author Sergey Sharybin Tue, 25 Apr 2017 09:42:36 +0000 (11:42 +0200) committer Sergey Sharybin Tue, 25 Apr 2017 09:43:20 +0000 (11:43 +0200)
Was doing lots of investigation recently, with need to have lots of things
side by side.

@@ -20,14 +20,15 @@ CCL_NAMESPACE_BEGIN

#ifdef __SOBOL__

-/* skip initial numbers that are not as well distributed, especially the
+/* Skip initial numbers that are not as well distributed, especially the
* first sequence is just 0 everywhere, which can be problematic for e.g.
- * path termination */
+ * path termination.
+ */
#define SOBOL_SKIP 64

-/* High Dimensional Sobol */
+/* High Dimensional Sobol. */

-/* van der corput radical inverse */
+/* Van der Corput radical inverse. */
ccl_device uint van_der_corput(uint bits)
{
bits = (bits << 16) | (bits >> 16);
@@ -38,58 +39,63 @@ ccl_device uint van_der_corput(uint bits)
return bits;
}

ccl_device uint sobol(uint i)
{
uint r = 0;
-
-       for(uint v = 1U << 31; i; i >>= 1, v ^= v >> 1)
-               if(i & 1)
+       for(uint v = 1U << 31; i; i >>= 1, v ^= v >> 1) {
+               if(i & 1) {
r ^= v;
-
+               }
+       }
return r;
}

-/* inverse of sobol radical inverse */
+/* Inverse of sobol radical inverse. */
ccl_device uint sobol_inverse(uint i)
{
const uint msb = 1U << 31;
uint r = 0;
-
-       for(uint v = 1; i; i <<= 1, v ^= v << 1)
-               if(i & msb)
+       for(uint v = 1; i; i <<= 1, v ^= v << 1) {
+               if(i & msb) {
r ^= v;
-
+               }
+       }
return r;
}

-/* multidimensional sobol with generator matrices
- * dimension 0 and 1 are equal to van_der_corput() and sobol() respectively */
+/* Multidimensional sobol with generator matrices
+ * dimension 0 and 1 are equal to van_der_corput() and sobol() respectively.
+ */
ccl_device uint sobol_dimension(KernelGlobals *kg, int index, int dimension)
{
uint result = 0;
uint i = index;
-
-       for(uint j = 0; i; i >>= 1, j++)
-               if(i & 1)
+       for(uint j = 0; i; i >>= 1, j++) {
+               if(i & 1) {
result ^= kernel_tex_fetch(__sobol_directions, 32*dimension + j);
-
+               }
+       }
return result;
}

-/* lookup index and x/y coordinate, assumes m is a power of two */
-ccl_device uint sobol_lookup(const uint m, const uint frame, const uint ex, const uint ey, uint *x, uint *y)
+/* Lookup index and x/y coordinate, assumes m is a power of two. */
+ccl_device uint sobol_lookup(const uint m,
+                             const uint frame,
+                             const uint ex,
+                             const uint ey,
+                             uint *x, uint *y)
{
-       /* shift is constant per frame */
+       /* Shift is constant per frame. */
const uint shift = frame << (m << 1);
const uint sobol_shift = sobol(shift);
-       /* van der Corput is its own inverse */
+       /* Van der Corput is its own inverse. */
const uint lower = van_der_corput(ex << (32 - m));
-       /* need to compensate for ey difference and shift */
+       /* Need to compensate for ey difference and shift. */
const uint sobol_lower = sobol(lower);
-       const uint mask = ~-(1 << m) << (32 - m); /* only m upper bits */
+       const uint mask = ~-(1 << m) << (32 - m);  /* Only m upper bits. */
const uint delta = ((ey << (32 - m)) ^ sobol_lower ^ sobol_shift) & mask;
-       /* only use m upper bits for the index (m is a power of two) */
+       /* Only use m upper bits for the index (m is a power of two). */
const uint sobol_result = delta | (delta >> m);
const uint upper = sobol_inverse(sobol_result);
const uint index = shift | upper | lower;
@@ -98,11 +104,14 @@ ccl_device uint sobol_lookup(const uint m, const uint frame, const uint ex, cons
return index;
}

