[Bf-blender-cvs] [8fbd71e] master: Cycles: improved Beckmann sampling using precomputed data
Brecht Van Lommel
noreply at git.blender.org
Sat Jun 21 22:36:27 CEST 2014
Commit: 8fbd71e5f2862e6efd0e3bcb015f1f6201ccca14
Author: Brecht Van Lommel
Date: Fri Jun 20 21:21:05 2014 +0200
https://developer.blender.org/rB8fbd71e5f2862e6efd0e3bcb015f1f6201ccca14
Cycles: improved Beckmann sampling using precomputed data
It turns out that the new Beckmann sampling function doesn't work well with
Quasi Monte Carlo sampling, mainly near normal incidence where it can be worse
than the previous sampler. In the new sampler the random number pattern gets
split in two, warped and overlapped, which hurts the stratification, see the
visualization in the differential revision.
Now we use a precomputed table, which is much better behaved. GGX does not seem
to benefit from using a precomputed table.
Disadvantage is that this table adds 1MB of memory usage and 0.03s startup time
to every render (on my quad core CPU).
Differential Revision: https://developer.blender.org/D614
===================================================================
M intern/cycles/kernel/closure/bsdf.h
M intern/cycles/kernel/closure/bsdf_microfacet.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/osl/osl_closures.cpp
M intern/cycles/kernel/osl/osl_closures.h
M intern/cycles/kernel/svm/svm_blackbody.h
M intern/cycles/render/shader.cpp
M intern/cycles/render/shader.h
===================================================================
diff --git a/intern/cycles/kernel/closure/bsdf.h b/intern/cycles/kernel/closure/bsdf.h
index 8ddf497..81c239e 100644
--- a/intern/cycles/kernel/closure/bsdf.h
+++ b/intern/cycles/kernel/closure/bsdf.h
@@ -85,13 +85,13 @@ ccl_device int bsdf_sample(KernelGlobals *kg, const ShaderData *sd, const Shader
case CLOSURE_BSDF_MICROFACET_GGX_ID:
case CLOSURE_BSDF_MICROFACET_GGX_ANISO_ID:
case CLOSURE_BSDF_MICROFACET_GGX_REFRACTION_ID:
- label = bsdf_microfacet_ggx_sample(sc, sd->Ng, sd->I, sd->dI.dx, sd->dI.dy, randu, randv,
+ label = bsdf_microfacet_ggx_sample(kg, sc, sd->Ng, sd->I, sd->dI.dx, sd->dI.dy, randu, randv,
eval, omega_in, &domega_in->dx, &domega_in->dy, pdf);
break;
case CLOSURE_BSDF_MICROFACET_BECKMANN_ID:
case CLOSURE_BSDF_MICROFACET_BECKMANN_ANISO_ID:
case CLOSURE_BSDF_MICROFACET_BECKMANN_REFRACTION_ID:
- label = bsdf_microfacet_beckmann_sample(sc, sd->Ng, sd->I, sd->dI.dx, sd->dI.dy, randu, randv,
+ label = bsdf_microfacet_beckmann_sample(kg, sc, sd->Ng, sd->I, sd->dI.dx, sd->dI.dy, randu, randv,
eval, omega_in, &domega_in->dx, &domega_in->dy, pdf);
break;
#ifdef __ANISOTROPIC__
diff --git a/intern/cycles/kernel/closure/bsdf_microfacet.h b/intern/cycles/kernel/closure/bsdf_microfacet.h
index 4a3d223..e130349 100644
--- a/intern/cycles/kernel/closure/bsdf_microfacet.h
+++ b/intern/cycles/kernel/closure/bsdf_microfacet.h
@@ -176,8 +176,8 @@ ccl_device float approx_erfinvf(float z)
* E. Heitz and E. d'Eon, EGSR 2014 */
ccl_device_inline void microfacet_beckmann_sample_slopes(
+ KernelGlobals *kg,
const float cos_theta_i, const float sin_theta_i,
- const float alpha_x, const float alpha_y,
float randu, float randv, float *slope_x, float *slope_y,
float *G1i)
{
@@ -200,10 +200,12 @@ ccl_device_inline void microfacet_beckmann_sample_slopes(
const float SQRT_PI_INV = 0.56418958354f;
const float Lambda = 0.5f*(erf_a - 1.0f) + (0.5f*SQRT_PI_INV)*(exp_a2*inv_a);
const float G1 = 1.0f/(1.0f + Lambda); /* masking */
- const float C = 1.0f - G1 * erf_a;
*G1i = G1;
+#if 0
+ const float C = 1.0f - G1 * erf_a;
+
/* sample slope X */
if(randu < C) {
/* rescale randu */
@@ -238,11 +240,20 @@ ccl_device_inline void microfacet_beckmann_sample_slopes(
/* sample slope Y */
*slope_y = approx_erfinvf(2.0f*randv - 1.0f);
+#else
+ /* use precomputed table, because it better preserves stratification
