Fix H-padding

KastnerRG · Oct 24, 2023 · 760448d · 760448d
1 parent 2f79ab8
commit 760448d
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 17 deletions.
diff --git a/c/model.h b/c/model.h
@@ -1,12 +1,12 @@
 #define N_BUNDLES 7
 Bundle_t bundles [N_BUNDLES] = {
-   {.n=8, .l=2, .kw=11, .coe=2, .coe_tl=2, .r_ll=8, .h=16, .w=8, .ci=3, .co=16, .w_kw2=3, .t=8, .p=3, .cm=1, .cm_p0=1, .w_bpt=272, .w_bpt_p0=272, .x_bpt=840, .x_bpt_p0=840, .is_bias=1, .b_offset=0, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=0, .ca_shift=12, .ca_pl_scale=0, .x_header=414341061322735616, .x_header_p0=414341061322735616, .w_header=414587437826703360, .w_header_p0=414341061322735616 },
-   {.n=8, .l=2, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=16, .w=8, .ci=16, .co=16, .w_kw2=8, .t=1, .p=1, .cm=20, .cm_p0=16, .w_bpt=392, .w_bpt_p0=392, .x_bpt=13320, .x_bpt_p0=13320, .is_bias=0, .b_offset=16, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=7, .ca_pl_scale=0, .x_header=8700964375684448256, .x_header_p0=8700964375684448256, .w_header=8701210795138088960, .w_header_p0=8700964375684448256 },
-   {.n=8, .l=2, .kw=7, .coe=3, .coe_tl=4, .r_ll=8, .h=16, .w=8, .ci=16, .co=16, .w_kw2=5, .t=6, .p=8, .cm=2, .cm_p0=2, .w_bpt=344, .w_bpt_p0=344, .x_bpt=1672, .x_bpt_p0=1672, .is_bias=1, .b_offset=16, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=1, .ca_shift=12, .ca_pl_scale=0, .x_header=846686625550303232, .x_header_p0=846686625550303232, .w_header=846933027824074752, .w_header_p0=846686625550303232 },
-   {.n=8, .l=2, .kw=5, .coe=4, .coe_tl=4, .r_ll=8, .h=16, .w=8, .ci=16, .co=16, .w_kw2=6, .t=4, .p=4, .cm=4, .cm_p0=4, .w_bpt=488, .w_bpt_p0=488, .x_bpt=3336, .x_bpt_p0=3336, .is_bias=0, .b_offset=34, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=10, .ca_pl_scale=3, .x_header=1927550536119222272, .x_header_p0=1927550536119222272, .w_header=1927796989932601344, .w_header_p0=1927550536119222272 },
-   {.n=8, .l=2, .kw=3, .coe=8, .coe_tl=8, .r_ll=8, .h=16, .w=8, .ci=16, .co=24, .w_kw2=7, .t=3, .p=3, .cm=6, .cm_p0=4, .w_bpt=440, .w_bpt_p0=296, .x_bpt=5000, .x_bpt_p0=3336, .is_bias=1, .b_offset=34, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=0, .ca_shift=12, .ca_pl_scale=0, .x_header=3008414446688141312, .x_header_p0=1855492942081294336, .w_header=3008660883321651200, .w_header_p0=1855492942081294336 },
-   {.n=8, .l=2, .kw=1, .coe=24, .coe_tl=2, .r_ll=8, .h=16, .w=8, .ci=24, .co=50, .w_kw2=8, .t=3, .p=2, .cm=20, .cm_p0=4, .w_bpt=488, .w_bpt_p0=104, .x_bpt=16648, .x_bpt_p0=3336, .is_bias=0, .b_offset=58, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=10, .ca_pl_scale=3, .x_header=11006807384898142208, .x_header_p0=1783435348043366400, .w_header=11007053838711521280, .w_header_p0=1783435348043366400 },
-   {.n=1, .l=1, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=8, .w=1, .ci=6400, .co=10, .w_kw2=1, .t=1, .p=320, .cm=20, .cm_p0=20, .w_bpt=488, .w_bpt_p0=488, .x_bpt=138, .x_bpt_p0=138, .is_bias=1, .b_offset=58, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=1, .ca_shift=15, .ca_pl_scale=3, .x_header=10952754293765046272, .x_header_p0=10952754293765046272, .w_header=10952754456973803520, .w_header_p0=10952754293765046272 }
