INT8 MLP example design for RTL/SystemC co-simulation

rad-sim/config.py

@@ -347,10 +347,10 @@ def prepare_build_dir(design_name):
     "noc_num_packet_types": [3],
     "noc_num_nodes": [0],
     "noc_max_num_router_dest_interfaces": 32,
-    "interfaces_max_axis_tdata_width": 1024,
+    "interfaces_max_axis_tdata_width": 512,
     "interfaces_axis_tkeep_width": 8,
     "interfaces_axis_tstrb_width": 8,
-    "interfaces_axis_tuser_width": 66,
+    "interfaces_axis_tuser_width": 75,
     "interfaces_axi_id_width": 8,
     "interfaces_axi_user_width": 64,
     "interfaces_max_axi_data_width": 512,

rad-sim/example-designs/mlp_int8/.gitignore

@@ -0,0 +1,6 @@
+compiler/input_mifs/*.mif
+compiler/inst_mifs/*.mif
+compiler/weight_mifs/*.mif
+golden_outputs.mif
+layer_mvm_config
+*.pdf

rad-sim/example-designs/mlp_int8/CMakeLists.txt

@@ -0,0 +1,52 @@
+cmake_minimum_required(VERSION 3.16)
+find_package(SystemCLanguage CONFIG REQUIRED)
+
+include_directories(
+  ./
+  modules
+  ../../sim
+  ../../sim/noc
+  ../../sim/noc/booksim
+  ../../sim/noc/booksim/networks
+  ../../sim/noc/booksim/routers
+)
+
+set(srcfiles
+  modules/fifo.cpp
+  modules/register_file.cpp
+  modules/sim_utils.cpp
+  modules/instructions.cpp
+  modules/dispatcher.cpp
+  modules/collector.cpp
+  modules/datapath.cpp
+  modules/pipeline.cpp
+  modules/mvm.cpp
+  modules/weight_loader.cpp
+  modules/inst_loader.cpp
+  mlp_top.cpp
+  mlp_driver.cpp
+  mlp_int8_system.cpp
+)
+
+set(hdrfiles
+  modules/fifo.hpp
+  modules/register_file.hpp
+  modules/sim_utils.hpp
+  modules/instructions.hpp
+  modules/dispatcher.hpp
+  modules/collector.hpp
+  modules/datapath.hpp
+  modules/pipeline.hpp
+  modules/mvm.hpp
+  modules/weight_loader.hpp
+  modules/inst_loader.hpp
+  modules/config.hpp
+  mlp_top.hpp
+  mlp_driver.hpp
+  mlp_int8_system.hpp
+)
+
+add_compile_options(-Wall -Wextra -pedantic)
+
+add_library(design STATIC ${srcfiles} ${hdrfiles})
+target_link_libraries(design PUBLIC SystemC::systemc booksim noc)

