feat: new Tracy Time preset and more efficient univariate extension

barretenberg/cpp/CMakeLists.txt

@@ -61,6 +61,15 @@ else()
     SET(TRACY_LIBS)
 endif()
 
+if(TRACY_PROFILE_MEMORY)
+  add_compile_options(-DTRACY_MEMORY)
+endif()
+
+if(TRACY_PROFILE_TIME)
+  add_compile_options(-DTRACY_TIME)
+endif()
+
+
 if(ENABLE_ASAN)
     add_compile_options(-fsanitize=address)
     add_link_options(-fsanitize=address)

barretenberg/cpp/CMakePresets.json

@@ -111,13 +111,31 @@
       }
     },
     {
-      "name": "tracy",
+      "name": "tracy-memory",
       "displayName": "Release build with tracy, optimized for memory tracking",
       "description": "Release build with tracy, optimized for memory tracking",
       "inherits": "clang16",
-      "binaryDir": "build-tracy",
+      "binaryDir": "build-tracy-memory",
       "cacheVariables": {
-        "ENABLE_TRACY": "ON"
+        "ENABLE_TRACY": "ON",
+        "TRACY_PROFILE_MEMORY": "ON"
+      }
+    },
+    {
+      "name": "tracy-time",
+      "displayName": "Build for tracy time profiling",
+      "description": "Build for tracy time profiling",
+      "binaryDir": "build-tracy-time",
+      "inherits": "clang16",
+      "environment": {
+        "CMAKE_BUILD_TYPE": "RelWithDebInfo",
+        "CFLAGS": "-g -fno-omit-frame-pointer",
+        "CXXFLAGS": "-g -fno-omit-frame-pointer",
+        "LDFLAGS": "-g -fno-omit-frame-pointer -rdynamic"
+      },
+      "cacheVariables": {
+        "ENABLE_TRACY": "ON",
+        "TRACY_PROFILE_TIME": "ON"
       }
     },
     {
@@ -472,9 +490,14 @@
       "configurePreset": "clang16-dbg"
     },
     {
-      "name": "tracy",
+      "name": "tracy-memory",
+      "inherits": "default",
+      "configurePreset": "tracy-memory"
+    },
+    {
+      "name": "tracy-time",
       "inherits": "default",
-      "configurePreset": "tracy"
+      "configurePreset": "tracy-time"
     },
     {
       "name": "clang16-pic",

barretenberg/cpp/scripts/benchmark_tracy_build_mainframe_run_local.sh

@@ -0,0 +1,49 @@
+
+# NOTE: intended to be ran from one's external computer, connecting to Aztec mainframe
+# IF ON YOUR LOCAL COMPUTER USE NORMAL INTERACTIVE TRACY WORKFLOW
+# the benchmark runs with headless capture and then we copy the trace file and run tracy profiler
+# This is thus only really useful internally at Aztec, sorry external folks. It can be easily tweaked
+# however for any SSH setup, especially an ubuntu one.
+# on local machine run:
+# export USER=...
+# export PRESET=...tracy for memory or tracy-gates for circuit gates...
+# ssh $USER-box "cat ~/aztec-packages/barretenberg/cpp/scripts/benchmark_tracy.sh" | bash /dev/stdin $USER
+set -eux
+USER=${1:-$USER}
+BOX=$USER-box
+BENCHMARK=${2:-protogalaxy_bench}
+COMMAND=${3:-./bin/$BENCHMARK --benchmark_filter=fold_k/17}
+
+# Can also set PRESET=tracy-gates env variable
+PRESET=${PRESET:-tracy-time}
+
+wait # TODO(AD) hack - not sure why needed
+! [ -d ~/tracy ] && git clone https://github.com/wolfpld/tracy ~/tracy
+cd ~/tracy
+git checkout 075395620a504c0cdcaf9bab3d196db16a043de7 # release 0.11.0
+cmake -B profiler/build -S profiler -DCMAKE_BUILD_TYPE=Release
+cmake --build profiler/build --parallel
+cd -
+
+ssh $BOX "
+	set -eux ;
+	cd ~/aztec-packages/barretenberg/cpp/ ;
+	cmake --preset $PRESET && cmake --build --preset $PRESET --target $BENCHMARK ;
+" &
+wait
+if [ ! -d build-$PRESET/bin ]; then
+  echo build-$PRESET/bin;
+  mkdir -p build-$PRESET/bin;
+fi
+scp $BOX:/mnt/user-data/$USER/aztec-packages/barretenberg/cpp/build-$PRESET/bin/$BENCHMARK build-$PRESET/bin/. ;
+! [ -d ~/tracy ] && git clone https://github.com/wolfpld/tracy ~/tracy ;
+cd ~/tracy/capture ;
+	git checkout 075395620a504c0cdcaf9bab3d196db16a043de7 ;
+mkdir -p build && cd build && cmake .. && make -j ;
+
+./tracy-capture -a 127.0.0.1 -f -o ../trace-$BENCHMARK &
+sleep 0.1 ;
+cd ~/aztec-packages/barretenberg/cpp/build-$PRESET/
+$COMMAND ;
+
+~/tracy/profiler/build/tracy-profiler ~/tracy/capture/trace-$BENCHMARK

