Merge pull request #1 from cms-hls4ml/jmduarte-patch-1

Updates

.gitignore

@@ -1,4 +1,4 @@
-TOoLLip_v1/NN/
 .gitignore
 *.so
 *.o
+*~

.gitmodules

@@ -1,3 +1,3 @@
 [submodule "TOoLLip_v1/L1JetTagger"]
-	path = TOoLLip_v1/L1JetTagger
+	path = TOoLLiP_v1/L1JetTagger
 	url = https://github.com/ucsd-hep-ex/L1JetTagger

Makefile

@@ -1,13 +1,13 @@
 CPP_STANDARD := c++17
 CXXFLAGS := -O3 -fPIC -std=$(CPP_STANDARD)
 PREFIX:=.
-EMULATOR_EXTRAS := /cvmfs/cms.cern.ch/el8_amd64_gcc11/external/hls4mlEmulatorExtras/1.1.1-6933fcc7cdb4cdd5a649bd6579151d1b/
+EMULATOR_EXTRAS := ../../hls4mlEmulatorExtras
 AP_TYPES := $(EMULATOR_EXTRAS)/include/ap_types
-HLS_ROOT := /cvmfs/cms.cern.ch/el8_amd64_gcc11/external/hls/2019.08-fd724004387c2a6770dc3517446d30d9
+HLS_ROOT := ../../hls
 HLS4ML_INCLUDE := $(EMULATOR_EXTRAS)/include/hls4ml
 INCLUDES := -I$(HLS4ML_INCLUDE) -I$(AP_TYPES) -I$(HLS_ROOT)/include
 LD_FLAGS := -L$(EMULATOR_EXTRAS)/lib64 -lemulator_interface
-ALL_VERSIONS:=TOoLLip_v1/TOoLLip_emulator_v1.so
+ALL_VERSIONS:=TOoLLiP_v1/TOoLLiP_v1.so
 
 .DEFAULT_GOAL := all
 .PHONY: all clean install
@@ -19,10 +19,10 @@ all: $(ALL_VERSIONS)
 install: all
 	@rm -rf $(PREFIX)/lib64
 	@mkdir -p $(PREFIX)/lib64
-	cp TOoLLip_*.so $(PREFIX)/lib64
+	cp TOoLLiP_*.so $(PREFIX)/lib64
 
 %.so:
 	$(MAKE) -C $(@D) INCLUDES="$(INCLUDES)" LD_FLAGS="$(LD_FLAGS)" CXXFLAGS="$(CXXFLAGS)"
 
 clean:
-	rm TOoLLip_*.so
+	rm TOoLLiP_*.so

README.md

@@ -0,0 +1,18 @@
+# TOoLLiP
+
+- Model training and conversion code: https://github.com/ucsd-hep-ex/L1JetTag
+- Emulator tutorial: https://gitlab.cern.ch/fastmachinelearning/cms_mlatl1t_tutorial/-/tree/master/part3
+
+## v1
+- Based on QKeras model: https://github.com/ucsd-hep-ex/L1JetTagger/blob/eb93e812690dcf1431bf5cb27bf063d154daed1e/qkL1JetTagModel.h5
+- Convert script: https://github.com/ucsd-hep-ex/L1JetTagger/blob/eb93e812690dcf1431bf5cb27bf063d154daed1e/HLS_qk_Roc.py
+- Converted HLS project: https://github.com/ucsd-hep-ex/L1JetTagger/tree/eb93e812690dcf1431bf5cb27bf063d154daed1e
+- hls4ml version/commit used: https://github.com/jmduarte/hls4ml/tree/8dea6ad51f78be4662513c1d4482b4b0ab94c5de
+- Model:
+![model_summary](https://github.com/cms-hls4ml/TOoLLiP/assets/4932543/788b4d4b-4268-4ad4-94ab-26b4edadb0de)
+
+- Expected performance:
+
+![hls_Qk_ROCCurve](https://github.com/cms-hls4ml/TOoLLiP/assets/4932543/1470ab5b-ec51-4741-a500-ace2a7a14b5d)
+
+![profiling_sigmoid](https://github.com/cms-hls4ml/TOoLLiP/assets/4932543/48042a16-5a9c-4697-b891-a1b57fa899ac)

