michele/FPGA-Neural
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
Files
a918c3f1e9dfdf7f6d4032cf898015f784e1edac
FPGA-Neural/rtl/layer_sequencer.v
T
micheleandClaude Sonnet 5 a918c3f1e9 feat: configurable activation functions + runtime-configurable network topology 
Two related Phase 5 additions, both threaded the same way (a new
runtime field defaulting to the pre-existing behavior, settable
per-layer via the descriptor table or per-run via SET_BASE):

Configurable activation functions:
- neuron_parallel.v gains a 2-bit `activation` port (ACT_NONE =
  linear + two-sided INT8 saturate, ACT_RELU = the original
  hardwired behavior, kept as the default so every pre-existing
  caller/testbench is unaffected), threaded through neuron_memory.v.
- spi_engine.v: SET_BASE sel=6 (single-layer path); the descriptor
  table gains a 7th byte (multi-layer path).
- Verified in neuron_parallel_tb.v (negative pass-through + negative
  saturation to -128) and end-to-end in
  spi_neuron_top_runnetwork_tb.v (a real negative accumulator that
  ACT_RELU would clamp to 0 comes through unclamped under ACT_NONE,
  over real SPI/RAM).

Runtime network width (one bitstream, any topology up to its
build-time max, entirely host-configured over SPI):
- neuron_parallel.v gains n_inputs_real, bounding its MAC group loop
  (n_inputs_real/PARALLEL groups instead of the fixed build-time
  count). neuron_memory.v gains n_inputs_real/n_neurons_real,
  bounding its X/W RAM-read loop and its neuron loop. All default to
  the build-time max, so unconnected callers are unaffected.
  n_inputs_real must stay a multiple of PARALLEL (same constraint
  N_INPUTS itself is held to at elaboration time, now the caller's
  runtime responsibility).
- spi_engine.v: SET_BASE sel=7/8 (single-layer path); the descriptor
  table grows to 11 bytes/layer (+n_inputs_real +n_neurons_real,
  multi-layer path) -- layer_sequencer.v also now copies only
  n_neurons_real bytes into the ping-pong buffer, not the full
  build width.
- This is real early termination, not bookkeeping: no RAM
  zero-padding needed for the unused tail, and it measurably
  completes faster. neuron_parallel_tb.v TEST 7: 3 cycles vs 6 for a
  reduced-vs-full run, with garbage loaded into the skipped lanes to
  prove they're never read. neuron_memory_tb.v TEST 5: through the
  real PSRAM stack, 209 cycles vs 788. layer_sequencer_tb.v proves a
  reduced n_neurons_real shortens the ping-pong copy-out itself
  (bytes beyond the real count stay untouched, not just differing).

docs/FPGA-NeuralNetwork-Engine.md: §8.1 opcode/SET_BASE table, new
"Runtime network width" subsection, Phase 5 checklist, Current
Status table, and the "Core architectural principle" statement
updated to reflect that topology (not just trained parameters) is
now host-configured at runtime up to a build-time ceiling.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
2026-09-02 20:18:24 +02:00