+   {.n=8, .l=2, .kw=11, .coe=2, .coe_tl=2, .r_ll=2, .h=10, .w=8, .ci=3, .co=16, .w_kw2=3, .t=8, .p=3, .cm=1, .cm_p0=1, .w_bpt=272, .w_bpt_p0=272, .x_bpt=840, .x_bpt_p0=840, .is_bias=1, .b_offset=0, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=0, .ca_shift=12, .ca_pl_scale=0, .x_header=414341061322735616, .x_header_p0=414341061322735616, .w_header=414587437826703360, .w_header_p0=414341061322735616 },
+   {.n=8, .l=2, .kw=1, .coe=24, .coe_tl=0, .r_ll=2, .h=10, .w=8, .ci=16, .co=16, .w_kw2=8, .t=1, .p=1, .cm=20, .cm_p0=16, .w_bpt=392, .w_bpt_p0=392, .x_bpt=13320, .x_bpt_p0=13320, .is_bias=0, .b_offset=16, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=7, .ca_pl_scale=0, .x_header=8700964375684448256, .x_header_p0=8700964375684448256, .w_header=8701210795138088960, .w_header_p0=8700964375684448256 },
+   {.n=8, .l=2, .kw=7, .coe=3, .coe_tl=4, .r_ll=2, .h=10, .w=8, .ci=16, .co=16, .w_kw2=5, .t=6, .p=8, .cm=2, .cm_p0=2, .w_bpt=344, .w_bpt_p0=344, .x_bpt=1672, .x_bpt_p0=1672, .is_bias=1, .b_offset=16, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=1, .ca_shift=12, .ca_pl_scale=0, .x_header=846686625550303232, .x_header_p0=846686625550303232, .w_header=846933027824074752, .w_header_p0=846686625550303232 },
+   {.n=8, .l=2, .kw=5, .coe=4, .coe_tl=4, .r_ll=2, .h=10, .w=8, .ci=16, .co=16, .w_kw2=6, .t=4, .p=4, .cm=4, .cm_p0=4, .w_bpt=488, .w_bpt_p0=488, .x_bpt=3336, .x_bpt_p0=3336, .is_bias=0, .b_offset=34, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=10, .ca_pl_scale=3, .x_header=1927550536119222272, .x_header_p0=1927550536119222272, .w_header=1927796989932601344, .w_header_p0=1927550536119222272 },
+   {.n=8, .l=2, .kw=3, .coe=8, .coe_tl=8, .r_ll=2, .h=10, .w=8, .ci=16, .co=24, .w_kw2=7, .t=3, .p=3, .cm=6, .cm_p0=4, .w_bpt=440, .w_bpt_p0=296, .x_bpt=5000, .x_bpt_p0=3336, .is_bias=1, .b_offset=34, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=0, .ca_shift=12, .ca_pl_scale=0, .x_header=3008414446688141312, .x_header_p0=1855492942081294336, .w_header=3008660883321651200, .w_header_p0=1855492942081294336 },
+   {.n=8, .l=2, .kw=1, .coe=24, .coe_tl=2, .r_ll=2, .h=10, .w=8, .ci=24, .co=50, .w_kw2=8, .t=3, .p=2, .cm=20, .cm_p0=4, .w_bpt=488, .w_bpt_p0=104, .x_bpt=16648, .x_bpt_p0=3336, .is_bias=0, .b_offset=58, .b_val_shift=0, .b_bias_shift=0, .ca_nzero=1, .ca_shift=10, .ca_pl_scale=3, .x_header=11006807384898142208, .x_header_p0=1783435348043366400, .w_header=11007053838711521280, .w_header_p0=1783435348043366400 },
+   {.n=1, .l=1, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=8, .w=1, .ci=4000, .co=10, .w_kw2=1, .t=1, .p=200, .cm=20, .cm_p0=20, .w_bpt=488, .w_bpt_p0=488, .x_bpt=138, .x_bpt_p0=138, .is_bias=1, .b_offset=58, .b_val_shift=5, .b_bias_shift=0, .ca_nzero=1, .ca_shift=15, .ca_pl_scale=3, .x_header=10952754293765046272, .x_header_p0=10952754293765046272, .w_header=10952754456973803520, .w_header_p0=10952754293765046272 }
 };
 
