Merge pull request #29 from negishubham/cnn-support

New feature added - weight and input quantization.

include/constants.py

@@ -148,25 +148,41 @@
 adc_lat_dict = {'1' : 12.5,
                 '2' : 25,
                 '4' : 50,
+		'5' : 62.5,			
+		'6' : 75,
+		'7' : 87.5,
                 '8' : 100,
+		'9' : 112.5,
                 '16': 200}
 
 adc_pow_dyn_dict = {'1' : 0.225,
                     '2' : 0.45,
                     '4' : 0.9,
+		    '5' : 1.125,			
+		    '6' : 1.35,
+		    '7' : 1.575,
                     '8' : 1.8,
+		    '9' : 2.025,
                     '16': 3.6}
 
 adc_pow_leak_dict = {'1' : 0.025,
                      '2' : 0.05,
                      '4' : 0.1,
+		     '5' : 0.125,			
+		     '6' : 0.150,
+		     '7' : 0.175,
                      '8' : 0.2,
+		     '9' : 0.225,
                      '16': 0.4}
 
 adc_area_dict = {'1' : 0.0012,
                  '2' : 0.0012,
                  '4' : 0.0012,
+		 '5' : 0.0012,			
+		 '6' : 0.0012,
+		 '7' : 0.0012,
                  '8' : 0.0012,
+		 '9' : 0.0012,
                  '16': 0.0012}
 
 # SNH (MVM pipeline)

include/example-configs/config-cnn.py

@@ -28,9 +28,11 @@
 # Fixed parameters
 addr_width = 22 # Added to address larger address space for conv layers (#TODO: Compiler needs to fix shared memory reuse)
 data_width = num_bits # (in bits)
-xbdata_width = data_width # (in bits)
+xbdata_width = data_width # (in bits), equivalent to input_prec
 instrn_width = 48 # (in bits)
-
+# Input and Weight parameters
+input_prec = 16
+weight_width = 16
 # Change here - Specify the IMA parameters here
 xbar_bits = 2
 num_matrix = 2 # each matrix is 1-fw logical xbar for inference and 1-fw, 1-bw, and 1 delta logical xbar for training. Each logical xbar for inference is 8-fw physical xbar and for training  8-fw, 8-bw and 16-delta physical xbars.

include/example-configs/config-mlp.py

@@ -30,7 +30,9 @@
 data_width = num_bits # (in bits)
 xbdata_width = data_width # (in bits)
 instrn_width = 48 # (in bits)
-
+# Input and Weight parameters
+input_prec = 16
+weight_width = 16
 # Change here - Specify the IMA parameters here
 xbar_bits = 2
 num_matrix = 2 # each matrix is 1-fw logical xbar for inference and 1-fw, 1-bw, and 1 delta logical xbar for training. Each logical xbar for inference is 8-fw physical xbar and for training  8-fw, 8-bw and 16-delta physical xbars.

src/ima.py

@@ -598,7 +598,7 @@ def inner_product (mat_id, key):
                     self.xb_outMem_list[mat_id][key].reset ()
 
                     ## Loop to cover all bits of inputs
-                    for k in xrange (cfg.xbdata_width / cfg.dac_res):
+                    for k in xrange (int(math.ceil(cfg.input_prec / cfg.dac_res))): #quantization affects the # of streams
                     #for k in xrange (1):
                         # read the values from the xbar's input register
                         out_xb_inMem = self.xb_inMem_list[mat_id][key].read (cfg.dac_res)
@@ -612,7 +612,7 @@ def inner_product (mat_id, key):
                         out_dac = self.dacArray_list[mat_id][key].propagate_dummy(out_xb_inMem) #pass through
 
                         # Do for (data_width/xbar_bits) xbars
-                        num_xb = cfg.data_width / cfg.xbar_bits
+                        num_xb = int(math.ceil(float(cfg.weight_width) / cfg.xbar_bits))  # # of XBs change with quantization
                         out_xbar = [[] for x in range(num_xb)]
                         out_snh = [[] for x in range(num_xb)]
                         for m in range (num_xb):
@@ -793,7 +793,8 @@ def xbComputeLatency (self, mask):
                     print("adccccc.adc_res", adccccc.adc_res)
                 print("---")
                 '''
-                latency_ip = lat_temp * ((cfg.xbdata_width / cfg.dac_res) + num_stage - 1) * float(int(fb_found>0))
+                latency_ip = lat_temp * ((cfg.input_prec / cfg.dac_res) + num_stage - 1) * float(int(fb_found>0))*(math.ceil(float(cfg.weight_width)/ \
+                cfg.xbar_bits) /math.ceil(float(cfg.data_width)/cfg.xbar_bits)) # last term to account for the effect of quantization on latency
                 ## MVM outer product occurs in 4 cycles to take care of all i/o polarities (++, +-, -+, --)
                 num_phase = 4
                 lat_temp = self.matrix_list[0]['f'][0].getOpLatency()