barretenberg/cpp/src/barretenberg/benchmark/client_ivc_bench/client_ivc.bench.cpp

@@ -45,7 +45,7 @@ BENCHMARK_DEFINE_F(ClientIVCBench, Full)(benchmark::State& state)
     }
 }
 
-#define ARGS Arg(ClientIVCBench::NUM_ITERATIONS_MEDIUM_COMPLEXITY)
+#define ARGS Arg(ClientIVCBench::NUM_ITERATIONS_MEDIUM_COMPLEXITY)->Arg(2)
 
 BENCHMARK_REGISTER_F(ClientIVCBench, Full)->Unit(benchmark::kMillisecond)->ARGS;
 

barretenberg/cpp/src/barretenberg/benchmark/protogalaxy_bench/protogalaxy.bench.cpp

@@ -75,7 +75,7 @@ void fold_k(State& state) noexcept
     }
 }
 
-BENCHMARK(vector_of_evaluations)->DenseRange(15, 21)->Unit(kMillisecond);
+BENCHMARK(vector_of_evaluations)->DenseRange(15, 21)->Unit(kMillisecond)->Iterations(1);
 BENCHMARK(compute_row_evaluations)->DenseRange(15, 21)->Unit(kMillisecond);
 // We stick to just k=1 for compile-time reasons.
 BENCHMARK(fold_k)->/* vary the circuit size */ DenseRange(14, 20)->Unit(kMillisecond);

barretenberg/cpp/src/barretenberg/benchmark/relations_bench/barycentric.bench.cpp

@@ -14,14 +14,50 @@ using bb::Univariate;
 
 namespace bb::benchmark {
 
-void extend_2_to_6(State& state) noexcept
+void extend_2_to_11(State& state) noexcept
 {
     auto univariate = Univariate<FF, 2>::get_random();
     for (auto _ : state) {
-        DoNotOptimize(univariate.extend_to<6>());
+        DoNotOptimize(univariate.extend_to<11>());
     }
 }
-BENCHMARK(extend_2_to_6);
+
+// 93.9s goes down to 62.7
+// Theoretical min: 1 sub, 9 additions at about 3.8ns each, 38ns
+void fake_extend_2_to_11(State& state) noexcept
+{
+    std::array<FF, 11> univariate;
+    std::generate(univariate.begin(), univariate.end(), [&]() { return FF::random_element(); });
+
+    const auto extend_to_11 = [](auto& arr) {
+        FF tmp = arr[1];
+        const FF delta = tmp - arr[0];
+        for (size_t idx = 2; idx < 10; idx++) {
+            arr[idx] = (tmp += delta); // fused ~> 62.9ns; non-fused ~>69.5ns
+        }
+        arr[10] = tmp; // save one +=;
+        return arr;
+    };
+
+    for (auto _ : state) {
+        DoNotOptimize(extend_to_11(univariate));
+    }
+}
+
+// 93.9s goes down to 62.7
+// Theoretical min: 1 sub, 9 additions at about 3.8ns each, 38ns
+void self_extend_2_to_11(State& state) noexcept
+{
+    auto univariate = Univariate<FF, 11>::get_random();
+
+    for (auto _ : state) {
+        univariate.self_extend_from<2>();
+    }
+}
+
+BENCHMARK(extend_2_to_11);
+BENCHMARK(fake_extend_2_to_11);
+BENCHMARK(self_extend_2_to_11);
 
