Get rid of indexed memory access to accumulators/accumulators_next. S…

…lices=2

src/1_58bit_mul.v

@@ -4,6 +4,8 @@
  */
 
 // TODOs:
+//  + remove indexed memory access in out_queue and accumulators/accumulators_next
+//  * get rid of (slice_count - 1) wait once indexed memory access is removed
 //  * special weights 243..255 can be used to   a) initiate readout,
 //                                              b) shift accumulators by 1,
 //                                              c) set beta&gamma,
@@ -162,18 +164,27 @@ module systolic_array #(
             wire [16:0] value_curr  = accumulators     [i*W+j];
             wire [16:0] value_next  = accumulators_next[i*W+j];
             wire [16:0] value_queue = out_queue        [i*W+j];
-            wire skip = (j != slice_counter) | arg_left_zero_curr[i];
+            // wire skip = (j != slice_counter) | arg_left_zero_curr[i];
+            wire skip = arg_left_zero_curr[i];
             wire sign = arg_left_sign_curr[i];
-            wire signed [7:0] addend = $signed(arg_top_curr[j*8 +: 8]);
-            assign accumulators_next[i*W+j] =
-                 reset  ? 0 :
-                 skip   ? accumulators[i*W+j] + 0 :
-                (sign   ? accumulators[i*W+j] - addend :
-                          accumulators[i*W+j] + addend);
+            // wire signed [7:0] addend = $signed(arg_top_curr[j*8 +: 8]);
+            wire signed [7:0] addend = $signed(arg_top_curr[slice_counter*8 +: 8]);
+            if (j == 0) begin
+                assign accumulators_next[i*W+W-1] =
+                     reset  ? 0 :
+                     skip   ? accumulators[i*W+j] + 0 :
+                    (sign   ? accumulators[i*W+j] - addend :
+                              accumulators[i*W+j] + addend);
+            end else begin
+                assign accumulators_next[i*W+j-1] =
+                    reset  ? 0 :
+                              accumulators[i*W+j];
+            end
         end
     endgenerate
 
     assign out = out_queue[0] >> 8;
+    // assign out = out_queue[0][7:0];
     // assign out = out_queue[out_queue_counter] >> 8;
     // assign out = out_queue[out_queue_counter][7:0];
 endmodule

test/Makefile

@@ -2,7 +2,7 @@
 # See https://docs.cocotb.org/en/stable/quickstart.html for more info
 
 # global parameters
-COMPUTE_SLICES ?= 1
+COMPUTE_SLICES ?= 2
 COMPILE_ARGS += -DCOMPUTE_SLICES=$(COMPUTE_SLICES)
 
 # defaults

test/test.py

@@ -120,7 +120,7 @@ async def gemm(dut, weights, inputs, weights_per_byte = 4, compute_block_width =
             # Wait until all slices have finished accunulating
             dut.ui_in.value = 0
             dut.uio_in.value = 0
-            await ClockCycles(dut.clk, compute_slices)
+            await ClockCycles(dut.clk, compute_slices + (compute_slices-1))
 
             # Move accumulators to output queue
             dut.ena.value = 0
@@ -225,7 +225,7 @@ async def test_gemm_small(dut):
     inputs   = random_matrix(-127, 127, (K, M))
     expected = matrix_mul(weights, inputs)
 
-    await reset_run_and_validate_gemm(dut, weights, inputs, expected)
+    await reset_run_and_validate_gemm(dut, weights, inputs, expected, verbose=True)
 
 @cocotb.test()
 async def test_gemm_large(dut):