Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .github/workflows/harmony-fast-dsp.yml
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,10 @@ jobs:
run: |
g++ -std=c++20 -O2 -Wall -Wextra tests/harmony_pitch_synchronous_test.cpp -o /tmp/arson-harmony-psola
/tmp/arson-harmony-psola
- name: Profile legacy versus exact-output epoch scan on same runner (non-gating)
run: |
g++ -std=c++20 -O2 -Wall -Wextra tests/harmonice_epoch_cpu_profile_test.cpp -o /tmp/harmonice-epoch-cpu
/tmp/harmonice-epoch-cpu
- name: Compile and run complete engine dry, pitch, stereo and graph-independent PCM fixtures
run: |
g++ -std=c++20 -O2 -Wall -Wextra tests/harmony_dsp_test.cpp -o /tmp/arson-harmony-engine
Expand Down
6 changes: 6 additions & 0 deletions CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -136,6 +136,12 @@ if(BUILD_TESTING)
add_executable(ARMixHarmonyPitchSyncTests tests/harmony_pitch_synchronous_test.cpp)
target_compile_features(ARMixHarmonyPitchSyncTests PRIVATE cxx_std_20)
add_test(NAME NativeHarmonyPitchSynchronousFormantCandidateAndBlockParity COMMAND ARMixHarmonyPitchSyncTests)
# Non-gating microprofile of exact-output reference versus ring-cursor
# epoch scan. Must not be represented as total mixer/ASIO timing.
add_executable(ARMixHarmoniceEpochCpuProfile tests/harmonice_epoch_cpu_profile_test.cpp)
target_compile_features(ARMixHarmoniceEpochCpuProfile PRIVATE cxx_std_20)
add_test(NAME NativeHarmoniceEpochSameRunnerCpuComparison
COMMAND ARMixHarmoniceEpochCpuProfile)
add_executable(ARMixHarmonySpatialTests tests/harmony_stage_spatial_test.cpp)
target_compile_features(ARMixHarmonySpatialTests PRIVATE cxx_std_20)
add_test(NAME NativeHarmonyBackingStageSpatialDepthAndMonoPDC COMMAND ARMixHarmonySpatialTests)
Expand Down
31 changes: 25 additions & 6 deletions engine/builtin_fx/HarmonyPitchSynchronous.h
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,10 @@ namespace armix::harmony {
// Positive waveform peaks are approximate epochs, NOT proven glottal closures.
// This must pass actual singer A/B before acoustic release qualification.
// All state is fixed; no allocation, locks, FFT or IO in process().
template <std::size_t Capacity, int Voices>
// The default ring-cursor walk is sample-exact to the original lookup.
// False compiles the historical age/subtraction loop for SAME-RUNNER offline
// A/B; this is a compile-time switch, never a GUI/session/RT branch.
template <std::size_t Capacity, int Voices, bool FastEpochWalk=true>
class PitchSynchronousVoice final {
static_assert(Capacity >= 4096 && (Capacity & (Capacity-1)) == 0);
static_assert(Voices > 0 && Voices <= 8);
Expand Down Expand Up @@ -95,11 +98,27 @@ class PitchSynchronousVoice final {
if(end<start)return -1;
std::int64_t peak=-1;
float best=.003f;
for(auto t=start;t<=end;++t) {
const auto age=static_cast<std::size_t>(
static_cast<std::int64_t>(time)-t);
const float value=input[(writeIndex+Capacity-age)&(Capacity-1)];
if(value>best){best=value;peak=t;}
if constexpr(FastEpochWalk) {
// Adjacent integer sample times occupy adjacent ring cells. One
// initial age/index calculation replaces the repeated 64-bit
// per-candidate age and index arithmetic, without changing sample
// access order, threshold or earliest-max tie behavior.
const auto firstAge=static_cast<std::size_t>(
static_cast<std::int64_t>(time)-start);
auto ring=(writeIndex+Capacity-firstAge)&(Capacity-1);
for(auto t=start;t<=end;++t) {
const float value=input[ring];
if(value>best){best=value;peak=t;}
ring=(ring+1)&(Capacity-1);
}
}else{
// Legacy reference, built only by the offline PCM/CPU verifier.
for(auto t=start;t<=end;++t) {
const auto age=static_cast<std::size_t>(
static_cast<std::int64_t>(time)-t);
const float value=input[(writeIndex+Capacity-age)&(Capacity-1)];
if(value>best){best=value;peak=t;}
}
}
return peak;
}
Expand Down
94 changes: 94 additions & 0 deletions tests/harmonice_epoch_cpu_profile_test.cpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,94 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Non-gating, same-runner old/new pitch-epoch workload comparison.
// Timing is synthetic PSOLA-only, not full mixer, driver or vocal naturalness.
