diff options
Diffstat (limited to 'intern/cycles/kernel/kernel_passes.h')
| -rw-r--r-- | intern/cycles/kernel/kernel_passes.h | 277 |
1 files changed, 232 insertions, 45 deletions
diff --git a/intern/cycles/kernel/kernel_passes.h b/intern/cycles/kernel/kernel_passes.h index 7aec47e4957..b31356905f2 100644 --- a/intern/cycles/kernel/kernel_passes.h +++ b/intern/cycles/kernel/kernel_passes.h @@ -16,19 +16,23 @@ CCL_NAMESPACE_BEGIN -ccl_device_inline void kernel_write_pass_float(ccl_global float *buffer, int sample, float value) +#if defined(__SPLIT_KERNEL__) || defined(__KERNEL_CUDA__) +#define __ATOMIC_PASS_WRITE__ +#endif + +ccl_device_inline void kernel_write_pass_float(ccl_global float *buffer, float value) { ccl_global float *buf = buffer; -#if defined(__SPLIT_KERNEL__) && defined(__WORK_STEALING__) +#ifdef __ATOMIC_PASS_WRITE__ atomic_add_and_fetch_float(buf, value); #else - *buf = (sample == 0)? value: *buf + value; -#endif // __SPLIT_KERNEL__ && __WORK_STEALING__ + *buf += value; +#endif } -ccl_device_inline void kernel_write_pass_float3(ccl_global float *buffer, int sample, float3 value) +ccl_device_inline void kernel_write_pass_float3(ccl_global float *buffer, float3 value) { -#if defined(__SPLIT_KERNEL__) && defined(__WORK_STEALING__) +#ifdef __ATOMIC_PASS_WRITE__ ccl_global float *buf_x = buffer + 0; ccl_global float *buf_y = buffer + 1; ccl_global float *buf_z = buffer + 2; @@ -38,13 +42,13 @@ ccl_device_inline void kernel_write_pass_float3(ccl_global float *buffer, int sa atomic_add_and_fetch_float(buf_z, value.z); #else ccl_global float3 *buf = (ccl_global float3*)buffer; - *buf = (sample == 0)? value: *buf + value; -#endif // __SPLIT_KERNEL__ && __WORK_STEALING__ + *buf += value; +#endif } -ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, int sample, float4 value) +ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, float4 value) { -#if defined(__SPLIT_KERNEL__) && defined(__WORK_STEALING__) +#ifdef __ATOMIC_PASS_WRITE__ ccl_global float *buf_x = buffer + 0; ccl_global float *buf_y = buffer + 1; ccl_global float *buf_z = buffer + 2; @@ -56,12 +60,137 @@ ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, int sa atomic_add_and_fetch_float(buf_w, value.w); #else ccl_global float4 *buf = (ccl_global float4*)buffer; - *buf = (sample == 0)? value: *buf + value; -#endif // __SPLIT_KERNEL__ && __WORK_STEALING__ + *buf += value; +#endif +} + +#ifdef __DENOISING_FEATURES__ +ccl_device_inline void kernel_write_pass_float_variance(ccl_global float *buffer, float value) +{ + kernel_write_pass_float(buffer, value); + + /* The online one-pass variance update that's used for the megakernel can't easily be implemented + * with atomics, so for the split kernel the E[x^2] - 1/N * (E[x])^2 fallback is used. */ + kernel_write_pass_float(buffer+1, value*value); } +# ifdef __ATOMIC_PASS_WRITE__ +# define kernel_write_pass_float3_unaligned kernel_write_pass_float3 +# else +ccl_device_inline void kernel_write_pass_float3_unaligned(ccl_global float *buffer, float3 value) +{ + buffer[0] += value.x; + buffer[1] += value.y; + buffer[2] += value.z; +} +# endif + +ccl_device_inline void kernel_write_pass_float3_variance(ccl_global float *buffer, float3 value) +{ + kernel_write_pass_float3_unaligned(buffer, value); + kernel_write_pass_float3_unaligned(buffer+3, value*value); +} + +ccl_device_inline void kernel_write_denoising_shadow(KernelGlobals *kg, ccl_global float *buffer, + int sample, float path_total, float path_total_shaded) +{ + if(kernel_data.film.pass_denoising_data == 0) + return; + + buffer += (sample & 1)? DENOISING_PASS_SHADOW_B : DENOISING_PASS_SHADOW_A; + + path_total = ensure_finite(path_total); + path_total_shaded = ensure_finite(path_total_shaded); + + kernel_write_pass_float(buffer, path_total); + kernel_write_pass_float(buffer+1, path_total_shaded); + + float value = path_total_shaded / max(path_total, 1e-7f); + kernel_write_pass_float(buffer+2, value*value); +} +#endif /* __DENOISING_FEATURES__ */ + +ccl_device_inline void kernel_update_denoising_features(KernelGlobals *kg, + ShaderData *sd, + ccl_addr_space PathState *state, + PathRadiance *L) +{ +#ifdef __DENOISING_FEATURES__ + if(state->denoising_feature_weight == 0.0f) { + return; + } + + L->denoising_depth += ensure_finite(state->denoising_feature_weight * sd->ray_length); + + /* Skip implicitly transparent surfaces. */ + if(sd->flag & SD_HAS_ONLY_VOLUME) { + return; + } + + float3 normal = make_float3(0.0f, 0.0f, 0.0f); + float3 albedo = make_float3(0.0f, 0.0f, 0.0f); + float sum_weight = 0.0f, sum_nonspecular_weight = 0.0f; + + for(int i = 0; i < sd->num_closure; i++) { + ShaderClosure *sc = &sd->closure[i]; + + if(!CLOSURE_IS_BSDF_OR_BSSRDF(sc->type)) + continue; + + /* All closures contribute to the normal feature, but only diffuse-like ones to the albedo. */ + normal += sc->N * sc->sample_weight; + sum_weight += sc->sample_weight; + if(!bsdf_is_specular_like(sc)) { + albedo += sc->weight; + sum_nonspecular_weight += sc->sample_weight; + } + } + + /* Wait for next bounce if 75% or more sample weight belongs to specular-like closures. */ + if((sum_weight == 0.0f) || (sum_nonspecular_weight*4.0f > sum_weight)) { + if(sum_weight != 0.0f) { + normal /= sum_weight; + } + L->denoising_normal += ensure_finite3(state->denoising_feature_weight * normal); + L->denoising_albedo += ensure_finite3(state->denoising_feature_weight * albedo); + + state->denoising_feature_weight = 0.0f; + } +#else + (void) kg; + (void) sd; + (void) state; + (void) L; +#endif /* __DENOISING_FEATURES__ */ +} + +#ifdef __KERNEL_DEBUG__ +ccl_device_inline void kernel_write_debug_passes(KernelGlobals *kg, + ccl_global float *buffer, + PathRadiance *L) +{ + int flag = kernel_data.film.pass_flag; + if(flag & PASS_BVH_TRAVERSED_NODES) { + kernel_write_pass_float(buffer + kernel_data.film.pass_bvh_traversed_nodes, + L->debug_data.num_bvh_traversed_nodes); + } + if(flag & PASS_BVH_TRAVERSED_INSTANCES) { + kernel_write_pass_float(buffer + kernel_data.film.pass_bvh_traversed_instances, + L->debug_data.num_bvh_traversed_instances); + } + if(flag & PASS_BVH_INTERSECTIONS) { + kernel_write_pass_float(buffer + kernel_data.film.pass_bvh_intersections, + L->debug_data.num_bvh_intersections); + } + if(flag & PASS_RAY_BOUNCES) { + kernel_write_pass_float(buffer + kernel_data.film.pass_ray_bounces, + L->debug_data.num_ray_bounces); + } +} +#endif /* __KERNEL_DEBUG__ */ + ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global float *buffer, PathRadiance *L, - ShaderData *sd, int sample, ccl_addr_space PathState *state, float3 throughput) + ShaderData *sd, ccl_addr_space PathState *state, float3 throughput) { #ifdef __PASSES__ int path_flag = state->flag; @@ -75,38 +204,37 @@ ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global fl return; if(!(path_flag & PATH_RAY_SINGLE_PASS_DONE)) { - if(!(ccl_fetch(sd, flag) & SD_TRANSPARENT) || + if(!