+ * of the random number pattern */
+ int beckmann_table_offset = kernel_data.tables.beckmann_offset;
+
+ *slope_x = lookup_table_read_2D(kg, randu, cos_theta_i,
+ beckmann_table_offset, BECKMANN_TABLE_SIZE, BECKMANN_TABLE_SIZE);
+ *slope_y = approx_erfinvf(2.0f*randv - 1.0f);
+#endif
+
}
ccl_device_inline void microfacet_ggx_sample_slopes(
const float cos_theta_i, const float sin_theta_i,
- const float alpha_x, const float alpha_y,
float randu, float randv, float *slope_x, float *slope_y,
float *G1i)
{
@@ -290,7 +301,8 @@ ccl_device_inline void microfacet_ggx_sample_slopes(
*slope_y = S * z * safe_sqrtf(1.0f + (*slope_x)*(*slope_x));
}
-ccl_device_inline float3 microfacet_sample_stretched(const float3 omega_i,
+ccl_device_inline float3 microfacet_sample_stretched(
+ KernelGlobals *kg, const float3 omega_i,
const float alpha_x, const float alpha_y,
const float randu, const float randv,
bool beckmann, float *G1i)
@@ -317,12 +329,14 @@ ccl_device_inline float3 microfacet_sample_stretched(const float3 omega_i,
/* 2. sample P22_{omega_i}(x_slope, y_slope, 1, 1) */
float slope_x, slope_y;
- if(beckmann)
- microfacet_beckmann_sample_slopes(costheta_, sintheta_,
- alpha_x, alpha_y, randu, randv, &slope_x, &slope_y, G1i);
- else
+ if(beckmann) {
+ microfacet_beckmann_sample_slopes(kg, costheta_, sintheta_,
+ randu, randv, &slope_x, &slope_y, G1i);
+ }
+ else {
microfacet_ggx_sample_slopes(costheta_, sintheta_,
- alpha_x, alpha_y, randu, randv, &slope_x, &slope_y, G1i);
+ randu, randv, &slope_x, &slope_y, G1i);
+ }
/* 3. rotate */
float tmp = cosphi_*slope_x - sinphi_*slope_y;
@@ -530,7 +544,7 @@ ccl_device float3 bsdf_microfacet_ggx_eval_transmit(const ShaderClosure *sc, con
return make_float3(out, out, out);
}
-ccl_device int bsdf_microfacet_ggx_sample(const ShaderClosure *sc, float3 Ng, float3 I, float3 dIdx, float3 dIdy, float randu, float randv, float3 *eval, float3 *omega_in, float3 *domega_in_dx, float3 *domega_in_dy, float *pdf)
+ccl_device int bsdf_microfacet_ggx_sample(KernelGlobals *kg, const ShaderClosure *sc, float3 Ng, float3 I, float3 dIdx, float3 dIdy, float randu, float randv, float3 *eval, float3 *omega_in, float3 *domega_in_dx, float3 *domega_in_dy, float *pdf)
{
float alpha_x = sc->data0;
float alpha_y = sc->data1;
@@ -552,7 +566,7 @@ ccl_device int bsdf_microfacet_ggx_sample(const ShaderClosure *sc, float3 Ng, fl
float3 local_m;
float G1o;
- local_m = microfacet_sample_stretched(local_I, alpha_x, alpha_y,
+ local_m = microfacet_sample_stretched(kg, local_I, alpha_x, alpha_y,
randu, randv, false, &G1o);
float3 m = X*local_m.x + Y*local_m.y + Z*local_m.z;
@@ -878,7 +892,7 @@ ccl_device float3 bsdf_microfacet_beckmann_eval_transmit(const ShaderClosure *sc
return make_float3(out, out, out);
}
-ccl_device int bsdf_microfacet_beckmann_sample(const ShaderClosure *sc, float3 Ng, float3 I, float3 dIdx, float3 dIdy, float randu, float randv, float3 *eval, float3 *omega_in, float3 *domega_in_dx, float3 *domega_in_dy, float *pdf)
+ccl_device int bsdf_microfacet_beckmann_sample(KernelGlobals *kg, const ShaderClosure *sc, float3 Ng, float3 I, float3 dIdx, float3 dIdy, float randu, float randv, float3 *eval, float3 *omega_in, float3 *domega_in_dx, float3 *domega_in_dy, float *pdf)
{
float alpha_x = sc->data0;
float alpha_y = sc->data1;
@@ -900,7 +914,7 @@ ccl_device int bsdf_microfacet_beckmann_sample(const ShaderClosure *sc, float3 N
float3 local_m;
float G1o;
- local_m = microfacet_sample_stretched(local_I, alpha_x, alpha_x,
+ local_m = microfacet_sample_stretched(kg, local_I, alpha_x, alpha_x,
randu, randv, true, &G1o);
float3 m = X*local_m.x + Y*local_m.y + Z*local_m.z;
diff --git a/intern/cycles/kernel/kernel_types.h b/intern/cycles/kernel/kernel_types.h
index 652a784..d81909a 100644
--- a/intern/cycles/kernel/kernel_types.h
+++ b/intern/cycles/kernel/kernel_types.h