rad-sim/example-designs/mlp_int8/compiler/.gitignore

@@ -0,0 +1,2 @@
+*.mif
+layer_mvm_config

rad-sim/example-designs/mlp_int8/compiler/gen_testcase.py

@@ -0,0 +1,244 @@
+import sys
+import math
+import os
+import glob
+import numpy as np
+
+# Defining constants
+native_dim = 64
+num_test_inputs = 1
+num_noc_routers = 16
+
+if ('-h' in sys.argv or '--help' in sys.argv):
+  print('python gen_testcase.py <num_layers> <input_dim> [<hidden_dims>] [<mvms_per_layer>]')
+  exit(1)
+
+# Parse command line arguments
+num_layers = int(sys.argv[1])
+print('Number of layers = ' + str(num_layers))
+input_dim = int(sys.argv[2])
+print('Input dimension = ' + str(input_dim))
+hidden_dims = np.zeros(shape=num_layers, dtype=int)
+print('Hidden dimensions = [ ', end='')
+for i in range(3, 3 + num_layers):
+  hidden_dims[i-3] = int(sys.argv[i])
+  print(str(hidden_dims[i-3]) + ' ', end='')
+print(']')
+num_mvms = np.ones(shape=num_layers, dtype=int)
+print('Number of MVMs = [ ', end='')
+for i in range(3 + num_layers, 3 + 2 * num_layers):
+  num_mvms[i-3-num_layers] = int(sys.argv[i])
+  print(str(num_mvms[i-3-num_layers]) + ' ', end='')
+print(']')
+
+# Generate placement and clock constraint files
+placement_file = open('../mlp.place', 'w')
+clocks_file = open('../mlp.clks', 'w')
+placement_dict = {}
+
+total_num_blocks = 0
+for mvm_count in num_mvms:
+  total_num_blocks = total_num_blocks + mvm_count
+total_num_blocks = total_num_blocks + num_mvms[0] + 3
+if (total_num_blocks > num_noc_routers):
+  print("Number of blocks is more than NoC routers!")
+  exit(1)
+router_ids = np.random.choice(num_noc_routers, size=total_num_blocks, replace=False)
+
+idx = 0
+for l in range(num_layers):
+  for m in range(num_mvms[l]):
+    mvm_name = 'layer' + str(l) + '_mvm' + str(m)
+    placement_file.write(mvm_name + ' 0 ' + str(router_ids[idx]) + ' axis\n')
+    placement_dict[mvm_name] = router_ids[idx]
+    idx = idx + 1
+    clocks_file.write(mvm_name + ' 0 0\n')
+
+for m in range(num_mvms[0]):
+  dispatcher_name = 'input_dispatcher' + str(m)
+  placement_file.write(dispatcher_name + ' 0 ' + str(router_ids[idx]) + ' axis\n')
+  placement_dict[dispatcher_name] = router_ids[idx]
+  idx = idx + 1
+  clocks_file.write(dispatcher_name + ' 0 0\n')
+
+placement_file.write('output_collector 0 ' + str(router_ids[idx]) + ' axis\n')
+placement_dict['output_collector'] = router_ids[idx]
+idx = idx + 1
+clocks_file.write('output_collector 0 0\n')
+
+placement_file.write('weight_loader 0 ' + str(router_ids[idx]) + ' axis\n')
+placement_dict['weight_loader'] = router_ids[idx]
+idx = idx + 1
+clocks_file.write('weight_loader 0 0\n')
+
+placement_file.write('inst_loader 0 ' + str(router_ids[idx]) + ' axis\n')
+placement_dict['inst_loader'] = router_ids[idx]
+idx = idx + 1
+clocks_file.write('inst_loader 0 0\n')
+
+placement_file.close()
+clocks_file.close()
+
+# Generate random padded weight matrices
+padded_weights = []
+for l in range(num_layers):
+  num_mvms_in = num_mvms[l]
+  if (l == num_layers-1):
+    num_mvms_out = 1 #num_mvms[0]
+  else:
+    num_mvms_out = num_mvms[l+1]
+  if (l == 0):
+    layer_input_dim = input_dim
+  else:
+    layer_input_dim = hidden_dims[l-1]
+  padded_dimx = int(math.ceil(layer_input_dim * 1.0 / native_dim / num_mvms_in) * native_dim * num_mvms_in)
+  padded_dimy = int(math.ceil(hidden_dims[l] * 1.0 / native_dim / num_mvms_out) * native_dim * num_mvms_out)
+  padded_weights.append(np.zeros(shape=(padded_dimy, padded_dimx), dtype=int))
+  padded_weights[l][:hidden_dims[l], :layer_input_dim] = np.random.randint(-2, 2, size=(hidden_dims[l], layer_input_dim))
+
+# Prepare weight MIFs directory
+if(not(os.path.exists('./weight_mifs'))):
+  os.mkdir('weight_mifs')
+else:
+  files = glob.glob('weight_mifs/*.mif')
+  for file in files:
+    os.remove(file)
+
+# Write weight MIFs
+for l in range(num_layers):
+  layer_mvms = num_mvms[l]
+  mvm_idx = 0
+  limx = int(padded_weights[l].shape[1] / native_dim)
+  limy = int(padded_weights[l].shape[0] / native_dim)
+  mifs = []
+  for m in range(layer_mvms):
+    mifs.append([])
+    for d in range(native_dim):
+      mifs[m].append(open('weight_mifs/layer'+str(l)+'_mvm'+str(m)+'_dot'+str(d)+'.mif', 'w'))
+
+  for i in range(limx):
+    for j in range(limy):
+      for d in range(native_dim):
+        for e in range(native_dim):