 } // namespace bb::benchmark
 

barretenberg/cpp/src/barretenberg/benchmark/ultra_bench/mock_circuits.hpp

@@ -54,7 +54,9 @@ Prover get_prover(void (*test_circuit_function)(typename Prover::Flavor::Circuit
         Composer composer;
         return composer.create_prover(builder);
     } else {
+#ifdef TRACY_MEMORY
         ZoneScopedN("creating prover");
+#endif
         return Prover(builder);
     }
 };

barretenberg/cpp/src/barretenberg/common/op_count.hpp

@@ -2,6 +2,7 @@
 #pragma once
 
 #include <memory>
+#include <tracy/Tracy.hpp>
 #ifndef BB_USE_OP_COUNT
 // require a semicolon to appease formatters
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
@@ -11,12 +12,19 @@
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define BB_OP_COUNT_CYCLES_NAME(name) (void)0
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define BB_OP_COUNT_TIME_NAME(name) (void)0
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define BB_OP_COUNT_CYCLES() (void)0
+#ifndef TRACY_TIME
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define BB_OP_COUNT_TIME_NAME(name) (void)0
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define BB_OP_COUNT_TIME() (void)0
 #else
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define BB_OP_COUNT_TIME_NAME(name) ZoneScopedN(name)
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define BB_OP_COUNT_TIME() BB_OP_COUNT_TIME_NAME(__func__)
+#endif
+#else
 /**
  * Provides an abstraction that counts operations based on function names.
  * For efficiency, we spread out counts across threads.

barretenberg/cpp/src/barretenberg/execution_trace/execution_trace.cpp

@@ -8,7 +8,9 @@ namespace bb {
 
 template <class Flavor> void ExecutionTrace_<Flavor>::populate_public_inputs_block(Builder& builder)
 {
+#ifdef TRACY_MEMORY
     ZoneScopedN("populate_public_inputs_block");
+#endif
     // Update the public inputs block
     for (const auto& idx : builder.public_inputs) {
         for (size_t wire_idx = 0; wire_idx < NUM_WIRES; ++wire_idx) {
@@ -27,7 +29,10 @@ template <class Flavor> void ExecutionTrace_<Flavor>::populate_public_inputs_blo
 template <class Flavor>
 void ExecutionTrace_<Flavor>::populate(Builder& builder, typename Flavor::ProvingKey& proving_key, bool is_structured)
 {
+
+#ifdef TRACY_MEMORY
     ZoneScopedN("trace populate");
+#endif
     // Share wire polynomials, selector polynomials between proving key and builder and copy cycles from raw circuit
     // data
     auto trace_data = construct_trace_data(builder, proving_key, is_structured);
@@ -36,18 +41,27 @@ void ExecutionTrace_<Flavor>::populate(Builder& builder, typename Flavor::Provin
         proving_key.pub_inputs_offset = trace_data.pub_inputs_offset;
     }
     if constexpr (IsUltraPlonkOrHonk<Flavor>) {
+
+#ifdef TRACY_MEMORY
         ZoneScopedN("add_memory_records_to_proving_key");
+#endif
         add_memory_records_to_proving_key(trace_data, builder, proving_key);
     }
 
     if constexpr (IsGoblinFlavor<Flavor>) {
+
+#ifdef TRACY_MEMORY
         ZoneScopedN("add_ecc_op_wires_to_proving_key");
+#endif
         add_ecc_op_wires_to_proving_key(builder, proving_key);
     }
 