-ccl_device_forceinline float path_rng_1D(KernelGlobals *kg, RNG *rng, int sample, int num_samples, int dimension)
+ccl_device_forceinline float path_rng_1D(KernelGlobals *kg,
+                                         RNG *rng,
+                                         int sample, int num_samples,
+                                         int dimension)
{
#ifdef __CMJ__
if(kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_CMJ) {
-               /* correlated multi-jittered */
+               /* Correlated multi-jitter. */
int p = *rng + dimension;
return cmj_sample_1D(sample, num_samples, p);
}
@@ -113,7 +122,7 @@ ccl_device_forceinline float path_rng_1D(KernelGlobals *kg, RNG *rng, int sample
float r = (float)result * (1.0f/(float)0xFFFFFFFF);
return r;
#else
-       /* compute sobol sequence value using direction vectors */
+       /* Compute sobol sequence value using direction vectors. */
uint result = sobol_dimension(kg, sample + SOBOL_SKIP, dimension);
float r = (float)result * (1.0f/(float)0xFFFFFFFF);

@@ -130,24 +139,33 @@ ccl_device_forceinline float path_rng_1D(KernelGlobals *kg, RNG *rng, int sample
#endif
}

-ccl_device_forceinline void path_rng_2D(KernelGlobals *kg, RNG *rng, int sample, int num_samples, int dimension, float *fx, float *fy)
+ccl_device_forceinline void path_rng_2D(KernelGlobals *kg,
+                                        RNG *rng,
+                                        int sample, int num_samples,
+                                        int dimension,
+                                        float *fx, float *fy)
{
#ifdef __CMJ__
if(kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_CMJ) {
-               /* correlated multi-jittered */
+               /* Correlated multi-jitter. */
int p = *rng + dimension;
cmj_sample_2D(sample, num_samples, p, fx, fy);
}
else
#endif
{
-               /* sobol */
+               /* Sobol. */
*fx = path_rng_1D(kg, rng, sample, num_samples, dimension);
*fy = path_rng_1D(kg, rng, sample, num_samples, dimension + 1);
}
}

-ccl_device_inline void path_rng_init(KernelGlobals *kg, ccl_global uint *rng_state, int sample, int num_samples, RNG *rng, int x, int y, float *fx, float *fy)
+ccl_device_inline void path_rng_init(KernelGlobals *kg,
+                                     ccl_global uint *rng_state,
+                                     int sample, int num_samples,
+                                     RNG *rng,
+                                     int x, int y,
+                                     float *fx, float *fy)
{
#ifdef __SOBOL_FULL_SCREEN__
uint px, py;
@@ -182,29 +200,43 @@ ccl_device_inline void path_rng_init(KernelGlobals *kg, ccl_global uint *rng_sta
#endif
}

-ccl_device void path_rng_end(KernelGlobals *kg, ccl_global uint *rng_state, RNG rng)
+ccl_device void path_rng_end(KernelGlobals *kg,
+                             ccl_global uint *rng_state,
+                             RNG rng)
{
/* nothing to do */
}

-#else
+#else  /* __SOBOL__ */

/* Linear Congruential Generator */

-ccl_device_forceinline float path_rng_1D(KernelGlobals *kg, RNG *rng, int sample, int num_samples, int dimension)
+ccl_device_forceinline float path_rng_1D(KernelGlobals *kg,
+                                         RNG *rng,
+                                         int sample, int num_samples,
+                                         int dimension)
{
/* implicit mod 2^32 */
*rng = (1103515245*(*rng) + 12345);
return (float)*rng * (1.0f/(float)0xFFFFFFFF);
}

