 0) ? valid1 : valid_tree[gl-1];
last_tree[gl] <= (gl == 0) ? last1 : last_tree[gl-1];
end
end
for (gn = 0; gn < (P_IN >> (gl+1)); gn = gn + 1) begin : GEN_TREE_NODE
if (gl == 0) begin : GEN_FROM_LEVEL0
always @(posedge clk) begin
treea[1][gn] <= level0a[2*gn] + level0a[2*gn+1];
treeb[1][gn] <= level0b[2*gn] + level0b[2*gn+1];
end
end else begin : GEN_FROM_TREE
always @(posedge clk) begin
treea[gl+1][gn] <= treea[gl][2*gn] + treea[gl][2*gn+1];
treeb[gl+1][gn] <= treeb[gl][2*gn] + treeb[gl][2*gn+1];
end
end
end
end
endgenerate
wire valid_tree_out = (TREE_LEVELS == 0) ? valid1 : valid_tree[TREE_LEVELS-1];
wire last_tree_out = (TREE_LEVELS == 0) ? last1 : last_tree[TREE_LEVELS-1];
wire signed [ACC_WIDTH-1:0] tile_sum_a = (TREE_LEVELS == 0) ? proda1[0] : treea[TREE_LEVELS][0];
wire signed [ACC_WIDTH-1:0] tile_sum_b = (TREE_LEVELS == 0) ? prodb1[0] : treeb[TREE_LEVELS][0];
// ============================================================
// STAGE (2+TREE_LEVELS) -- TWO accumulators
// ============================================================
reg signed [ACC_WIDTH-1:0] acc_reg_a, acc_reg_b;
reg valid5, last5;
always @(posedge clk) begin
if (rst) begin
acc_reg_a <= {ACC_WIDTH{1'b0}};
acc_reg_b <= {ACC_WIDTH{1'b0}};
valid5 <= 1'b0;
last5 <= 1'b0;
end else begin
valid5 <= valid_tree_out;
last5 <= last_tree_out;
if (np_state == NP_LOAD_JOB) begin
acc_reg_a <= {ACC_WIDTH{1'b0}};
acc_reg_b <= {ACC_WIDTH{1'b0}};
end else if (valid_tree_out) begin
acc_reg_a <= acc_reg_a + tile_sum_a;
acc_reg_b <= acc_reg_b + tile_sum_b;
end
end
end
// ============================================================
// STAGE (3+TREE_LEVELS) -- bias add + activation (shared bias/act)
// ============================================================
wire signed [ACC_WIDTH-1:0] bias_ext =
{{(ACC_WIDTH-DATA_WIDTH){bias_reg[DATA_WIDTH-1]}}, bias_reg};
reg valid6, last6;
reg signed [ACC_WIDTH-1:0] final_acc_a, final_acc_b;
always @(posedge clk) begin
if (rst) begin
valid6 <= 1'b0;
last6 <= 1'b0;
end else begin
valid6 <= valid5;
last6 <= last5;
final_acc_a <= acc_reg_a + bias_ext;
final_acc_b <= acc_reg_b + bias_ext;
end
end
function automatic signed [DATA_WIDTH-1:0] saturate_activate(
input signed [ACC_WIDTH-1:0] final_acc,
input [1:0] activation
);
reg sign;
reg upper_all0, upper_all1, in_range, le_zero;
reg signed [DATA_WIDTH-1:0] y_none, y_relu;
begin
sign = final_acc[ACC_WIDTH-1];
upper_all0 = ~(|final_acc[ACC_WIDTH-1:DATA_WIDTH-1]);
upper_all1 = &final_acc[ACC_WIDTH-1:DATA_WIDTH-1];
in_range = upper_all0 | upper_all1;
le_zero = sign | ~(|final_acc);
y_none = in_range ? final_acc[DATA_WIDTH-1:0]
: (sign ? {1'b1, {(DATA_WIDTH-1){1'b0}}}
: {1'b0, {(DATA_WIDTH-1){1'b1}}});
y_relu = le_zero ? {DATA_WIDTH{1'b0}}
: (upper_all0 ? final_acc[DATA_WIDTH-1:0]
: {1'b0, {(DATA_WIDTH-1){1'b1}}});
saturate_activate = (activation == ACT_NONE) ? y_none : y_relu;
end
endfunction
// ============================================================
// STAGE (4+TREE_LEVELS) -- output register / saturation, per job
// ============================================================
reg valid7;
reg signed [DATA_WIDTH-1:0] y7_a, y7_b;
always @(posedge clk) begin
if (rst) begin
valid7 <= 1'b0;
end else begin
valid7 <= last6;
y7_a <= saturate_activate(final_acc_a, activation_reg);
y7_b <= saturate_activate(final_acc_b, activation_reg);
end
end
wire pipeline_busy = valid0 || valid1 || (|valid_tree) || valid5 || valid6 || valid7;
assign job_ready = (np_state == NP_IDLE) && !pipeline_busy;
// ============================================================
// OUTER FSM -- identical shape to V2, both result channels together
// ============================================================
always @(posedge clk) begin
if (rst) begin
np_state <= NP_IDLE;
np_error <= 1'b0;
result_valid <= 1'b0;
result_data_a <= {DATA_WIDTH{1'b0}};
result_data_b <= {DATA_WIDTH{1'b0}};
result_node_id_a <= 16'h0;
result_node_id_b <= 16'h0;
bias_reg <= {DATA_WIDTH{1'b0}};
activation_reg <= ACT_RELU;
node_id_a_reg <= 16'h0;
node_id_b_reg <= 16'h0;
end else begin
case (np_state)
NP_IDLE: begin
if (job_valid && job_ready) begin
bias_reg <= job_bias;
activation_reg <= job_activation;
node_id_a_reg <= job_node_id_a;
node_id_b_reg <= job_node_id_b;
np_state <= NP_LOAD_JOB;
end
end
NP_LOAD_JOB: begin
np_state <= NP_WAIT_OPERANDS;
end
NP_WAIT_OPERANDS: begin
if (operand_valid && operand_ready && tile_last) begin
np_state <= NP_FINISH;
end
end
NP_FINISH: begin
if (valid7) begin
result_valid <= 1'b1;
result_data_a <= y7_a;
result_data_b <= y7_b;
result_node_id_a <= node_id_a_reg;
result_node_id_b <= node_id_b_reg;
np_state <= NP_WRITE_RESULT;
end
end
NP_WRITE_RESULT: begin
if (result_valid && result_ready) begin
result_valid <= 1'b0;
np_state <= NP_DONE;
end
end
NP_DONE: begin
np_state <= NP_IDLE;
end
NP_ERROR: begin
end
default: np_state <= NP_ERROR;
endcase
end
end
endmodule