#include "../engine/builtin_fx/HarmonyPitchSynchronous.h"
#include <algorithm>
#include <array>
#include <chrono>
#include <cmath>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <vector>
namespace {
using Clock=std::chrono::steady_clock;
constexpr int block=64,warmup=128,measured=1024;
constexpr double pi2=6.283185307179586;
struct Result {
double p50=0.,p95=0.,p99=0.,p999=0.,maximum=0.,checksum=0.;
};
template <bool Fast>
Result measure(double rate,double f0){
using Impl=armix::harmony::PitchSynchronousVoice<16384,6,Fast>;
Impl shifter;
std::array<float,16384> ring{};
std::array<double,measured> times{};
const int frames=(warmup+measured)*block;
std::vector<float> input(static_cast<std::size_t>(frames));
for(int t=0;t<frames;++t){
const double phase=pi2*f0*t/rate;
input[static_cast<std::size_t>(t)]=static_cast<float>(
.17*std::sin(phase)+.06*std::sin(2.*phase+.3)+
.028*std::sin(3.*phase+.17));
}
double sum=0.;
for(int b=0;b<warmup+measured;++b){
const auto before=Clock::now();
for(int j=0;j<block;++j){
const int t=b*block+j;
const std::size_t at=static_cast<std::size_t>(t)&16383;
ring[at]=input[static_cast<std::size_t>(t)];
for(int v=0;v<5;++v){
const auto output=shifter.process(
v,static_cast<std::uint64_t>(t),at,ring,
static_cast<float>(rate/f0),1.2f,
static_cast<float>(rate*.032)+static_cast<float>(v*rate*.004));
if(!std::isfinite(output.sample)){
std::cerr<<"PSOLA nonfinite\n";std::exit(2);
}
sum+=static_cast<double>(output.sample);
}
}
const auto after=Clock::now();
if(b>=warmup)
times[static_cast<std::size_t>(b-warmup)]=
std::chrono::duration<double,std::micro>(after-before).count();
}
std::sort(times.begin(),times.end());
const auto at=[&](double p)->double{
return times[static_cast<std::size_t>(std::ceil(p*measured)-1)];
};
return {at(.5),at(.95),at(.99),at(.999),times.back(),sum};
}
void compare(double rate,double f0){
// Order is constant, same binary and runner; scheduling noise remains.
const auto original=measure<false>(rate,f0);
const auto optimized=measure<true>(rate,f0);
if(std::abs(original.checksum-optimized.checksum)>1.e-9){
std::cerr<<"SAME INPUT PSOLA reference checksum changed\n";
std::exit(2);
}
const double deadline=1.e6*block/rate;
std::cout<<std::fixed<<std::setprecision(2)
<<"PSOLA_EPOCH_SYNTHETIC_ONLY rate="<<rate<<" f0="<<f0
<<" block="<<block<<" reference_p50_us="<<original.p50
<<" optimized_p50_us="<<optimized.p50
<<" reference_p95_us="<<original.p95
<<" optimized_p95_us="<<optimized.p95
<<" reference_p99_us="<<original.p99
<<" optimized_p99_us="<<optimized.p99
<<" reference_p99_9_us="<<original.p999
<<" optimized_p99_9_us="<<optimized.p999
<<" reference_max_us="<<original.maximum
<<" optimized_max_us="<<optimized.maximum
<<" optimized_p99_9_deadline_pct="<<100.*optimized.p999/deadline
<<" hardware_ASIO_qualified=false\n";
}
}
int main(){
for(double rate:{48000.,96000.})
for(double f0:{82.406889,220.})
compare(rate,f0);
std::cout<<"PSOLA same-source reference/optimized checksums match; "
"CPU comparisons non-gating\n";
}
14 changes: 10 additions & 4 deletions tests/harmony_pitch_synchronous_test.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -7,11 +7,13 @@
#include <vector>
using armix::harmony::PitchSynchronousVoice;
using Shifter=PitchSynchronousVoice<16384,6>;
using ReferenceShifter=PitchSynchronousVoice<16384,6,false>;
static void need(bool yes,const char* reason){
if(!yes){std::cerr<<reason<<"\n";std::exit(2);}
}
template <bool Fast>
static std::vector<float> render(double rate,double f0,double ratio,int block){
auto p=Shifter{};
auto p=PitchSynchronousVoice<16384,6,Fast>{};
std::array<float,16384> ring{};
const int n=static_cast<int>(rate*.65);
std::vector<float> out(static_cast<std::size_t>(n));
Expand Down Expand Up @@ -60,11 +62,15 @@ int main(){
{{48000.,82.40689,1.2}},{{96000.,440.,.75}},
{{192000.,880.,1.4}}
}}){
const auto a=render(v[0],v[1],v[2],64);
const auto b=render(v[0],v[1],v[2],257);
const auto a=render<true>(v[0],v[1],v[2],64);
const auto b=render<true>(v[0],v[1],v[2],257);
const auto legacy=render<false>(v[0],v[1],v[2],64);
need(a.size()==b.size(),"different source length");
for(std::size_t i=0;i<a.size();++i)
for(std::size_t i=0;i<a.size();++i){
need(a[i]==b[i],"callback partition changed PSOLA PCM");
need(a[i]==legacy[i],
"optimized epoch cursor changed historical PSOLA PCM");
}
const auto result=measuredHz(a,v[0],v[1]*v[2]);
need(std::abs(result-v[1]*v[2])/(v[1]*v[2])<.024,
"synchronous target pitch mismatch");
Expand Down
Loading