Merge pull request #15 from pulp-platform/chi/increase_ce_array

Allow larger accelerator dimensions.

rtl/redmule_ctrl.sv

@@ -74,7 +74,7 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
   logic [4:0]  w_computed;
   logic [15:0] w_rows;
   logic [15:0] w_rows_iter, w_row_count_d, w_row_count_q;
-  logic [15:0] z_storings_d, z_storings_q, tot_stores;
+  logic [15:0] z_storings_d, z_storings_q, tot_stores, issued_store_d, issued_store_q;
 
   typedef enum logic [2:0] {REDMULE_IDLE, REDMULE_STARTING, REDMULE_COMPUTING, REDMULE_BUFFERING, REDMULE_STORING, REDMULE_FINISHED} redmule_ctrl_state;
   redmule_ctrl_state current, next;
@@ -191,6 +191,18 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
   end
   assign accumulate_o =  accumulate_q & !accumulate_ctrl_q;
 
+  logic finish_d, finish_q;
+  always_ff @(posedge clk_i or negedge rst_ni) begin : finish_sampler
+    if(~rst_ni) begin
+      finish_q <= 1'b0;
+    end else begin
+      if (clear)
+        finish_q <= 1'b0;
+      else
+        finish_q <= finish_d;
+    end
+  end
+
   always_ff @(posedge clk_i or negedge rst_ni) begin : last_w_row_reg
     if(~rst_ni) begin
       last_w_row <= 1'b0;
@@ -202,6 +214,17 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
     end
   end
 
+  always_ff @(posedge clk_i or negedge rst_ni) begin : issued_store
+    if(~rst_ni) begin
+      issued_store_q <= '0;
+    end else begin
+      if (clear)
+        issued_store_q <= '0;
+      else
+        issued_store_q <= issued_store_d;
+    end
+  end
+
   // This register counts the number of times we exit from the REDMULE_STORING
   // state and go to the REDMULE_COMPUTING one. Every time this happens, it
   // means that a piece of computation fas done, and we can track the number
@@ -258,6 +281,8 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
     last_w_row_rst     = 1'b0;
     w_row_count_d      = w_row_count_q;
     z_storings_d       = z_storings_q;
+    issued_store_d     = issued_store_q;
+    finish_d           = finish_q;
     accumulate_en      = 1'b0;
     accumulate_rst     = 1'b0;
     storing_rst        = 1'b0;
@@ -276,6 +301,7 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
           z_buffer_clk_en = 1'b1;
         if ( (slave_start & tiler_valid) || test_mode_i) begin
           tiler_setback = 1'b1;
+          finish_d = 1'b0;
           next = REDMULE_STARTING;
         end
         else 
@@ -307,20 +333,23 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
 
         case (last_w_row)
           1'b0: begin
-            if (w_computed == Height - 1) begin
-              if (!accumulate_q)
+            if (w_computed == NumPipeRegs) begin
+              if (!accumulate_q && !finish_q)
                   accumulate_en = 1'b1;
               if (count_w_q)
                   w_computed_rst = 1'b1;
             end
           end
 
           1'b1: begin
-            if (w_computed == Height - 2 && reg_enable_i) begin
+            if (w_computed == NumPipeRegs - 1 && reg_enable_i) begin
               w_row_count_d = 16'd1;
+              issued_store_d = issued_store_q + 'd1;
               next = REDMULE_BUFFERING;
-              if (accumulate_q)
+              if (accumulate_q) begin
                 accumulate_rst = 1'b1;
+                finish_d = (issued_store_q == tot_stores - 1) ? 1'b1 : 1'b0;
+              end
               if (count_w_q)
                 w_computed_rst = 1'b1;
             end else
@@ -337,7 +366,7 @@ localparam int unsigned LEFT_PARAMS   = LEFT_PARAMS
           w_row_count_d = w_row_count_q + 1;
           z_fill_o = reg_enable_i;
         if (flgs_z_buffer_i.full) begin
-          accumulate_en = 1'b1;
+          accumulate_en = finish_q ? 1'b0 : 1'b1;
           next = REDMULE_STORING;
         end
         else

rtl/redmule_pkg.sv

@@ -18,6 +18,7 @@
  * 
  * RedMulE Package
  */
+
 `include "hci/typedef.svh"
 `include "hci/assign.svh"
 `include "hwpe-ctrl/typedef.svh"
@@ -28,7 +29,7 @@ import hwpe_stream_package::*;
 
 package redmule_pkg;
 
