Portwidth, Elementwise, Pooling improvements (#8)

* fix data overlap in mem_write * enable passing DMA_WIDTH and DATA_WIDTH in partition_template * Enable passing port_width, DMA_WIDTH and DATA_WIDTH in GeneratePartition * rewrite _fixed_point_stream_to_dat helper function, allow port_width = 128, 256 and 512 * Add templete for common.hpp and common_tb.hpp * Update default common_tb.hpp * Allow passing port_width, format common.hpp and common_tb.hpp in GenerateNetwork * changed datatype in common.hpp * add library in common.hpp * implemented elementwise_add module * Add configuration file of elementwise_add * fix typo * fix hls template for no weight reloading * implement elementwise_add,elemenwise_mul,split layers and elementwis_add,elementwise_mul modules * implement testing files for elementwise_add,elementwise_mul,split layers and elementwise_add,elementwise_mul modules * Add configuration files for elementwise_mul module * Add configuration files for elementwise_add layer * Add configuration files for elementwise_mul layer * fix layer testing scripts * dev: modules and layers to support ResNet8 * testing files for new modules and layers * modify gitignore
AlexMontgomerie · Sep 20, 2024 · 81b6f27 · 81b6f27
1 parent 8624942
commit 81b6f27
Show file tree

Hide file tree

Showing 96 changed files with 4,081 additions and 124 deletions.
diff --git a/fpgaconvnet/hls/generate/layers/avg_pooling.py b/fpgaconvnet/hls/generate/layers/avg_pooling.py
@@ -0,0 +1,194 @@
+import fpgaconvnet.hls.generate.modules.sliding_window as generate_sliding_window
+import fpgaconvnet.hls.generate.modules.avg_pool as generate_avg_pool
+
+avg_pooling_layer_template_header = """#ifndef {NAME}_HPP_
+#define {NAME}_HPP_
+
+#include "sliding_window.hpp"
+#include "avg_pool.hpp"
+
+#define name        {name}
+#define NAME        {NAME}
+#define {NAME}_ID   {id}
+
+#define {NAME}_BATCH_SIZE    {batch_size}
+#define {NAME}_ROWS          {rows}
+#define {NAME}_COLS          {cols}
+#define {NAME}_CHANNELS      {channels}
+#define {NAME}_COARSE        {coarse}
+#define {NAME}_KERNEL_SIZE_X {kernel_size_x}
+#define {NAME}_KERNEL_SIZE_Y {kernel_size_y}
+#define {NAME}_STRIDE_X      {stride_x}
+#define {NAME}_STRIDE_Y      {stride_y}
+#define {NAME}_FINE          {fine}
+
+#define {NAME}_COARSE_IN    {NAME}_COARSE
+#define {NAME}_COARSE_OUT   {NAME}_COARSE
+
+#define {NAME}_ROWS_OUT     {rows_out}
+#define {NAME}_COLS_OUT     {cols_out}
+#define {NAME}_CHANNELS_OUT {channels_out}
+
+// define the data type
+typedef ap_fixed<{data_width},{data_int_width},AP_RND> {name}_data_t;
+typedef {name}_data_t {name}_input_t;
+typedef {name}_data_t {name}_output_t;
+
+// SLIDING WINDOW
+#define {NAME}_SLIDING_WINDOW_BATCH_SIZE    {batch_size}
+#define {NAME}_SLIDING_WINDOW_ROWS          {rows}
+#define {NAME}_SLIDING_WINDOW_COLS          {cols}
+#define {NAME}_SLIDING_WINDOW_CHANNELS      {channels_per_module}
+#define {NAME}_SLIDING_WINDOW_KERNEL_SIZE_X {kernel_size_x}
+#define {NAME}_SLIDING_WINDOW_KERNEL_SIZE_Y {kernel_size_y}
+#define {NAME}_SLIDING_WINDOW_STRIDE_X      {stride_x}
+#define {NAME}_SLIDING_WINDOW_STRIDE_Y      {stride_y}
+#define {NAME}_SLIDING_WINDOW_PAD_LEFT      {pad_left}
+#define {NAME}_SLIDING_WINDOW_PAD_RIGHT     {pad_right}
+#define {NAME}_SLIDING_WINDOW_PAD_TOP       {pad_top}
+#define {NAME}_SLIDING_WINDOW_PAD_BOTTOM    {pad_bottom}
+
+// AVG_POOL
+#define {NAME}_AVG_POOL_BATCH_SIZE   {batch_size}
+#define {NAME}_AVG_POOL_ROWS         {rows_out}
+#define {NAME}_AVG_POOL_COLS         {cols_out}
+#define {NAME}_AVG_POOL_CHANNELS     {channels_per_module}
+#define {NAME}_AVG_POOL_KERNEL_SIZE_X {kernel_size_x}
+#define {NAME}_AVG_POOL_KERNEL_SIZE_Y {kernel_size_y}
+#define {NAME}_AVG_POOL_FINE         {fine}
+
+/**
+ * FUNCTION DEFINITION
+ */
+ 
+void {name}(
+    stream_t({name}_data_t) in[{NAME}_COARSE],
+    stream_t({name}_data_t) out[{NAME}_COARSE], 
+    int mode
+);
+
+#undef name
+#undef NAME
+#endif
+"""
+
