Fix bits=8

KastnerRG · Oct 19, 2023 · 2f79ab8 · 2f79ab8
1 parent 0b4fe23
commit 2f79ab8
Show file tree

Hide file tree

Showing 6 changed files with 383 additions and 60 deletions.
diff --git a/c/model.h b/c/model.h
@@ -1,23 +1,23 @@
 #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=140, .w_bpt_p0=140, .x_bpt=840, .x_bpt_p0=840, .is_bias=1, .b_offset=0, .b_val_shift=9, .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=200, .w_bpt_p0=200, .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=3, .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=176, .w_bpt_p0=176, .x_bpt=1672, .x_bpt_p0=1672, .is_bias=1, .b_offset=16, .b_val_shift=9, .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=248, .w_bpt_p0=248, .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=6, .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=224, .w_bpt_p0=152, .x_bpt=5000, .x_bpt_p0=3336, .is_bias=1, .b_offset=34, .b_val_shift=9, .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=248, .w_bpt_p0=56, .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=6, .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=248, .w_bpt_p0=248, .x_bpt=138, .x_bpt_p0=138, .is_bias=1, .b_offset=58, .b_val_shift=9, .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=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 }
 };
 
 #define X_BITS_L2   2
-#define W_BITS_L2   2
+#define W_BITS_L2   3
 #define X_PAD       5
 #define KH_MAX      11
 #define PE_ROWS     8
 #define PE_COLS     24
 
-#define WB_BYTES    98212
-#define W_BYTES     98048
+#define WB_BYTES    192868
+#define W_BYTES     192704
 #define X_BYTES     2520
 #define X_BYTES_ALL 120040
 #define Y_BYTES     294920

diff --git a/fpga/scripts/vivado_config.tcl b/fpga/scripts/vivado_config.tcl
@@ -4,7 +4,7 @@
     set ROWS               8
     set COLS               24
     set X_BITS             4
-    set K_BITS             4
+    set K_BITS             8
     set Y_BITS             24
     set DELAY_W_RAM        2
     set RAM_EDGES_DEPTH    288

diff --git a/rtl/include/params_input.svh b/rtl/include/params_input.svh
@@ -4,7 +4,7 @@
     `define ROWS                8                 	// PE rows, constrained by resources
     `define COLS                24                 	// PE cols, constrained by resources
     `define X_BITS              4               	// Bits per word in input
-    `define K_BITS              4               	// Bits per word in input
+    `define K_BITS              8               	// Bits per word in input
     `define Y_BITS              24               	// Bits per word in output of conv
 
     `define KH_MAX              11               	// max of kernel height, across layers

diff --git a/test/py/bundle.py b/test/py/bundle.py
@@ -706,4 +706,4 @@ def pack_words_into_bytes (arr, bits):
         arr = arr.reshape(arr.size//w_words_per_byte, w_words_per_byte)
         for i_word in range(1, w_words_per_byte):
             arr[:,0] += arr[:,i_word] << (i_word * bits) # pack multiple words into a byte
-        return arr[:,0] # packed byte
+        return arr[:,0].astype(np.uint8) # packed byte
diff --git a/test/py/param_test.py b/test/py/param_test.py
@@ -167,7 +167,7 @@ class Config:
 
 @pytest.mark.parametrize("COMPILE", list(product_dict(
                                                 X_BITS     = [4    ], 
-                                                K_BITS     = [4    ], 
+                                                K_BITS     = [8    ], 
                                                 B_BITS     = [16   ], 
                                                 Y_BITS     = [24   ], 
                                                 INT_BITS   = [32   ], # size of integer in target CPU
@@ -183,17 +183,17 @@ class Config:
                                                 RAM_WEIGHTS_DEPTH = [20  ],  # KH*CI + Config beats
                                                 RAM_EDGES_DEPTH   = [288 ], # max(CI * XW * (XH/ROWS-1))
 
-                                                VALID_PROB = [100],
-                                                READY_PROB = [1],
+                                                VALID_PROB = [1],
+                                                READY_PROB = [100],
                                             )))
 def test_dnn_engine(COMPILE):
     c = make_compile_params(COMPILE)
 
     input_shape = (8,16,8,3) # (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.K_BITS},0,False,False,1)'),
-        Config(7 , 16, True , f'quantized_bits({c.K_BITS},0,False,True,1)'),
+        Config(1 , 16, False, f'quantized_bits({c.X_BITS},0,False,False,1)'),
+        Config(7 , 16, True , f'quantized_bits({c.X_BITS},0,False,True,1)'),
         Config(5 , 16, False, f'quantized_relu({c.X_BITS},0,negative_slope=0.125)'),
         Config(3 , 24, True , f'quantized_relu({c.X_BITS},0,negative_slope=0)'),
         Config(1 , 50, False, f'quantized_relu({c.X_BITS},0,negative_slope=0.125)', flatten=True),