-ccl_device_inline void path_rng_2D(KernelGlobals *kg, RNG *rng, int sample, int num_samples, int dimension, float *fx, float *fy)
+ccl_device_inline void path_rng_2D(KernelGlobals *kg,
+                                   RNG *rng,
+                                   int sample, int num_samples,
+                                   int dimension,
+                                   float *fx, float *fy)
{
*fx = path_rng_1D(kg, rng, sample, num_samples, dimension);
*fy = path_rng_1D(kg, rng, sample, num_samples, dimension + 1);
}

-ccl_device void path_rng_init(KernelGlobals *kg, ccl_global uint *rng_state, int sample, int num_samples, RNG *rng, int x, int y, float *fx, float *fy)
+ccl_device void path_rng_init(KernelGlobals *kg,
+                              ccl_global uint *rng_state,
+                              int sample, int num_samples,
+                              RNG *rng,
+                              int x, int y,
+                              float *fx, float *fy)
{
*rng = *rng_state;
@@ -220,13 +252,15 @@ ccl_device void path_rng_init(KernelGlobals *kg, ccl_global uint *rng_state, int
}
}

-ccl_device void path_rng_end(KernelGlobals *kg, ccl_global uint *rng_state, RNG rng)
+ccl_device void path_rng_end(KernelGlobals *kg,
+                             ccl_global uint *rng_state,
+                             RNG rng)
{
/* store state for next sample */
*rng_state = rng;
}

-#endif
+#endif  /* __SOBOL__ */

/* Linear Congruential Generator */

@@ -257,49 +291,108 @@ ccl_device uint lcg_init(uint seed)
* dimension to avoid using the same sequence twice.
*
* For branches in the path we must be careful not to reuse the same number
- * in a sequence and offset accordingly. */
+ * in a sequence and offset accordingly.
+ */

-ccl_device_inline float path_state_rng_1D(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int dimension)
+ccl_device_inline float path_state_rng_1D(KernelGlobals *kg,
+                                          RNG *rng,
+                                          int dimension)
{
-       return path_rng_1D(kg, rng, state->sample, state->num_samples, state->rng_offset + dimension);
+       return path_rng_1D(kg,
+                          rng,
+                          state->sample, state->num_samples,
+                          state->rng_offset + dimension);
}

-ccl_device_inline float path_state_rng_1D_for_decision(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int dimension)
+ccl_device_inline float path_state_rng_1D_for_decision(
+        KernelGlobals *kg,
+        RNG *rng,
+        int dimension)
{
-       /* the rng_offset is not increased for transparent bounces. if we do then
+       /* The rng_offset is not increased for transparent bounces. if we do then
* fully transparent objects can become subtly visible by the different
* sampling patterns used where the transparent object is.
*
* however for some random numbers that will determine if we next bounce
* is transparent we do need to increase the offset to avoid always making
-        * the same decision */
-       int rng_offset = state->rng_offset + state->transparent_bounce*PRNG_BOUNCE_NUM;
-       return path_rng_1D(kg, rng, state->sample, state->num_samples, rng_offset + dimension);
+        * the same decision. */
+       const int rng_offset = state->rng_offset + state->transparent_bounce * PRNG_BOUNCE_NUM;
+       return path_rng_1D(kg,
+                          rng,
+                          state->sample, state->num_samples,
+                          rng_offset + dimension);
}

-ccl_device_inline void path_state_rng_2D(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int dimension, float *fx, float *fy)
+ccl_device_inline void path_state_rng_2D(KernelGlobals *kg,
+                                         RNG *rng,
+                                         int dimension,
+                                         float *fx, float *fy)
{
-       path_rng_2D(kg, rng, state->sample, state->num_samples, state->rng_offset + dimension, fx, fy);
+       path_rng_2D(kg,
+                   rng,
+                   state->sample, state->num_samples,
+                   state->rng_offset + dimension,
+                   fx, fy);
}

-ccl_device_inline float path_branched_rng_1D(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int branch, int num_branches, int dimension)
+ccl_device_inline float path_branched_rng_1D(
+        KernelGlobals *kg,
+        RNG *rng,
+        int branch,
+        int num_branches,
+        int dimension)
{
-       return path_rng_1D(kg, rng, state->sample*num_branches + branch, state->num_samples*num_branches, state->rng_offset + dimension);
+       return path_rng_1D(kg,
+                          rng,
+                          state->sample * num_branches + branch,
+                          state->num_samples * num_branches,
+                          state->rng_offset + dimension);
}