+avg_pooling_layer_template_src = """#include "{name}.hpp"
+
+void {name}_sliding_window(
+    stream_t({name}_data_t) &in,
+    stream_t({name}_data_t) out[{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y]
+) {{
+#pragma HLS INLINE OFF
+{sliding_window}
+}}
+
+void {name}_avg_pool(
+    stream_t({name}_data_t) in[{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y],
+    stream_t({name}_data_t) &out
+) {{
+#pragma HLS INLINE OFF
+{avg_pool}
+}}
+
+void {name}(
+    stream_t({name}_data_t) in[{NAME}_COARSE],
+    stream_t({name}_data_t) out[{NAME}_COARSE],
+    int mode
+)
+{{
+    
+#pragma HLS INLINE OFF
+#pragma HLS DATAFLOW
+
+#pragma HLS STREAM variable=in depth={buffer_depth}
+#pragma HLS STREAM variable=out 
+
+#pragma HLS ARRAY_PARTITION variable=in complete dim=0
+#pragma HLS ARRAY_PARTITION variable=out complete dim=0
+
+    stream_t({name}_data_t) sw_out[{NAME}_COARSE][{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y]; //sliding window output
+    
+#pragma HLS STREAM variable=sw_out
+#pragma HLS ARRAY_PARTITION variable=sw_out complete dim=0
+
+    for(unsigned int coarse_index=0; coarse_index<{NAME}_COARSE; coarse_index++)
+    {{
+#pragma HLS UNROLL
+        {name}_sliding_window(in[coarse_index],sw_out[coarse_index]);
+        {name}_avg_pool(sw_out[coarse_index],out[coarse_index]);
+    }}
+}}
+
+"""
+
+def gen_avg_pooling_layer(name,param,src_path,header_path):
+
+    # get sliding window type
+    single_channel = True if param['channels_in'] == 1 else False
+
+    # SLIDING WINDOW MODULE INIT
+    sliding_window = generate_sliding_window.gen_sliding_window_module(
+        name+"_sliding_window", 
+        "in", "out",
+        sliding_window_t=f"{name}_data_t",
+        indent=4
+    )
+
+    # AVG_POOL MODULE INIT
+    avg_pool = generate_avg_pool.gen_avg_pool_module(
+        name+"_avg_pool",
+        "in", 
+        "out",
+        avg_pool_t=f"{name}_data_t",
+        indent=4
+    )
+
+    # src
+    avg_pooling_layer_src = avg_pooling_layer_template_src.format(
+        name            =name,
+        NAME            =name.upper(),
+        # buffer_depth=max(param['buffer_depth'],2),
+        buffer_depth    =2,
+        sliding_window  =sliding_window,
+        avg_pool            =avg_pool
+    )
+
+    # header
+    avg_pooling_layer_header = avg_pooling_layer_template_header.format(
+        name                =name,
+        NAME                =name.upper(),
+        id                  =0, # param['id'],
+        batch_size          =param['batch_size'],
+        rows                =param['rows_in'],
+        cols                =param['cols_in'],
+        channels            =param['channels_in'],
+        channels_per_module =param['channels_in']//param['coarse'],
+        coarse              =param['coarse'],
+        kernel_size_x       =param['kernel_size'][0],
+        kernel_size_y       =param['kernel_size'][1],
+        stride_x            =param['stride'][0],
+        stride_y            =param['stride'][1],
+        pad_left            =param['pad_left'],
+        pad_right           =param['pad_right'],
+        pad_top             =param['pad_top'],
+        pad_bottom          =param['pad_bottom'],
+        fine                =param['fine'] if 'fine' in param else "1",
+        rows_out            =param['rows_out'],
+        cols_out            =param['cols_out'],
+        channels_out        =param['channels_out'],
+        data_width          =param['data_t']['width'],
+        data_int_width      =(param['data_t']['width']-param['data_t']['binary_point'])
+    )
+
+    # write source file
+    with open(src_path,'w') as src_file:
+        src_file.write(avg_pooling_layer_src)
+
+    # write header file
+    with open(header_path,'w') as header_file:
+        header_file.write(avg_pooling_layer_header)
+
+    return 
+
+
+
diff --git a/fpgaconvnet/hls/generate/layers/convolution.py b/fpgaconvnet/hls/generate/layers/convolution.py
@@ -147,7 +147,7 @@
 }}
 