346 lines
12 KiB
Verilog
Raw Blame History

`timescale 1ns/1ps
// ================================================================
// LAYER_SEQUENCER (Phase 5 - Multi-Layer Network)
//
// Chains up to N_LAYERS runs of a single, reused neuron_memory
// instance to execute a feedforward network of N_LAYERS dense
// layers, without touching neuron_memory.v or the validated compute
// core (neuron_parallel/mac8/mac_unit) at all.
//
// KEY DESIGN CHOICE: neuron_memory's N_INPUTS and N_NEURONS are both
// fixed at synthesis to the SAME value (this module's N_WIDTH
// parameter, e.g. 256). A logical layer with fewer real inputs or
// neurons than N_WIDTH is handled entirely by DATA convention, not
// RTL: the host zero-pads that layer's weight matrix beyond its
// real input count (so the extra MAC lanes contribute 0 regardless
// of input value) and its bias beyond its real neuron count. This
// sequencer then always reads/writes the FULL N_WIDTH-byte buffer
// for every layer transition -- it does not need to know any
// layer's "real" input/neuron count at all. Trade-off: a layer with
// few real inputs still takes as long as a full N_WIDTH-wide layer
// (wasted MAC cycles on zero-weighted padding); documented as a
// known Phase 7 (Optimization) follow-up, not solved here.
//
// Layer descriptor table (host-written via WRITE_RAM, read-only to
// this module): N_LAYERS entries of 11 bytes each, MSB-first,
// starting at `table_base`:
// w_base(3B), bias_addr(3B), activation(1B, low 2 bits --
// see rtl/neuron_parallel.v's ACT_* localparams),
// n_inputs_real(2B), n_neurons_real(2B)
// Layer 0's input is the external `x_base` (same register used for
// single-layer/manual mode). Layer k>0's input is the ping-pong
// output buffer (`buf_a_base`/`buf_b_base`) written by layer k-1.
// The final layer's output is left both in neuron_memory's own
// y_bus (readable via the existing READ_OUTPUT opcode, unchanged)
// and in the ping-pong buffer it was copied to.
//
// n_inputs_real/n_neurons_real let ONE synthesized bitstream (fixed
// N_WIDTH = neuron_memory's build-time max) serve any real network
// topology up to that width: forwarded to neuron_memory verbatim
// (see rtl/neuron_memory.v), and this sequencer copies exactly
// n_neurons_real bytes of y_bus into the ping-pong buffer -- NOT the
// full N_WIDTH -- so a narrower layer both computes AND is copied
// out faster, no zero-padding required in RAM.
// ================================================================
module layer_sequencer #(
parameter ADDR_WIDTH = 22,
parameter DATA_WIDTH = 8,
parameter N_WIDTH = 256, // = neuron_memory's N_INPUTS = N_NEURONS
parameter N_LAYERS = 4
)(
input wire clk,
input wire rst,
// ------------------------------------------------------------
// Trigger (from spi_engine's RUN_NETWORK opcode)
// ------------------------------------------------------------
input wire run_start, // one-cycle pulse
input wire [7:0] run_num_layers, // 1..N_LAYERS
output reg seq_busy,
output reg seq_done, // one-cycle pulse, mirrors neuron_memory.done
// ------------------------------------------------------------
// Config registers (from spi_engine's SET_BASE)
// ------------------------------------------------------------
input wire [ADDR_WIDTH-1:0] x_base, // layer 0's external input
input wire [ADDR_WIDTH-1:0] table_base,
input wire [ADDR_WIDTH-1:0] buf_a_base,
input wire [ADDR_WIDTH-1:0] buf_b_base,
// ------------------------------------------------------------
// neuron_memory control (sequencer-owned; only meaningful while
// seq_busy -- the top level muxes these against spi_engine's
// own direct-drive outputs based on seq_busy)
// ------------------------------------------------------------
output reg [ADDR_WIDTH-1:0] nm_x_base,
output reg [ADDR_WIDTH-1:0] nm_w_base,
output reg [ADDR_WIDTH-1:0] nm_bias_addr,
output reg [1:0] nm_activation,
output reg [15:0] nm_n_inputs,
output reg [15:0] nm_n_neurons,
output reg nm_start,
input wire nm_busy,
input wire nm_done, // one-cycle pulse
input wire signed [DATA_WIDTH*N_WIDTH-1:0] y_bus,
// ------------------------------------------------------------
// Byte-level RAM master port (own arbiter port)
// ------------------------------------------------------------
output reg ram_req,
output reg ram_wr,
output reg [ADDR_WIDTH-1:0] ram_addr,
output reg signed [7:0] ram_wdata,
input wire signed [7:0] ram_rdata,
input wire ram_ready
);
// ============================================================
// STATES
// ============================================================