+          mifs[mvm_idx][d].write(str(padded_weights[l][(j * native_dim) + d][(i * native_dim) + e]) + ' ')
+        mifs[mvm_idx][d].write('\n')
+    if (mvm_idx == layer_mvms-1):
+      mvm_idx = 0
+    else:
+      mvm_idx = mvm_idx + 1
+
+  for mvm_mifs in mifs:
+    for mif in mvm_mifs:
+      mif.close()
+
+# Prepare instruction MIFs directory
+if(not(os.path.exists('./inst_mifs'))):
+  os.mkdir('inst_mifs')
+else:
+  files = glob.glob('inst_mifs/*.mif')
+  for file in files:
+    os.remove(file)
+
+# Generate instruction MIFs
+# release_op, release_dest, rf_raddr, accum_raddr, last, release, accum_en, reduce
+for l in range(num_layers):
+  layer_mvms = num_mvms[l]
+  limx = int(padded_weights[l].shape[1] / native_dim / layer_mvms)
+  limy = int(padded_weights[l].shape[0] / native_dim)
+  for m in range(layer_mvms):
+    inst_mif = open('inst_mifs/layer'+str(l)+'_mvm'+str(m)+'.mif', 'w')
+    for i in range(limx):
+      for j in range(limy):
+        # release_op
+        if (m == layer_mvms - 1):
+          inst_mif.write('1 ')
+        else:
+          inst_mif.write('0 ')
+        # release_dest
+        if ((l == num_layers - 1) and (m == layer_mvms - 1)):
+          inst_mif.write(str(placement_dict['output_collector']) + ' ')
+        elif (m == layer_mvms - 1):
+          dest_layer = l + 1
+          dest_mvm = j % num_mvms[l+1]
+          dest_str = 'layer' + str(dest_layer) + '_mvm' + str(dest_mvm)
+          inst_mif.write(str(placement_dict[dest_str]) + ' ')
+        else:
+          dest_layer = l
+          dest_mvm = m + 1
+          dest_str = 'layer' + str(dest_layer) + '_mvm' + str(dest_mvm)
+          inst_mif.write(str(placement_dict[dest_str]) + ' ')
+        # rf_raddr
+        inst_mif.write(str(i * limy + j) + ' ')
+        # accum_raddr
+        inst_mif.write(str(j) + ' ')
+        # last
+        if (j == limy-1):
+          inst_mif.write('1 ')
+        else:
+          inst_mif.write('0 ')
+        # release
+        if (i == limx-1):
+          inst_mif.write('1 ')
+        else:
+          inst_mif.write('0 ')
+        # accum_en
+        if (i == 0):
+          inst_mif.write('0 ')
+        else:
+          inst_mif.write('1 ')
+        # reduce
+        if (m == 0 or i < limx-1):
+          inst_mif.write('0\n')
+        else:
+          inst_mif.write('1\n')
+    inst_mif.close()
+
+# Prepare input MIFs directory
+if(not(os.path.exists('./input_mifs'))):
+  os.mkdir('input_mifs')
+else:
+  files = glob.glob('input_mifs/*.mif')
+  for file in files:
+    os.remove(file)
+
+# Generate test input MIFs
+padded_input_dim = int(math.ceil(input_dim * 1.0 / native_dim / num_mvms[0]) * native_dim * num_mvms[0])
+test_inputs = np.zeros(shape=(num_test_inputs, padded_input_dim), dtype=int)
+test_inputs[:, :input_dim] = np.random.randint(-2, 2, size=(num_test_inputs, input_dim))
+input_files = []
+for i in range(num_mvms[0]):
+  input_files.append(open('input_mifs/inputs_mvm' + str(i) + '.mif', 'w'))
+for i in range(num_test_inputs):
+  for c in range(int(padded_input_dim / native_dim)):
+    for e in range(native_dim):
+      input_files[c % num_mvms[0]].write(str(test_inputs[i][(c * native_dim) + e]) + ' ')
+    input_files[c % num_mvms[0]].write('\n')
+for file in input_files:
+  file.close()
+
+# Compute test outputs
+test_inputs = np.transpose(test_inputs)
+test_outputs = np.dot(padded_weights[0], test_inputs)
+test_outputs = test_outputs.astype(np.int8)
+#test_outputs = np.maximum(test_outputs, np.zeros(shape=test_outputs.shape, dtype=int))
+for l in range(1, num_layers):
+  test_outputs = np.dot(padded_weights[l], test_outputs)
+  test_outputs = test_outputs.astype(np.int8)
+  #test_outputs = np.maximum(test_outputs, np.zeros(shape=test_outputs.shape, dtype=int))
+test_outputs = np.transpose(test_outputs)
+
+# Generate test output MIFs
+output_file = open('./golden_outputs.mif', 'w')
+for o in range(test_outputs.shape[0]):
+  for c in range(int(test_outputs.shape[1] / native_dim)):
+    for e in range(native_dim):
+      output_file.write(str(test_outputs[o][(c * native_dim) + e]) + ' ')
+    output_file.write('\n')
+output_file.close()
+
+# Generate layer/MVM configuration
+config_file = open('./layer_mvm_config', 'w')
+config_file.write(str(num_layers) + ' ')
+for mvm_count in num_mvms:
+  config_file.write(str(mvm_count) + ' ')
+config_file.close()