 void {name}_fork(
-#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1
+#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1 && {NAME}_STRIDE_X == 1 && {NAME}_STRIDE_Y == 1
     stream_t({name}_input_t)  &in,
     stream_t({name}_output_t) out[{NAME}_COARSE_OUT]
 #else
@@ -162,7 +162,7 @@
 
 void {name}_conv(
     const {name}_weight_t weights[DIVIDE({NAME}_WEIGHTS,{NAME}_COARSE_IN*{NAME}_COARSE_GROUP*{NAME}_COARSE_OUT*{NAME}_KERNEL_SIZE_X*{NAME}_KERNEL_SIZE_Y)][{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y],
-#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1
+#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1 && {NAME}_STRIDE_X == 1 && {NAME}_STRIDE_Y == 1
     stream_t({name}_input_t) &in,
 #else
     stream_t({name}_input_t)  in[{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y],
@@ -230,7 +230,7 @@
     #pragma HLS ARRAY_PARTITION variable=sw_out complete dim=0
 #endif
 
-#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1
+#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1 && {NAME}_STRIDE_X == 1 && {NAME}_STRIDE_Y == 1
     stream_t({name}_input_t) fork_out[{NAME}_COARSE_IN*{NAME}_COARSE_GROUP][{NAME}_COARSE_OUT];
 #else
     stream_t({name}_input_t) fork_out[{NAME}_COARSE_IN*{NAME}_COARSE_GROUP][{NAME}_COARSE_OUT][{NAME}_KERNEL_SIZE_X][{NAME}_KERNEL_SIZE_Y];
@@ -256,7 +256,7 @@
 
     {name}_coarse_in_loop: for(unsigned int i=0;i<{NAME}_COARSE_IN*{NAME}_COARSE_GROUP;i++) {{
         #pragma HLS unroll
-#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1
+#if {NAME}_KERNEL_SIZE_X == 1 && {NAME}_KERNEL_SIZE_Y == 1 && {NAME}_STRIDE_X == 1 && {NAME}_STRIDE_Y == 1
         {name}_fork(in[i], fork_out[i]);
 #else
         {name}_sliding_window(in[i], sw_out[i]);
@@ -318,15 +318,17 @@ def gen_convolution_layer(name, param, src_path, header_path):
     )
 
     # FORK MODULE INIT
-    fork = generate_fork.gen_fork_module(
+    fork = generate_fork.gen_conv_fork_module(
+        name, 
         name+"_fork",
         "in", "out",
         fork_t=f"{name}_input_t",
         indent=4
     )
 
     # CONV MODULE INIT
-    conv = generate_conv.gen_conv_module(
+    conv = generate_conv.gen_convolution_conv_module(
+        name, 
         name+"_conv",
         "in", "weights", "out",
         data_t=f"{name}_input_t",

