use pcg32 for pssmlt

LuisaGroup · Jun 2, 2024 · 5e5f383 · 5e5f383
1 parent 163bb6f
commit 5e5f383
Showing 1 changed file with 10 additions and 10 deletions.
diff --git a/src/integrators/pssmlt.cpp b/src/integrators/pssmlt.cpp
@@ -44,7 +44,7 @@ class PSSMLTSampler {
 
 public:
     struct State {
-        UInt rng_state;
+        PCG32 rng;
         ULong current_iteration;
         Bool large_step;
         ULong last_large_step_iteration;
@@ -62,7 +62,7 @@ class PSSMLTSampler {
 private:
     uint _chains{};
     uint _pss_dim{};
-    Buffer<uint> _rng_buffer;
+    Buffer<ulong2> _rng_buffer;
     Buffer<ulong> _current_iteration_buffer;
     Buffer<uint> _large_step_and_initialized_dimensions_buffer;
     Buffer<ulong> _last_large_step_iteration_buffer;
@@ -77,7 +77,7 @@ class PSSMLTSampler {
                      "Too many primary samples.");
         command_buffer << synchronize();
         if (auto n = next_pow2(chains); n > _rng_buffer.size()) {
-            _rng_buffer = _device.create_buffer<uint>(n);
+            _rng_buffer = _device.create_buffer<ulong2>(n);
             _current_iteration_buffer = _device.create_buffer<ulong>(n);
             _large_step_and_initialized_dimensions_buffer = _device.create_buffer<uint>(n);
             _last_large_step_iteration_buffer = _device.create_buffer<ulong>(n);
@@ -152,19 +152,19 @@ class PSSMLTSampler {
         };
         // Reset Xi if a large step took place in the meantime
         $if (Xi.last_modification < _state->last_large_step_iteration) {
-            Xi.value = lcg(_state->rng_state);
+            Xi.value = _state->rng.uniform_float();
             Xi.last_modification = _state->last_large_step_iteration;
         };
         // Apply remaining sequence of mutations to _sample_
         Xi->backup();
         $if (_state->large_step) {
-            Xi.value = lcg(_state->rng_state);
+            Xi.value = _state->rng.uniform_float();
         }
         $else {
             auto nSmall = compute::cast<uint>(_state->current_iteration - Xi.last_modification);
             // Apply _nSmall_ small step mutations
             // Sample the standard normal distribution N(0, 1)
-            auto normalSample = sqrt_two * _erf_inv(2.f * lcg(_state->rng_state) - 1.f);
+            auto normalSample = sqrt_two * _erf_inv(2.f * _state->rng.uniform_float() - 1.f);
             // Compute the effective standard deviation and apply perturbation to Xi
             auto effSigma = _sigma * sqrt(cast<float>(nSmall));
             Xi.value = fract(Xi.value + normalSample * effSigma);
@@ -181,7 +181,7 @@ class PSSMLTSampler {
 
     void create(Expr<uint> chain_index, Expr<uint> rng_sequence) noexcept {
         _state = luisa::make_unique<State>(State{
-            .rng_state = xxhash32(rng_sequence),
+            .rng = PCG32{rng_sequence},
             .current_iteration = 0ull,
             .large_step = true,
             .last_large_step_iteration = 0ull,
@@ -196,7 +196,7 @@ class PSSMLTSampler {
         auto large_step_and_dimensions = _large_step_and_initialized_dimensions_buffer->read(chain_index);
         auto last_large_step_iteration = _last_large_step_iteration_buffer->read(chain_index);
         _state = luisa::make_unique<State>(State{
-            .rng_state = rng_state,
+            .rng = PCG32{rng_state.x, rng_state.y},
             .current_iteration = current_iteration,
             .large_step = (large_step_and_dimensions & 1u) != 0u,
             .last_large_step_iteration = last_large_step_iteration,
@@ -206,7 +206,7 @@ class PSSMLTSampler {
     }
 
     void save() noexcept {
-        _rng_buffer->write(_state->chain_index, _state->rng_state);
+        _rng_buffer->write(_state->chain_index, make_ulong2(_state->rng.state(), _state->rng.inc()));
         _current_iteration_buffer->write(_state->chain_index, _state->current_iteration);
         _large_step_and_initialized_dimensions_buffer->write(
             _state->chain_index, ite(_state->large_step, 1u, 0u) | (_state->initialized_dimensions << 1u));
@@ -249,7 +249,7 @@ class PSSMLTSampler {
 
     void start_iteration() noexcept {
         _state->current_iteration = _state->current_iteration + 1ull;
-        _state->large_step = lcg(_state->rng_state) < _large_step_probability;
+        _state->large_step = _state->rng.uniform_float() < _large_step_probability;
     }
 };