@@ -44,6 +44,8 @@ CCL_NAMESPACE_BEGIN
#define BB_TABLE_YPOWER 5.0f
#define BB_TABLE_SPACING 2.0f
+#define BECKMANN_TABLE_SIZE 256
+
#define TEX_NUM_FLOAT_IMAGES 5
#define SHADER_NONE (~0)
@@ -933,11 +935,11 @@ typedef struct KernelCurves {
float maximum_width;
} KernelCurves;
-typedef struct KernelBlackbody {
- int table_offset;
- int pad1, pad2, pad3;
-} KernelBlackbody;
-
+typedef struct KernelTables {
+ int blackbody_offset;
+ int beckmann_offset;
+ int pad1, pad2;
+} KernelTables;
typedef struct KernelData {
KernelCamera cam;
@@ -946,7 +948,7 @@ typedef struct KernelData {
KernelIntegrator integrator;
KernelBVH bvh;
KernelCurves curve;
- KernelBlackbody blackbody;
+ KernelTables tables;
} KernelData;
CCL_NAMESPACE_END
diff --git a/intern/cycles/kernel/osl/osl_closures.cpp b/intern/cycles/kernel/osl/osl_closures.cpp
index c1560ea..f102e79 100644
--- a/intern/cycles/kernel/osl/osl_closures.cpp
+++ b/intern/cycles/kernel/osl/osl_closures.cpp
@@ -41,6 +41,8 @@
#include "util_param.h"
#include "kernel_types.h"
+#include "kernel_compat_cpu.h"
+#include "kernel_globals.h"
#include "kernel_montecarlo.h"
#include "closure/bsdf_util.h"
diff --git a/intern/cycles/kernel/osl/osl_closures.h b/intern/cycles/kernel/osl/osl_closures.h
index 218cf1c..a543907 100644
--- a/intern/cycles/kernel/osl/osl_closures.h
+++ b/intern/cycles/kernel/osl/osl_closures.h
@@ -149,17 +149,18 @@ public: \
\
void blur(float roughness) \
{ \
- bsdf_##svmlower##_blur(&sc, roughness); \
} \
\
float3 eval_reflect(const float3 &omega_out, const float3 &omega_in, float& pdf) const \
{ \
- return bsdf_##svmlower##_eval_reflect(&sc, omega_out, omega_in, &pdf); \
+ pdf = 0; \
+ return make_float3(0, 0, 0); \
} \
\
float3 eval_transmit(const float3 &omega_out, const float3 &omega_in, float& pdf) const \
{ \
- return bsdf_##svmlower##_eval_transmit(&sc, omega_out, omega_in, &pdf); \
+ pdf = 0; \
+ return make_float3(0, 0, 0); \
} \
\
int sample(const float3 &Ng, \
@@ -168,8 +169,8 @@ public: \
float3 &omega_in, float3 &domega_in_dx, float3 &domega_in_dy, \
float &pdf, float3 &eval) const \
{ \
- return bsdf_##svmlower##_sample(&sc, Ng, omega_out, domega_out_dx, domega_out_dy, \
- randu, randv, &eval, &omega_in, &domega_in_dx, &domega_in_dy, &pdf); \
+ pdf = 0; \
+ return LABEL_NONE; \
} \
}; \
\
diff --git a/intern/cycles/kernel/svm/svm_blackbody.h b/intern/cycles/kernel/svm/svm_blackbody.h
index 63dbf27..15257ae 100644
--- a/intern/cycles/kernel/svm/svm_blackbody.h
+++ b/intern/cycles/kernel/svm/svm_blackbody.h
@@ -55,7 +55,7 @@ ccl_device void svm_node_blackbody(KernelGlobals *kg, ShaderData *sd, float *sta
just one (the OSL-lerp is also automatically done for us by "lookup_table_read") */
float t = powf((temperature - BB_DRAPPER) * (1.0f / BB_TABLE_SPACING), (1.0f / BB_TABLE_XPOWER));
- int blackbody_table_offset = kernel_data.blackbody.table_offset;
+ int blackbody_table_offset = kernel_data.tables.blackbody_offset;
/* Retrieve colors from the lookup table */
float lutval = t*lookuptablenormalize;
diff --git a/intern/cycles/render/shader.cpp b/intern/cycles/render/shader.cpp
index 662caed..87107ae 100644
--- a/intern/cycles/render/shader.cpp
+++ b/intern/cycles/render/shader.cpp
@@ -31,6 +31,95 @@
CCL_NAMESPACE_BEGIN
+/* Beckmann sampling precomputed table, see bsdf_microfacet.h */
+
+/* 2D slope distribution (alpha = 1.0) */
+static float beckmann_table_P22(const float slope_x, const float slope_y)
+{
+ return expf(-(slope_x*slope_x + slope_y*slope_y));
+}
+
+/* maximal slope amplitude (range that contains 99.99% of the distribution) */
+static float beckmann_table_slope_max()
+{
+ return 6.0;
+}
+
+static void beckmann_table_rows(float *table, int row_from, int row_to)
+{
+ /* allocate temporary data */
+ const int DATA_TMP_
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list