TOoLLip_v1/Makefile → TOoLLiP_v1/Makefile

@@ -1,6 +1,6 @@
 .PHONY: clean
 
-TOoLLip_emulator_v1.so: TOoLLip_emulator_v1.o myproject.o
+TOoLLiP_v1.so: TOoLLiP_emulator_v1.o TOoLLiP_v1.o
 	$(CXX) $(CXXFLAGS) $(LD_FLAGS) -shared $^ -o $@
 
 %.o: NN/%.cpp
@@ -10,4 +10,4 @@ TOoLLip_emulator_v1.so: TOoLLip_emulator_v1.o myproject.o
 	$(CXX) $(CXXFLAGS) $(INCLUDES) -c $< -o $@
 
 clean:
-	rm TOoLLip_emulator_v1.so TOoLLip_emulator_v1.o myproject.o
+	rm TOoLLiP_emulator_v1.o TOoLLiP_v1.o TOoLLiP_v1.so

TOoLLiP_v1/NN/TOoLLiP_v1.cpp

@@ -0,0 +1,73 @@
+#include <iostream>
+
+#include "TOoLLiP_v1.h"
+#include "parameters.h"
+
+void TOoLLiP_v1(
+    input_t input_1[N_INPUT_1_1*N_INPUT_2_1],
+    result_t layer14_out[N_LAYER_12]
+) {
+
+    // hls-fpga-machine-learning insert IO
+    #pragma HLS ARRAY_RESHAPE variable=input_1 complete dim=0
+    #pragma HLS ARRAY_PARTITION variable=layer14_out complete dim=0
+    #pragma HLS INTERFACE ap_vld port=input_1,layer14_out 
+    #pragma HLS DATAFLOW 
+
+#ifdef __HLS4ML_LOAD_TXT_WEIGHTS__
+    static bool loaded_weights = false;
+    if (!loaded_weights) {
+        // hls-fpga-machine-learning insert load weights
+        nnet::load_weights_from_txt<weight15_t, 140>(w15, "w15.txt");
+        nnet::load_weights_from_txt<bias15_t, 10>(b15, "b15.txt");
+        nnet::load_weights_from_txt<weight16_t, 100>(w16, "w16.txt");
+        nnet::load_weights_from_txt<bias16_t, 10>(b16, "b16.txt");
+        nnet::load_weights_from_txt<weight9_t, 100>(w9, "w9.txt");
+        nnet::load_weights_from_txt<bias9_t, 10>(b9, "b9.txt");
+        nnet::load_weights_from_txt<weight12_t, 10>(w12, "w12.txt");
+        nnet::load_weights_from_txt<bias12_t, 1>(b12, "b12.txt");
+        loaded_weights = true;
+    }
+#endif
+
+    // ****************************************
+    // NETWORK INSTANTIATION
+    // ****************************************
+
+    // hls-fpga-machine-learning insert layers
+
+    layer15_t layer15_out[N_OUTPUTS_15*N_FILT_15];
+    #pragma HLS ARRAY_PARTITION variable=layer15_out complete dim=0
+    nnet::pointwise_conv_1d_cl<input_t, layer15_t, config15>(input_1, layer15_out, w15, b15); // q_conv1d
+
+    layer4_t layer4_out[N_OUTPUTS_2*N_FILT_2];
+    #pragma HLS ARRAY_PARTITION variable=layer4_out complete dim=0
+    nnet::relu<layer15_t, layer4_t, relu_config4>(layer15_out, layer4_out); // q_activation
+
+    layer16_t layer16_out[N_OUTPUTS_16*N_FILT_16];
+    #pragma HLS ARRAY_PARTITION variable=layer16_out complete dim=0
+    nnet::pointwise_conv_1d_cl<layer4_t, layer16_t, config16>(layer4_out, layer16_out, w16, b16); // q_conv1d_1
+
+    layer7_t layer7_out[N_OUTPUTS_5*N_FILT_5];
+    #pragma HLS ARRAY_PARTITION variable=layer7_out complete dim=0
+    nnet::relu<layer16_t, layer7_t, relu_config7>(layer16_out, layer7_out); // q_activation_1
+
+    layer8_t layer8_out[N_FILT_8];
+    #pragma HLS ARRAY_PARTITION variable=layer8_out complete dim=0
+    nnet::global_pooling1d_cl<layer7_t, layer8_t, config8>(layer7_out, layer8_out); // global_average_pooling1d
+
+    layer9_t layer9_out[N_LAYER_9];
+    #pragma HLS ARRAY_PARTITION variable=layer9_out complete dim=0
+    nnet::dense<layer8_t, layer9_t, config9>(layer8_out, layer9_out, w9, b9); // q_dense
+
+    layer11_t layer11_out[N_LAYER_9];
+    #pragma HLS ARRAY_PARTITION variable=layer11_out complete dim=0
+    nnet::relu<layer9_t, layer11_t, relu_config11>(layer9_out, layer11_out); // q_activation_2
+
+    layer12_t layer12_out[N_LAYER_12];
+    #pragma HLS ARRAY_PARTITION variable=layer12_out complete dim=0
+    nnet::dense<layer11_t, layer12_t, config12>(layer11_out, layer12_out, w12, b12); // q_dense_1
+
+    nnet::sigmoid<layer12_t, result_t, sigmoid_config14>(layer12_out, layer14_out); // sigmoid
+
+}