 #define X_BITS_L2   2
@@ -16,10 +16,10 @@ Bundle_t bundles [N_BUNDLES] = {
 #define PE_ROWS     8
 #define PE_COLS     24
 
-#define WB_BYTES    192868
-#define W_BYTES     192704
+#define WB_BYTES    134308
+#define W_BYTES     134144
 #define X_BYTES     2520
-#define X_BYTES_ALL 120040
+#define X_BYTES_ALL 103480
 #define Y_BYTES     294920
 #define B_TYPE      signed short
 #define B_WORDS     82

diff --git a/c/runtime.h b/c/runtime.h
@@ -23,6 +23,7 @@ typedef struct {
   B_TYPE b  [B_WORDS     ]; // keep next to w. weights are loaded to w_ptr
   char   x  [X_BYTES_ALL ];
   int    y  [Y_BYTES/4   ];
+  int p_sum [Y_BYTES/4   ];
 } Memory_st;
 Memory_st mem;
 
@@ -43,12 +44,12 @@ static inline void process_y(int val, int i_py, Bundle_t *p_bundle, int ip, int
   // ------ ADD P PASSES ------ 
   if (p_bundle->p == 1) {          // only p  : proceed with value
   } else if (ip == p_bundle->p-1) {// last p  : read, add, proceed
-    val += mem.y[i_py];
+    val += mem.p_sum[i_py];
   } else if (ip == 0) {            // first p : overwrite memory, return
-    mem.y[i_py] = val;
+    mem.p_sum[i_py] = val;
     return;
   } else {                         // middle p: read, add, store, return
-    mem.y[i_py] += val;
+    mem.p_sum[i_py] += val;
     return;
   }
 
@@ -95,10 +96,17 @@ extern EXT_C void load_y (unsigned char *p_done, unsigned char *pt_done_proc,  c
         for (int ir=0; ir<PE_ROWS; ir++) {
           // Index: [b, p, t, l, n, w | coe, w_last, r]
 
-          int val = p_sram[sram_addr];
-          fprintf(fp,"%d\n", val);
 
-          process_y(val, i_py, p_bundle, ip, i_bias);
+          int i_yh = il*PE_ROWS + ir;
+
+          if (i_yh < p_bundle->h) {
+            int val = p_sram[sram_addr];
+            fprintf(fp,"%d\n", val);
+            process_y(val, i_py, p_bundle, ip, i_bias);
+          } else {
+            mem.y[i_py] = 0;
+            fprintf(fp,"%d\n", 0);
+          }
 
           i_py += 1;
           sram_addr += 1;

diff --git a/test/py/param_test.py b/test/py/param_test.py
@@ -189,7 +189,7 @@ class Config:
 def test_dnn_engine(COMPILE):
     c = make_compile_params(COMPILE)
 
-    input_shape = (8,16,8,3) # (XN, XH, XW, CI)
+    input_shape = (8,10,8,3) # (XN, XH, XW, CI)
     model_config = [
         Config(11, 16, True , f'quantized_relu({c.X_BITS},0,negative_slope=0)'),
         Config(1 , 16, False, f'quantized_bits({c.X_BITS},0,False,False,1)'),