     // Compute the permutation argument polynomials (sigma/id) and add them to proving key
     {
+
+#ifdef TRACY_MEMORY
         ZoneScopedN("compute_permutation_argument_polynomials");
+#endif
         compute_permutation_argument_polynomials<Flavor>(builder, &proving_key, trace_data.copy_cycles);
     }
 }
@@ -73,7 +87,10 @@ template <class Flavor>
 typename ExecutionTrace_<Flavor>::TraceData ExecutionTrace_<Flavor>::construct_trace_data(
     Builder& builder, typename Flavor::ProvingKey& proving_key, bool is_structured)
 {
+
+#ifdef TRACY_MEMORY
     ZoneScopedN("construct_trace_data");
+#endif
 
     if constexpr (IsPlonkFlavor<Flavor>) {
         // Complete the public inputs execution trace block from builder.public_inputs
@@ -91,7 +108,10 @@ typename ExecutionTrace_<Flavor>::TraceData ExecutionTrace_<Flavor>::construct_t
         // Update wire polynomials and copy cycles
         // NB: The order of row/column loops is arbitrary but needs to be row/column to match old copy_cycle code
         {
+
+#ifdef TRACY_MEMORY
             ZoneScopedN("populating wires and copy_cycles");
+#endif
             for (uint32_t block_row_idx = 0; block_row_idx < block_size; ++block_row_idx) {
                 for (uint32_t wire_idx = 0; wire_idx < NUM_WIRES; ++wire_idx) {
                     uint32_t var_idx = block.wires[wire_idx][block_row_idx]; // an index into the variables array

barretenberg/cpp/src/barretenberg/execution_trace/execution_trace.hpp

@@ -28,7 +28,10 @@ template <class Flavor> class ExecutionTrace_ {
 
         TraceData(Builder& builder, ProvingKey& proving_key)
         {
+
+#ifdef TRACY_MEMORY
             ZoneScopedN("TraceData constructor");
+#endif
             if constexpr (IsHonkFlavor<Flavor>) {
                 // Initialize and share the wire and selector polynomials
                 for (auto [wire, other_wire] : zip_view(wires, proving_key.polynomials.get_wires())) {
@@ -45,7 +48,10 @@ template <class Flavor> class ExecutionTrace_ {
                     proving_key.polynomial_store.put(wire_tag, wires[idx].share());
                 }
                 {
+
+#ifdef TRACY_MEMORY
                     ZoneScopedN("selector initialization");
+#endif
                     for (size_t idx = 0; idx < Builder::Arithmetization::NUM_SELECTORS; ++idx) {
                         selectors[idx] = Polynomial(proving_key.circuit_size);
                         std::string selector_tag = builder.selector_names[idx] + "_lagrange";
@@ -54,7 +60,10 @@ template <class Flavor> class ExecutionTrace_ {
                 }
             }
             {
+
+#ifdef TRACY_MEMORY
                 ZoneScopedN("copy cycle initialization");
+#endif
                 copy_cycles.resize(builder.variables.size());
             }
         }

barretenberg/cpp/src/barretenberg/flavor/flavor.hpp

@@ -129,7 +129,10 @@ template <typename FF, typename CommitmentKey_> class ProvingKey_ {
                 std::shared_ptr<CommitmentKey_> commitment_key = nullptr)
     {
         if (commitment_key == nullptr) {
+
+#ifdef TRACY_MEMORY
             ZoneScopedN("init commitment key");
+#endif
             this->commitment_key = std::make_shared<CommitmentKey_>(circuit_size);
         } else {
             // Don't create another commitment key if we already have one