diff --git a/fpgaconvnet/hls/generate/layers/elementwise_add.py b/fpgaconvnet/hls/generate/layers/elementwise_add.py
@@ -0,0 +1,143 @@
+# import modules 
+import fpgaconvnet.hls.generate.modules.elementwise_add as generate_elementwise_add
+
+elementwise_add_layer_template_header = """#ifndef {NAME}_HPP_
+#define {NAME}_HPP_
+
+#include "elementwise_add.hpp"
+
+#define name        {name}
+#define NAME        {NAME}
+#define {NAME}_ID   {id}
+
+#define {NAME}_BATCH_SIZE   {batch_size}
+#define {NAME}_ROWS         {rows}
+#define {NAME}_COLS         {cols}
+#define {NAME}_CHANNELS     {channels}
+#define {NAME}_COARSE       {coarse}
+
+#define {NAME}_COARSE_IN    {NAME}_COARSE
+#define {NAME}_COARSE_OUT   {NAME}_COARSE
+
+#define {NAME}_ROWS_OUT     {rows_out}
+#define {NAME}_COLS_OUT     {cols_out}
+#define {NAME}_CHANNELS_OUT {channels_out}
+
+#define {NAME}_ELEMENTWISE_ADD_BATCH_SIZE   {batch_size}
+#define {NAME}_ELEMENTWISE_ADD_ROWS         {rows}
+#define {NAME}_ELEMENTWISE_ADD_COLS         {cols}
+#define {NAME}_ELEMENTWISE_ADD_CHANNELS     {channels_per_module}
+
+typedef ap_fixed<{data_width},{data_int_width},AP_RND> {name}_data_t;
+typedef {name}_data_t {name}_input_t;
+typedef {name}_data_t {name}_output_t;
+
+/**
+  * FUNCTION DEFINITION
+  */
+  
+void {name}(
+    stream_t({name}_data_t) in_1[{NAME}_COARSE],
+    stream_t({name}_data_t) in_2[{NAME}_COARSE],
+    stream_t({name}_data_t) out[{NAME}_COARSE],
+    int mode
+);
+
+#undef name
+#undef NAME
+#endif
+"""
+
+elementwise_add_layer_template_src = """#include "{name}.hpp"
+
+void {name}_elementwise_add(
+    stream_t({name}_data_t) &in_1,
+    stream_t({name}_data_t) &in_2,
+    stream_t({name}_data_t) &out
+) {{
+
+#pragma HLS INLINE OFF
+{elementwise_add}
+}}
+
+void {name}(
+    stream_t({name}_data_t) in_1[{NAME}_COARSE],
+    stream_t({name}_data_t) in_2[{NAME}_COARSE],
+    stream_t({name}_data_t) out[{NAME}_COARSE],
+    int mode
+) 
+{{
+    
+#pragma HLS INLINE OFF
+
+// depth may need to be adjusted to avoid deadlocks
+#pragma HLS STREAM variable=in_1 depth=16284
+#pragma HLS STREAM variable=in_2 depth=16284
+#pragma HLS STREAM variable=out 
+
+#pragma HLS ARRAY_PARTITION variable=in_1 complete dim=0
+#pragma HLS ARRAY_PARTITION variable=in_2 complete dim=0
+#pragma HLS ARRAY_PARTITION variable=out complete dim=0
+
+#pragma HLS DATAFLOW
+
+    for (unsigned int coarse_index=0; coarse_index<{NAME}_COARSE; coarse_index++)
+    {{
+#pragma HLS unroll
+        {name}_elementwise_add(in_1[coarse_index], in_2[coarse_index], out[coarse_index]);
+    }}
+}}
+
+"""
+
+def gen_elementwise_add_layer(name,param,src_path,header_path):
+
+    # ELEMENTWISE_ADD MODULE INIT
+    elementwise_add = generate_elementwise_add.gen_elementwise_add_module(
+        name+"_elementwise_add",
+        "in_1",
+        "in_2",
+        "out",
+        elementwise_add_t="data_t",
+        indent=4
+    )
+
+    # src
+    elementwise_add_layer_src = elementwise_add_layer_template_src.format(
+        name = name, 
+        NAME = name.upper(),
+        # buffer_depth=max(param['buffer_depth'],2),
+        elementwise_add = elementwise_add
+    )
+
+    # header
+    elementwise_add_layer_header = elementwise_add_layer_template_header.format(
+        name = name,
+        NAME = name.upper(),
+        id = 0, #param['id'],
+        batch_size = param['batch_size'],
+        rows = param['rows_in'],
+        cols = param['cols_in'],
+        channels = param['channels_in'],
+        channels_per_module = param['channels_in']//param['coarse_in'],
+        coarse = param['coarse_in'],
+        rows_out = param['rows_out'],
+        cols_out = param['cols_out'],
+        channels_out = param['channels_out'],
+        data_width = param['data_t']['width'],
+        data_int_width = param['data_t']['width']-param['data_t']['binary_point']
+        # data_width = 16,
+        # data_int_width = 8
+    )
+
+    # write source file 
+    with open(src_path,'w') as src_file:
+        src_file.write(elementwise_add_layer_src)
+
+    # write header file
+    with open(header_path,'w') as header_file:
+        header_file.write(elementwise_add_layer_header)
+
+    return
+
+