-  parameter int unsigned            DATA_W       = 288; // TCDM port dimension (in bits)
+  parameter int unsigned            DATA_W       = 544; // TCDM port dimension (in bits)
   parameter int unsigned            MemDw        = 32;
   parameter int unsigned            NumByte      = MemDw/8;
   parameter int unsigned            ADDR_W       = hci_package::DEFAULT_AW;
@@ -37,9 +38,9 @@ package redmule_pkg;
   parameter int unsigned            N_CONTEXT    = 2;
   parameter fpnew_pkg::fp_format_e  FPFORMAT     = fpnew_pkg::FP16;
   parameter int unsigned            BITW         = fpnew_pkg::fp_width(FPFORMAT);
-  parameter int unsigned            ARRAY_HEIGHT = 4;
+  parameter int unsigned            ARRAY_HEIGHT = 8;
   parameter int unsigned            PIPE_REGS    = 3;
-  parameter int unsigned            ARRAY_WIDTH  = 12; /* Superior limit is ARRAY_HEIGHT*PIPE_REGS */
+  parameter int unsigned            ARRAY_WIDTH  = ARRAY_HEIGHT*PIPE_REGS; // Superior limit, smaller values are allowed.
   parameter int unsigned            TOT_DEPTH    = DATAW/BITW;
   parameter int unsigned            DEPTH        = TOT_DEPTH/ARRAY_HEIGHT;
   parameter int unsigned            STRB         = DATA_W/8;

rtl/redmule_scheduler.sv

@@ -717,6 +717,18 @@ always_ff @(posedge clk_i or negedge rst_ni) begin : tot_stores_counter
   end
 end
 
+logic [15:0] consumed_y_d, consumed_y_q;
+always_ff @(posedge clk_i or negedge rst_ni) begin : y_consumed_counter
+  if(~rst_ni) begin
+    consumed_y_q <= '0;
+  end else begin
+    if (clear_i || clear_regs)
+      consumed_y_q <= '0;
+    else
+      consumed_y_q <= consumed_y_d;
+  end
+end
+
 always_comb begin
   if (store_rows_lftovr_q == '0 || tot_z_stored_q < store_rows_lftovr_q) begin
     if (store_cols_lftovr_q) begin
@@ -989,6 +1001,7 @@ transfer_count_d = transfer_count_q;
 gate_count_d     = gate_count_q;
 store_count_d    = store_count_q;
 tot_store_d      = tot_store_q;
+consumed_y_d     = consumed_y_q;
 y_rows_iter_d    = y_rows_iter_q;
 y_cols_iter_d    = y_cols_iter_q;
 
@@ -1116,6 +1129,7 @@ clear_regs       = 1'b0;
       if (reg_file_i.hwpe_params[OP_SELECTION][0]) begin
         y_push_rst = (flgs_z_buffer_i.y_pushed) ?  1'b1 : 1'b0;
         consume_y_rst = (flgs_z_buffer_i.y_pushed && consume_y_q) ? 1'b1 : 1'b0;
+        consumed_y_d = consume_y_rst ? consumed_y_q + 'd1 : consumed_y_q;
         if (!accumulate_i && consume_y_q && !skip_w_q) begin
           flgs_scheduler_o.y_push_enable = 1'b1;
           z_buffer_clk_en          = 1'b1;
@@ -1183,7 +1197,8 @@ clear_regs       = 1'b0;
       if (w_valid_i == 1'b1 && w_strb_i == '1) begin
         w_loaded = 1'b1;
         count_w_cycles_en = (!count_w_cycles_q & x_preloaded_q) ? 1'b1 : 1'b0;
-        if (reg_file_i.hwpe_params[OP_SELECTION][0]) begin
+        if (reg_file_i.hwpe_params[OP_SELECTION][0] &&
+            consumed_y_q < reg_file_i.hwpe_params[LEFT_PARAMS][31:16]) begin
           y_push_en = (!y_push_q) ? 1'b1 : 1'b0;
           consume_y_en = (!consume_y_q) ? 1'b1 : 1'b0;
         end
@@ -1210,9 +1225,9 @@ clear_regs       = 1'b0;
             if ((reg_file_i.hwpe_params[LEFTOVERS][7:0] != '0) && (y_cols_lftovr_q == '0))
               y_cols_lftovr_en = 1'b1;
           end
-        end else if (d_shift_d == H && !flgs_streamer_i.x_stream_source_flags.ready_start) begin
+        end else if ((d_shift_d == NumPipeRegs + 'd1) && !flgs_streamer_i.x_stream_source_flags.ready_start) begin
           load_x_en = 1'b1;
-            d_shift_d = '0;
+          d_shift_d = '0;
           next = LOAD_X;
           if (x_cols_iter_q == reg_file_i.hwpe_params[X_ITERS][15:0]) begin
             if ( (reg_file_i.hwpe_params[LEFTOVERS][23:16] != '0) && (x_cols_lftovr_q == '0) )
@@ -1245,7 +1260,7 @@ clear_regs       = 1'b0;
       x_buffer_clk_en = 1'b1;
       if (reg_file_i.hwpe_params[OP_SELECTION][0]) begin
         if (!accumulate_i & !skip_w_q) begin
-          flgs_scheduler_o.y_push_enable = 1'b1;
+          flgs_scheduler_o.y_push_enable = consume_y_q ? 1'b1 : 1'b0;
           z_buffer_clk_en          = 1'b1;
         end
         consume_y_rst = (flgs_z_buffer_i.y_pushed && consume_y_q) ? 1'b1 : 1'b0;
@@ -1340,7 +1355,7 @@ clear_regs       = 1'b0;
       gate_count_d = '0;
 