TOoLLiP_v1/NN/TOoLLiP_v1.h

@@ -0,0 +1,16 @@
+#ifndef TOOLLIP_V1_H_
+#define TOOLLIP_V1_H_
+
+#include "ap_fixed.h"
+#include "ap_int.h"
+#include "hls_stream.h"
+
+#include "defines.h"
+
+// Prototype of top level function for C-synthesis
+void TOoLLiP_v1(
+    input_t input_1[N_INPUT_1_1*N_INPUT_2_1],
+    result_t layer14_out[N_LAYER_12]
+);
+
+#endif

TOoLLiP_v1/NN/defines.h

@@ -0,0 +1,54 @@
+#ifndef DEFINES_H_
+#define DEFINES_H_
+
+#include "ap_fixed.h"
+#include "ap_int.h"
+#include "nnet_utils/nnet_types.h"
+#include <cstddef>
+#include <cstdio>
+
+// hls-fpga-machine-learning insert numbers
+#define N_INPUT_1_1 10
+#define N_INPUT_2_1 14
+#define N_OUTPUTS_15 10
+#define N_FILT_15 10
+#define N_OUTPUTS_2 10
+#define N_FILT_2 10
+#define N_OUTPUTS_16 10
+#define N_FILT_16 10
+#define N_OUTPUTS_5 10
+#define N_FILT_5 10
+#define N_FILT_8 10
+#define N_LAYER_9 10
+#define N_LAYER_9 10
+#define N_LAYER_12 1
+#define N_LAYER_12 1
+
+// hls-fpga-machine-learning insert layer-precision
+typedef ap_fixed<16,6> input_t;
+typedef ap_fixed<16,6> model_default_t;
+typedef ap_fixed<16,6> layer15_t;
+typedef ap_fixed<8,4> weight15_t;
+typedef ap_fixed<8,4> bias15_t;
+typedef ap_ufixed<8,0,AP_RND_CONV,AP_SAT> layer4_t;
+typedef ap_fixed<18,8> q_activation_table_t;
+typedef ap_fixed<16,6> layer16_t;
+typedef ap_fixed<8,4> weight16_t;
+typedef ap_fixed<8,4> bias16_t;
+typedef ap_ufixed<8,0,AP_RND_CONV,AP_SAT> layer7_t;
+typedef ap_fixed<18,8> q_activation_1_table_t;
+typedef ap_fixed<16,6> layer8_t;
+typedef ap_fixed<16,6> layer9_t;
+typedef ap_fixed<8,4> weight9_t;
+typedef ap_fixed<8,4> bias9_t;
+typedef ap_uint<1> layer9_index;
+typedef ap_ufixed<8,0,AP_RND_CONV,AP_SAT> layer11_t;
+typedef ap_fixed<18,8> q_activation_2_table_t;
+typedef ap_fixed<16,6> layer12_t;
+typedef ap_fixed<8,4> weight12_t;
+typedef ap_fixed<8,4> bias12_t;
+typedef ap_uint<1> layer12_index;
+typedef ap_fixed<16,6> result_t;
+typedef ap_fixed<18,8> sigmoid_table_t;
+
+#endif