Root-causes and fixes the real, disclosed defect left open at the end of the previous STEP20 commit: the board-level SPI host interface produced wrong compute results when jobs were dispatched with realistic (widely time-separated) pacing, even though job registration itself was already confirmed correct at the dependency_manager handshake. Root cause: nms_weight_packed.v and nms_activation_replicated.v both used a REGISTERED SRAM read (rd_data_reg <= mem[addr], one full clock of latency), but nms_memory_manager_stream_wide.v's own read-ahead pipeline (its `rd_pending` bit) is designed around a COMBINATIONAL read -- a request issued this cycle produces data already valid to capture the very next cycle. A busy, multi-tile job (e.g. the STEP19 D-Stress regression, 16 tiles/neuron) never exposes the mismatch, since its own weight/activation prefetch always runs far enough ahead that any given tile has been sitting stable in the SRAM for many cycles by the time it's actually consumed. An uncontested single-tile job has zero such margin: its one tile's read fires on the exact edge the data nominally becomes ready, landing squarely on the missing cycle and permanently latching stale/zero data. Fixed by making both SRAMs' reads combinational, with an explicit same-cycle fill/read address-match bypass for the one hazard a plain combinational read alone would still miss. No FSM, arbiter, or SDRAM controller logic was touched. Verified (Verilator, per this project's own standing DEC-0004 protocol): - tb_fpga_neural_v2_top_smoke.v: 11/11 PASS -- single job, back-to-back jobs, a realistic ~85us-gap job pair, and a parametric sweep of inter-job gaps (100ns/5000ns/50000ns). - STEP19 D-Stress N=2: 49788 cycles, 256/256 bit-exact -- identical cycle count to before this fix (zero regression). - STEP19 D-Stress N=4: 49771 cycles, 256/256 bit-exact -- identical cycle count to before this fix (zero regression). - tb_sdram_unified_backend.v (40/40) and tb_spi_host_bridge.v (18/18) reconfirmed unaffected. The physical SPI host interface is now verified correct end-to-end. Real synthesis/P&R of the board-level top (fpga_neural_v2_top.v) is the deliberate next step, not yet performed this round. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
71 lines
3.4 KiB
Verilog
71 lines
3.4 KiB
Verilog
// ============================================================
|
|
// Neural Memory System (NMS) -- Weight SRAM candidate W2 "packed":
|
|
// same private-per-slot semantics as nms_weight_direct.v, but each
|
|
// slot's P_IN*DATA_WIDTH-wide tile storage is decomposed into P_IN
|
|
// separate, narrow (DATA_WIDTH=8-bit-wide) single-port memories (one
|
|
// per MAC lane) instead of one wide 64-bit memory. Reassembly into the
|
|
// full tile word is a static concatenation of P_IN registered
|
|
// per-lane outputs -- no runtime mux, no real LUT cost expected there.
|
|
// Exists to measure whether narrower-but-more-numerous memories pack
|
|
// into fewer total DP16KD blocks than nms_weight_direct.v's wider-but-
|
|
// fewer instances, per the user's own explicit ask (NMS spec S6) to
|
|
// measure width/depth/packing real DP16KD cost rather than assume it.
|
|
// ============================================================
|
|
module nms_weight_packed #(
|
|
parameter DATA_WIDTH = 8,
|
|
parameter P_IN = 8,
|
|
parameter N_SLOTS = 4,
|
|
parameter MAX_TILES = 16,
|
|
parameter TIW = (MAX_TILES <= 1) ? 1 : $clog2(MAX_TILES)
|
|
)(
|
|
input clk,
|
|
input rst,
|
|
|
|
input [N_SLOTS-1:0] fill_we,
|
|
input [N_SLOTS*TIW-1:0] fill_addr_flat,
|
|
input [N_SLOTS*DATA_WIDTH*P_IN-1:0] fill_data_flat,
|
|
|
|
input [N_SLOTS-1:0] rd_en,
|
|
input [N_SLOTS*TIW-1:0] rd_addr_flat,
|
|
output [N_SLOTS*DATA_WIDTH*P_IN-1:0] rd_data_flat
|
|
);
|
|
|
|
genvar g, p;
|
|
generate
|
|
for (g = 0; g < N_SLOTS; g = g + 1) begin : GEN_SLOT
|
|
for (p = 0; p < P_IN; p = p + 1) begin : GEN_LANE
|
|
reg [DATA_WIDTH-1:0] mem [0:MAX_TILES-1];
|
|
|
|
always @(posedge clk) begin
|
|
if (fill_we[g])
|
|
mem[fill_addr_flat[g*TIW +: TIW]] <=
|
|
fill_data_flat[g*DATA_WIDTH*P_IN + p*DATA_WIDTH +: DATA_WIDTH];
|
|
end
|
|
|
|
// ROOT CAUSE (STEP20, ERR-0025 Part B) -- see
|
|
// nms_activation_replicated.v's own header for the
|
|
// full writeup: this read must be COMBINATIONAL, not
|
|
// registered, to match nms_memory_manager_stream_wide.v's
|
|
// own `rd_pending` pipeline's actual 1-cycle latency
|
|
// assumption (issue this cycle, capture next cycle). A
|
|
// registered read added a second, uncounted cycle of
|
|
// latency that a busy multi-tile job's own prefetch
|
|
// lead time always absorbed invisibly, but an
|
|
// uncontested single-tile job's first (only) tile does
|
|
// not -- permanently latching stale/zero data. The
|
|
// same-cycle fill/read bypass covers the one case a
|
|
// combinational read alone would still miss: a fill
|
|
// and a read to the identical address landing on the
|
|
// identical edge (mem[] itself only reflects a
|
|
// same-edge write starting the NEXT cycle).
|
|
wire rd_bypass = fill_we[g] &&
|
|
(fill_addr_flat[g*TIW +: TIW] == rd_addr_flat[g*TIW +: TIW]);
|
|
assign rd_data_flat[g*DATA_WIDTH*P_IN + p*DATA_WIDTH +: DATA_WIDTH] =
|
|
rd_bypass ? fill_data_flat[g*DATA_WIDTH*P_IN + p*DATA_WIDTH +: DATA_WIDTH]
|
|
: mem[rd_addr_flat[g*TIW +: TIW]];
|
|
end
|
|
end
|
|
endgenerate
|
|
|
|
endmodule
|