Update runtime & tb to work in absolute addresses, not offsets

KastnerRG · Nov 7, 2023 · 2820d01 · 2820d01
1 parent 50e56ab
commit 2820d01
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 30 deletions.
diff --git a/c/runtime.h b/c/runtime.h
@@ -396,21 +396,19 @@ extern EXT_C void load_y (uint8_t *p_done, uint8_t *pt_done_proc,  const uint32_
 }
 
 
-extern EXT_C void load_x (uint8_t *p_done, uint8_t *bundle_read_done, int32_t *p_offset, int32_t *p_bpt) {
+extern EXT_C void load_x (uint8_t *p_done, uint8_t *bundle_read_done, uint64_t *p_base_addr, int32_t *p_bpt) {
 
   static int32_t ib=0, ip=0, it=0, offset_next=0;
   static char bundle_read_done_next = 0;
-  int32_t offset = offset_next;
-  int32_t bpt = ip == 0 ? bundles[ib].x_bpt_p0 : bundles[ib].x_bpt;
 
-  *p_offset = offset;
-  *p_bpt = bpt;
+  *p_base_addr = (uint64_t)&mem.x + offset_next;
+  *p_bpt = ip == 0 ? bundles[ib].x_bpt_p0 : bundles[ib].x_bpt;
   *bundle_read_done = bundle_read_done_next;
 
   bundle_read_done_next = 0;
   // Nested for loop [for ib: for ip: for it: {}] inverted to increment once per call
   ++ it; if (it >= bundles[ib].t) { it = 0;
-    offset_next += bpt;
+    offset_next += *p_bpt;
     ++ ip; if (ip >= bundles[ib].p) { ip = 0;
 
       bundle_read_done_next = 1;
@@ -423,14 +421,14 @@ extern EXT_C void load_x (uint8_t *p_done, uint8_t *bundle_read_done, int32_t *p
 }
 
 
-extern EXT_C void load_w (uint8_t *p_done, int32_t *p_offset, int32_t *p_bpt) {
+extern EXT_C void load_w (uint8_t *p_done, uint64_t *p_base_addr, int32_t *p_bpt) {
 
   static int32_t ib=0, ip=0, it=0, offset_next=0;
 
   int32_t offset = offset_next;
   int32_t bpt = ip == 0 ? bundles[ib].w_bpt_p0 : bundles[ib].w_bpt;
 
-  *p_offset = offset;
+  *p_base_addr = (uint64_t)&mem.w + offset;
   *p_bpt = bpt;
 
   // Nested for loop [for ib: for ip: for it: {}] inverted to increment once per call
@@ -444,7 +442,7 @@ extern EXT_C void load_w (uint8_t *p_done, int32_t *p_offset, int32_t *p_bpt) {
 }
 
 
-extern EXT_C void fill_memory (){
+extern EXT_C void fill_memory (uint64_t *p_w_base, uint64_t *p_x_base){
   FILE *fp;
   char f_path [1000];
 
@@ -464,13 +462,18 @@ extern EXT_C void fill_memory (){
 
   for (int32_t i=0; i<B_WORDS; i++)
     printf("i:%d, bias:%d\n", i, mem.b[i]);
+
+  *p_w_base = (uint64_t)&mem.w;
+  *p_x_base = (uint64_t)&mem.x;
 }
 
-extern EXT_C int8_t get_byte_wx (int32_t addr, int32_t mode){
-  if      (mode==0) return mem.w[addr];
-  else if (mode==1) return mem.x[addr];
+extern EXT_C int8_t get_byte (uint64_t addr){
+  return *(int8_t*)addr;
 }
 
 extern EXT_C char get_is_bundle_write_done(){
   return is_bundle_write_done;
+}
+extern EXT_C void set_is_bundle_write_done(uint8_t val){
+  is_bundle_write_done = val;
 }
diff --git a/test/sv/dma.sv b/test/sv/dma.sv
@@ -18,13 +18,14 @@ module DMA_M2S #(
   logic [BYTES_PER_BEAT-1:0][7:0] s_data_val;
   logic [BYTES_PER_BEAT-1:0] s_keep_val;
 
-  int status, i_bytes=0;
+  int status;
+  longint unsigned i_bytes=0;
   bit prev_handshake=1; // data is released first
   bit prev_slast=0;
 
-  import "DPI-C" function byte get_byte_wx (int addr, int mode);
+  import "DPI-C" function byte get_byte (longint unsigned addr);
 