       if (!accumulate_i && reg_file_i.hwpe_params[OP_SELECTION][0] && !skip_w_q) begin
-        flgs_scheduler_o.y_push_enable = 1'b1;
+        flgs_scheduler_o.y_push_enable = consume_y_q ? 1'b1 : 1'b0;
         z_buffer_clk_en          = 1'b1;
       end
 

rtl/redmule_z_buffer.sv

@@ -66,11 +66,11 @@ always_ff @(posedge buffer_clock or negedge rst_ni) begin : z_buffer
       z_buffer_q <= '0;
     else if (ctrl_i.fill || ctrl_i.y_push_enable) begin
         if (reg_enable_i) begin
-          for (int d = 0; d < D; d++) begin    
+          for (int d = 0; d < D; d++) begin
             for (int w = 0; w < W; w++)
               z_buffer_q[d][w] <= (d == 0) ? z_buffer_i[w] : z_buffer_q[d-1][w];
           end
-        end else 
+        end else
           z_buffer_q <= z_buffer_q;
     end else if (ctrl_i.store && ctrl_i.ready) begin
       for (int w = 0; w < W; w++) begin

sw/archi_redmule.h

@@ -68,11 +68,11 @@
 
 // RedMulE architecture
 #define ADDR_WIDTH   32
-#define DATA_WIDTH   256
+#define DATA_WIDTH   512
 #define REDMULE_FMT  16
-#define ARRAY_HEIGHT 4
+#define ARRAY_HEIGHT 8
 #define PIPE_REGS    3
-#define ARRAY_WIDTH  12 /* Superior limit is ARRAY_HEIGHT*PIPE_REGS */
+#define ARRAY_WIDTH  24 /* Superior limit is ARRAY_HEIGHT*PIPE_REGS */
 
 // Base address
 #define REDMULE_BASE_ADD 0x00100000

tb/redmule_complex_tb.sv

@@ -22,12 +22,13 @@ timeunit 1ps;
 timeprecision 1ps;
 
 module redmule_complex_tb;
+import redmule_pkg::*;
 
   // parameters
   parameter int unsigned PROB_STALL = 0;
   parameter int unsigned NC = 1;
   parameter int unsigned ID = 10;
-  parameter int unsigned DW = 288;
+  parameter int unsigned DW = redmule_pkg::DATA_W;
   parameter int unsigned MP = DW/32;
   parameter int unsigned MEMORY_SIZE = 192*1024;
   parameter int unsigned STACK_MEMORY_SIZE = 192*1024;

tb/redmule_tb.sv

@@ -23,12 +23,13 @@ timeunit 1ps;
 timeprecision 1ps;
 
 module redmule_tb;
+import redmule_pkg::*;
 
   // parameters
   parameter int unsigned PROB_STALL = 0;
   parameter int unsigned NC = 1;
   parameter int unsigned ID = 10;
-  parameter int unsigned DW = 288;
+  parameter int unsigned DW = redmule_pkg::DATA_W;
   parameter int unsigned MP = DW/32;
   parameter int unsigned MEMORY_SIZE = 192*1024;
   parameter int unsigned STACK_MEMORY_SIZE = 192*1024;