localparam ST_IDLE = 4'd0;
localparam ST_READ_DESC = 4'd1;
localparam ST_READ_WAIT = 4'd2;
localparam ST_START_LAYER = 4'd3;
localparam ST_WAIT_LAYER = 4'd4;
localparam ST_COPY_ISSUE = 4'd5;
localparam ST_COPY_WAIT = 4'd6;
reg [3:0] state;
reg [7:0] layer_idx;
reg [7:0] num_layers_reg;
reg [ADDR_WIDTH-1:0] desc_table_addr;
reg [3:0] desc_byte_idx; // 0..10
reg [23:0] w_base_acc;
reg [23:0] bias_addr_acc;
reg [7:0] activation_acc;
reg [15:0] n_inputs_acc;
reg [15:0] n_neurons_acc;
reg cur_sel; // which ping-pong buffer to READ from for this layer (layer_idx>0)
reg write_sel; // which ping-pong buffer to WRITE this layer's output to
reg [$clog2(N_WIDTH+1)-1:0] copy_idx;
always @(posedge clk) begin
if (rst) begin
state <= ST_IDLE;
layer_idx <= 8'd0;
num_layers_reg <= 8'd0;
desc_table_addr <= {ADDR_WIDTH{1'b0}};
desc_byte_idx <= 4'd0;
w_base_acc <= 24'h0;
bias_addr_acc <= 24'h0;
activation_acc <= 8'h0;
n_inputs_acc <= 16'h0;
n_neurons_acc <= 16'h0;
cur_sel <= 1'b0;
write_sel <= 1'b0;
copy_idx <= 0;
nm_x_base <= {ADDR_WIDTH{1'b0}};
nm_w_base <= {ADDR_WIDTH{1'b0}};
nm_bias_addr <= {ADDR_WIDTH{1'b0}};
nm_activation <= 2'd1; // ACT_RELU
nm_n_inputs <= 16'h0;
nm_n_neurons <= 16'h0;
nm_start <= 1'b0;
ram_req <= 1'b0;
ram_wr <= 1'b0;
ram_addr <= {ADDR_WIDTH{1'b0}};
ram_wdata <= 8'sd0;
seq_busy <= 1'b0;
seq_done <= 1'b0;
end else begin
// --------------------------------------------------
// Default pulses
// --------------------------------------------------
nm_start <= 1'b0;
ram_req <= 1'b0;
seq_done <= 1'b0;
case (state)
// =================================================
// IDLE
// =================================================
ST_IDLE: begin
if (run_start) begin
seq_busy <= 1'b1;
layer_idx <= 8'd0;
num_layers_reg <= run_num_layers;
desc_table_addr <= table_base;
desc_byte_idx <= 4'd0;
write_sel <= 1'b0;
state <= ST_READ_DESC;
end
end
// =================================================
// READ DESCRIPTOR (6 bytes: w_base, bias_addr)
// =================================================
ST_READ_DESC: begin
ram_req <= 1'b1;
ram_wr <= 1'b0;
ram_addr <= desc_table_addr + desc_byte_idx;
state <= ST_READ_WAIT;
end
ST_READ_WAIT: begin
if (ram_ready) begin
case (desc_byte_idx)
4'd0: w_base_acc[23:16] <= ram_rdata;
4'd1: w_base_acc[15:8] <= ram_rdata;
4'd2: w_base_acc[7:0] <= ram_rdata;
4'd3: bias_addr_acc[23:16] <= ram_rdata;
4'd4: bias_addr_acc[15:8] <= ram_rdata;
4'd5: bias_addr_acc[7:0] <= ram_rdata;
4'd6: activation_acc <= ram_rdata;
4'd7: n_inputs_acc[15:8] <= ram_rdata;
4'd8: n_inputs_acc[7:0] <= ram_rdata;
4'd9: n_neurons_acc[15:8] <= ram_rdata;
4'd10: n_neurons_acc[7:0] <= ram_rdata;
endcase
if (desc_byte_idx == 4'd10) begin
desc_byte_idx <= 4'd0;
state <= ST_START_LAYER;
end else begin
desc_byte_idx <= desc_byte_idx + 4'd1;
state <= ST_READ_DESC;
end
end
end
// =================================================
// START LAYER
// =================================================
ST_START_LAYER: begin
nm_w_base <= w_base_acc[ADDR_WIDTH-1:0];
nm_bias_addr <= bias_addr_acc[ADDR_WIDTH-1:0];
nm_activation <= activation_acc[1:0];
nm_n_inputs <= n_inputs_acc;
nm_n_neurons <= n_neurons_acc;
nm_x_base <= (layer_idx == 8'd0)
? x_base
: (cur_sel ? buf_b_base : buf_a_base);
nm_start <= 1'b1;
state <= ST_WAIT_LAYER;
end
// =================================================
// WAIT FOR THIS LAYER TO FINISH
// =================================================
ST_WAIT_LAYER: begin
if (nm_done) begin
copy_idx <= 0;
state <= ST_COPY_ISSUE;
end
end
// =================================================
// COPY y_bus INTO THE PING-PONG OUTPUT BUFFER
// =================================================
ST_COPY_ISSUE: begin
ram_req <= 1'b1;
ram_wr <= 1'b1;
ram_addr <= (write_sel ? buf_b_base : buf_a_base) + copy_idx;
ram_wdata <= y_bus[copy_idx*DATA_WIDTH +: DATA_WIDTH];
state <= ST_COPY_WAIT;
end
ST_COPY_WAIT: begin
if (ram_ready) begin
if (copy_idx == n_neurons_acc[$clog2(N_WIDTH+1)-1:0] - 1'b1) begin
if (layer_idx == num_layers_reg - 8'd1) begin
// Last layer done.
seq_busy <= 1'b0;
seq_done <= 1'b1;
state <= ST_IDLE;
end else begin
// The buffer just written becomes
// the next layer's input.
cur_sel <= write_sel;
write_sel <= ~write_sel;
layer_idx <= layer_idx + 8'd1;
desc_table_addr <= desc_table_addr + 22'd11;
desc_byte_idx <= 4'd0;
state <= ST_READ_DESC;
end
end else begin
copy_idx <= copy_idx + 1'b1;
state <= ST_COPY_ISSUE;
end
end
end
default: begin
state <= ST_IDLE;
end
endcase
end
end
endmodule
Reference in New Issue View Git Blame Copy Permalink