rad-sim/example-designs/mlp_int8/compiler/input_mifs/.gitignore

@@ -0,0 +1 @@
+*.mif

rad-sim/example-designs/mlp_int8/compiler/input_mifs/README.md

@@ -0,0 +1 @@
+Directory for dumping input MIFs

rad-sim/example-designs/mlp_int8/compiler/inst_mifs/.gitignore

@@ -0,0 +1 @@
+*.mif

rad-sim/example-designs/mlp_int8/compiler/inst_mifs/README.md

@@ -0,0 +1 @@
+Directory for dumping instruction MIFs

rad-sim/example-designs/mlp_int8/compiler/weight_mifs/.gitignore

@@ -0,0 +1 @@
+*.mif

rad-sim/example-designs/mlp_int8/compiler/weight_mifs/README.md

@@ -0,0 +1 @@
+Directory for dumping weight MIFs

rad-sim/example-designs/mlp_int8/config.yml

@@ -0,0 +1,37 @@
+noc:
+  type: ['2d']
+  num_nocs: 1
+  clk_period: [1.0]
+  payload_width: [166]
+  topology: ['mesh']
+  dim_x: [4]
+  dim_y: [4] 
+  routing_func: ['dim_order']
+  vcs: [5]
+  vc_buffer_size: [8]
+  output_buffer_size: [8]
+  num_packet_types: [5]
+  router_uarch: ['iq']
+  vc_allocator: ['islip']
+  sw_allocator:  ['islip']
+  credit_delay: [1]
+  routing_delay: [1]
+  vc_alloc_delay: [1]
+  sw_alloc_delay: [1]
+
+noc_adapters:
+  clk_period: [1.25]
+  fifo_size: [16]
+  obuff_size: [2]
+  in_arbiter: ['fixed_rr']
+  out_arbiter: ['priority_rr']
+  vc_mapping: ['direct']
+
+design:
+  name: 'mlp_int8'
+  noc_placement: ['mlp.place']
+  clk_periods: [5.0]
+
+telemetry:
+  log_verbosity: 2
+  traces: []

rad-sim/example-designs/mlp_int8/mlp.clks

@@ -0,0 +1,10 @@
+layer0_mvm0 0 0
+layer0_mvm1 0 0
+layer1_mvm0 0 0
+layer1_mvm1 0 0
+layer2_mvm0 0 0
+input_dispatcher0 0 0
+input_dispatcher1 0 0
+output_collector 0 0
+weight_loader 0 0
+inst_loader 0 0