-ccl_device_inline float path_branched_rng_1D_for_decision(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int branch, int num_branches, int dimension)
+ccl_device_inline float path_branched_rng_1D_for_decision(
+        KernelGlobals *kg,
+        RNG *rng,
+        int branch,
+        int num_branches,
+        int dimension)
{
-       int rng_offset = state->rng_offset + state->transparent_bounce*PRNG_BOUNCE_NUM;
-       return path_rng_1D(kg, rng, state->sample*num_branches + branch, state->num_samples*num_branches, rng_offset + dimension);
+       const int rng_offset = state->rng_offset + state->transparent_bounce * PRNG_BOUNCE_NUM;
+       return path_rng_1D(kg,
+                          rng,
+                          state->sample * num_branches + branch,
+                          state->num_samples * num_branches,
+                          rng_offset + dimension);
}

-ccl_device_inline void path_branched_rng_2D(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int branch, int num_branches, int dimension, float *fx, float *fy)
+ccl_device_inline void path_branched_rng_2D(
+        KernelGlobals *kg,
+        RNG *rng,
+        int branch,
+        int num_branches,
+        int dimension,
+        float *fx, float *fy)
{
-       path_rng_2D(kg, rng, state->sample*num_branches + branch, state->num_samples*num_branches, state->rng_offset + dimension, fx, fy);
+       path_rng_2D(kg,
+                   rng,
+                   state->sample * num_branches + branch,
+                   state->num_samples * num_branches,
+                   state->rng_offset + dimension,
+                   fx, fy);
}

-/* Utitility functions to get light termination value, since it might not be needed in many cases. */
-ccl_device_inline float path_state_rng_light_termination(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state)
+/* Utitility functions to get light termination value,
+ * since it might not be needed in many cases.
+ */
+ccl_device_inline float path_state_rng_light_termination(
+        KernelGlobals *kg,
+        RNG *rng,
{
if(kernel_data.integrator.light_inv_rr_threshold > 0.0f) {
return path_state_rng_1D_for_decision(kg, rng, state, PRNG_LIGHT_TERMINATE);
@@ -307,15 +400,27 @@ ccl_device_inline float path_state_rng_light_termination(KernelGlobals *kg, RNG
return 0.0f;
}

-ccl_device_inline float path_branched_rng_light_termination(KernelGlobals *kg, RNG *rng, const ccl_addr_space PathState *state, int branch, int num_branches)
+ccl_device_inline float path_branched_rng_light_termination(
+        KernelGlobals *kg,
+        RNG *rng,
+        int branch,
+        int num_branches)
{
if(kernel_data.integrator.light_inv_rr_threshold > 0.0f) {
-               return path_branched_rng_1D_for_decision(kg, rng, state, branch, num_branches, PRNG_LIGHT_TERMINATE);
+               return path_branched_rng_1D_for_decision(kg,
+                                                        rng,
+                                                        state,
+                                                        branch,
+                                                        num_branches,
+                                                        PRNG_LIGHT_TERMINATE);
}
return 0.0f;
}

-ccl_device_inline void path_state_branch(ccl_addr_space PathState *state, int branch, int num_branches)
+                                         int branch,
+                                         int num_branches)
{
/* path is splitting into a branch, adjust so that each branch
* still gets a unique sample from the same sequence */
@@ -324,14 +429,17 @@ ccl_device_inline void path_state_branch(ccl_addr_space PathState *state, int br
state->num_samples = state->num_samples*num_branches;
}

-ccl_device_inline uint lcg_state_init(RNG *rng, int rng_offset, int sample, uint scramble)
+ccl_device_inline uint lcg_state_init(RNG *rng,
+                                      int rng_offset,
+                                      int sample,
+                                      uint scramble)
{
return lcg_init(*rng + rng_offset + sample*scramble);
}