test/mlp_l4_mnist/fully-connected-layer.cpp

@@ -0,0 +1,57 @@
+/*
+ *  Copyright (c) 2019 IMPACT Research Group, University of Illinois.
+ *  All rights reserved.
+ *
+ *  This file is covered by the LICENSE.txt license file in the root directory.
+ *
+ */
+
+#include <assert.h>
+#include <string>
+#include <vector>
+#include <iostream>
+
+#include "puma.h"
+#include "fully-connected-layer.h"
+
+int main(int argc, char** argv) {
+
+    //Model model = Model::create("fully-connected-layer");
+
+    // Process parameters
+    unsigned int in_size;
+    unsigned int out_size;
+    if(argc == 4) {
+        in_size = atoi(argv[1]);
+        out_size = atoi(argv[2]);
+	}
+
+    std:: string str=std::string("fully") + argv[3] + std::string("-connected-layer");
+    Model model = Model::create(str);
+
+    // Input
+    auto in = InputVector::create(model, "in", in_size);
+
+    // Output
+    auto out = OutputVector::create(model, "out", out_size);
+
+    // Layer
+    out = fully_connected_layer(model, "", in_size, out_size, in);
+
+    // Compile
+    model.compile();
+
+    // Bind data
+    ModelInstance modelInstance = ModelInstance::create(model);
+    float* weights = new float[in_size*out_size];
+    fully_connected_layer_bind(modelInstance, "", weights);
+    modelInstance.generateData();
+
+    // Destroy model
+    model.destroy();
+    delete[] weights;
+
+    return 0;
+
+}
+

test/mlp_l4_mnist/fully-connected-layer.h

@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2019 IMPACT Research Group, University of Illinois.
+ *  All rights reserved.
+ *
+ *  This file is covered by the LICENSE.txt license file in the root directory.
+ *
+ */
+
+#ifndef _PUMA_TEST_FULLY_CONNECTED_LAYER_
+#define _PUMA_TEST_FULLY_CONNECTED_LAYER_
+
+#include "puma.h"
+
+static Vector fully_connected_layer(Model model, std::string layerName, unsigned int in_size, unsigned int out_size, Vector in) {
+
+    ConstantMatrix mat = ConstantMatrix::create(model, layerName + "mat", in_size, out_size);
+
+    return sig(mat*in);
+
+}
+
+static void fully_connected_layer_bind(ModelInstance modelInstance, std::string layerName, float* weights) {
+    modelInstance.bind(layerName + "mat", weights);
+}
+
+#endif
+

test/utils/run-mlp-layer.sh

@@ -0,0 +1,46 @@
+set -v
+set -e
+path=`pwd` #path to your puma directory
+echo $path
+cppfile=fully-connected-layer #name for cpp file that you want to compile ex- mlp_l4_mnist.cpp, conv-layer.cpp, convmax-layer.cpp
+name=fully #name for the folder generated by compiler
+pumaenv=pumaenv #name for the environment 
+fileno=0
+name=$name$fileno
+
+#layer parameter
+in=64
+out=10
+
+
+#copying mlp config file
+rm ${path}/puma-simulator/include/config.py #remove existing config file
+cp ${path}/puma-simulator/include/example-configs/config-mlp.py  ${path}/puma-simulator/include/config.py #copy the mlp config file to include
+#copying model file
+rm ${path}/puma-compiler/test/${cppfile}.cpp ${path}/puma-compiler/test/${cppfile}.h   
+cp ${path}/puma-simulator/test/mlp_l4_mnist/${cppfile}.cpp  ${path}/puma-compiler/test/${cppfile}.cpp #copy the mlp config file to include 
+cp ${path}/puma-simulator/test/mlp_l4_mnist/${cppfile}.h  ${path}/puma-compiler/test/${cppfile}.h #copy the mlp config file to include 
+
+cd ${path}/puma-compiler/src
+source ~/.bashrc
+conda activate ${pumaenv}
+
+make clean
+make
+
+cd ${path}/puma-compiler/test
+make clean
+make ${cppfile}.test
+export LD_LIBRARY_PATH=`pwd`/../src:$LD_LIBRARY_PATH
+./${cppfile}.test ${in} ${out} ${fileno}
+echo $cppfile  
+./generate-py.sh 
+cp -r ${name} ../../puma-simulator/test/testasm
+
+cd ${path}/puma-simulator/src
+
+
+python dpe.py -n ${name} 
+
+
+