-  task axis_push (input int offset, input int bytes_per_transfer);
+  task axis_push (input longint unsigned base_addr, input int bytes_per_transfer);
     {s_valid, s_data, s_last, s_keep} = '0;
 
     wait(aresetn); // wait for slave to begin
@@ -38,7 +39,7 @@ module DMA_M2S #(
             s_keep_val[i] = 0;
           end
           else begin
-            s_data_val[i] = get_byte_wx(offset + i_bytes, MODE);
+            s_data_val[i] = get_byte(base_addr + i_bytes);
             // $display("DMA: start:%d, i_bytes:%d, val:%d", offset, i_bytes, $signed(s_data_val[i]));
             s_keep_val[i] = 1;
             i_bytes  += 1;

diff --git a/test/sv/dnn_engine_tb.sv b/test/sv/dnn_engine_tb.sv
@@ -50,32 +50,34 @@ module dnn_engine_tb;
   DMA_M2S #(S_WEIGHTS_WIDTH_LF, VALID_PROB, 0) source_k (aclk, aresetn, s_axis_weights_tready, s_axis_weights_tvalid, s_axis_weights_tlast, s_axis_weights_tdata, s_axis_weights_tkeep);
 
   bit y_done=0, x_done=0, w_done=0, bundle_read_done=0, bundle_write_done=0;
-  int w_offset=0, w_bpt=0, x_offset=0, x_bpt=0;
+  longint unsigned w_base=0, x_base=0;
+  int w_bpt=0, x_bpt=0;
 
-  import "DPI-C" function void load_x(inout bit x_done, bundle_read_done, inout int x_offset, x_bpt);
-  import "DPI-C" function void load_w(inout bit w_done, inout int w_offset, w_bpt);
+  import "DPI-C" function void load_x(inout bit x_done, bundle_read_done, inout longint unsigned x_base, inout int x_bpt);
+  import "DPI-C" function void load_w(inout bit w_done, inout longint unsigned w_base, inout int w_bpt);
   import "DPI-C" function void load_y(inout bit y_done, inout bit m_t_done_proc, inout bit [31:0] y_sram [ROWS*COLS-1:0]);
-  import "DPI-C" function void fill_memory();
-  import "DPI-C" function byte get_byte_wx (int addr, int mode);
+  import "DPI-C" function void fill_memory(inout longint unsigned w_base, x_base);
+  import "DPI-C" function byte get_byte (longint unsigned addr);
   import "DPI-C" function byte get_is_bundle_write_done();
+  import "DPI-C" function void set_is_bundle_write_done(input bit val);
 
 
   // W DMA
   initial 
     while (1) begin
-      load_w (w_done, w_offset, w_bpt);
-      source_k.axis_push(w_offset, w_bpt);
-      $display("Done weights dma at offset=%d, bpt=%d \n", w_offset, w_bpt);
+      load_w (w_done, w_base, w_bpt);
+      source_k.axis_push(w_base, w_bpt);
+      $display("Done weights dma at offset=%h, bpt=%d \n", w_base, w_bpt);
       if (w_done) break;
     end
 
   // X DMA
   initial 
     while (1) begin
-      load_x (x_done, bundle_read_done, x_offset, x_bpt);
-      source_x.axis_push(x_offset, x_bpt);
+      load_x (x_done, bundle_read_done, x_base, x_bpt);
+      source_x.axis_push(x_base, x_bpt);
       while(bundle_read_done && !get_is_bundle_write_done()) #10ps;
-      $display("Done input dma at offset=%d, bpt=%d \n", x_offset, x_bpt);
+      $display("Done input dma at offset=%h, bpt=%d \n", x_base, x_bpt);
       if (x_done) break;
     end
 
@@ -111,12 +113,12 @@ module dnn_engine_tb;
   initial begin
     aresetn = 0;
 
-    fill_memory();
+    fill_memory(w_base, x_base);
 
     for (int i=0; i<50; i++)
-      $display("weights: i:%d, w:%b", i, get_byte_wx(i, 0));
+      $display("weights: i:%h, w:%b", i, get_byte(w_base + i));
     for (int i=0; i<10; i++)
-      $display("inputs: i:%d, w:%b", i, get_byte_wx(i,1));
+      $display("inputs : i:%h, w:%b", i, get_byte(x_base + i));
 
     repeat(2) @(posedge aclk) #1;
     aresetn = 1;