(sd->flag & SD_TRANSPARENT) || kernel_data.film.pass_alpha_threshold == 0.0f || average(shader_bsdf_alpha(kg, sd)) >= kernel_data.film.pass_alpha_threshold) { - - if(sample == 0) { + if(state->sample == 0) { if(flag & PASS_DEPTH) { - float depth = camera_distance(kg, ccl_fetch(sd, P)); - kernel_write_pass_float(buffer + kernel_data.film.pass_depth, sample, depth); + float depth = camera_distance(kg, sd->P); + kernel_write_pass_float(buffer + kernel_data.film.pass_depth, depth); } if(flag & PASS_OBJECT_ID) { - float id = object_pass_id(kg, ccl_fetch(sd, object)); - kernel_write_pass_float(buffer + kernel_data.film.pass_object_id, sample, id); + float id = object_pass_id(kg, sd->object); + kernel_write_pass_float(buffer + kernel_data.film.pass_object_id, id); } if(flag & PASS_MATERIAL_ID) { float id = shader_pass_id(kg, sd); - kernel_write_pass_float(buffer + kernel_data.film.pass_material_id, sample, id); + kernel_write_pass_float(buffer + kernel_data.film.pass_material_id, id); } } if(flag & PASS_NORMAL) { - float3 normal = ccl_fetch(sd, N); - kernel_write_pass_float3(buffer + kernel_data.film.pass_normal, sample, normal); + float3 normal = shader_bsdf_average_normal(kg, sd); + kernel_write_pass_float3(buffer + kernel_data.film.pass_normal, normal); } if(flag & PASS_UV) { float3 uv = primitive_uv(kg, sd); - kernel_write_pass_float3(buffer + kernel_data.film.pass_uv, sample, uv); + kernel_write_pass_float3(buffer + kernel_data.film.pass_uv, uv); } if(flag & PASS_MOTION) { float4 speed = primitive_motion_vector(kg, sd); - kernel_write_pass_float4(buffer + kernel_data.film.pass_motion, sample, speed); - kernel_write_pass_float(buffer + kernel_data.film.pass_motion_weight, sample, 1.0f); + kernel_write_pass_float4(buffer + kernel_data.film.pass_motion, speed); + kernel_write_pass_float(buffer + kernel_data.film.pass_motion_weight, 1.0f); } state->flag |= PATH_RAY_SINGLE_PASS_DONE; @@ -127,7 +255,7 @@ ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global fl float mist_start = kernel_data.film.mist_start; float mist_inv_depth = kernel_data.film.mist_inv_depth; - float depth = camera_distance(kg, ccl_fetch(sd, P)); + float depth = camera_distance(kg, sd->P); float mist = saturate((depth - mist_start)*mist_inv_depth); /* falloff */ @@ -149,7 +277,7 @@ ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global fl #endif } -ccl_device_inline void kernel_write_light_passes(KernelGlobals *kg, ccl_global float *buffer, PathRadiance *L, int sample) +ccl_device_inline void kernel_write_light_passes(KernelGlobals *kg, ccl_global float *buffer, PathRadiance *L) { #ifdef __PASSES__ int flag = kernel_data.film.pass_flag; @@ -158,44 +286,103 @@ ccl_device_inline void kernel_write_light_passes(KernelGlobals *kg, ccl_global f return; if(flag & PASS_DIFFUSE_INDIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_indirect, sample, L->indirect_diffuse); + kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_indirect, L->indirect_diffuse); if(flag & PASS_GLOSSY_INDIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_indirect, sample, L->indirect_glossy); + kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_indirect, L->indirect_glossy); if(flag & PASS_TRANSMISSION_INDIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_indirect, sample, L->indirect_transmission); + kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_indirect, L->indirect_transmission); if(flag & PASS_SUBSURFACE_INDIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_indirect, sample, L->indirect_subsurface); + kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_indirect, L->indirect_subsurface); if(flag & PASS_DIFFUSE_DIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_direct, sample, L->direct_diffuse); + kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_direct, L->direct_diffuse); if(flag & PASS_GLOSSY_DIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_direct, sample, L->direct_glossy); + kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_direct, L->direct_glossy); if(flag & PASS_TRANSMISSION_DIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_direct, sample, L->direct_transmission); + kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_direct, L->direct_transmission); if(flag & PASS_SUBSURFACE_DIRECT) - kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_direct, sample, L->direct_subsurface); + kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_direct, L->direct_subsurface); if(flag & PASS_EMISSION) - kernel_write_pass_float3(buffer + kernel_data.film.pass_emission, sample, L->emission); + kernel_write_pass_float3(buffer + kernel_data.film.pass_emission, L->emission); if(flag & PASS_BACKGROUND) - kernel_write_pass_float3(buffer + kernel_data.film.pass_background, sample, L->background); + kernel_write_pass_float3(buffer + kernel_data.film.pass_background, L->background); if(flag & PASS_AO) - kernel_write_pass_float3(buffer + kernel_data.film.pass_ao, sample, L->ao); + kernel_write_pass_float3(buffer + kernel_data.film.pass_ao, L->ao); if(flag & PASS_DIFFUSE_COLOR) - kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_color, sample, L->color_diffuse); + kernel_write_pass_float3(buffer + kernel_data.film.pass_diffuse_color, L->color_diffuse); if(flag & PASS_GLOSSY_COLOR) - kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_color, sample, L->color_glossy); + kernel_write_pass_float3(buffer + kernel_data.film.pass_glossy_color, L->color_glossy); if(flag & PASS_TRANSMISSION_COLOR) - kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_color, sample, L->color_transmission); + kernel_write_pass_float3(buffer + kernel_data.film.pass_transmission_color, L->color_transmission); if(flag & PASS_SUBSURFACE_COLOR) - kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_color, sample, L->color_subsurface); + kernel_write_pass_float3(buffer + kernel_data.film.pass_subsurface_color, L->color_subsurface); if(flag & PASS_SHADOW) { float4 shadow = L->shadow; shadow.w = kernel_data.film.pass_shadow_scale; - kernel_write_pass_float4(buffer + kernel_data.film.pass_shadow, sample, shadow); + kernel_write_pass_float4(buffer + kernel_data.film.pass_shadow, shadow); } if(flag & PASS_MIST) - kernel_write_pass_float(buffer + kernel_data.film.pass_mist, sample, 1.0f - L->mist); + kernel_write_pass_float(buffer + kernel_data.film.pass_mist, 1.0f - L->mist); +#endif +} + +ccl_device_inline void kernel_write_result(KernelGlobals *kg, + ccl_global float *buffer, + int sample, + PathRadiance *L) +{ + float alpha; + float3 L_sum = path_radiance_clamp_and_sum(kg, L, &alpha); + + kernel_write_pass_float4(buffer, make_float4(L_sum.x, L_sum.y, L_sum.z, alpha)); + + kernel_write_light_passes(kg, buffer, L); + +#ifdef __DENOISING_FEATURES__ + if(kernel_data.film.pass_denoising_data) { +# ifdef __SHADOW_TRICKS__ + kernel_write_denoising_shadow(kg, + buffer + kernel_data.film.pass_denoising_data, + sample, + average(L->path_total), + average(L->path_total_shaded)); +# else + kernel_write_denoising_shadow(kg, + buffer + kernel_data.film.pass_denoising_data, + sample, + 0.0f, 0.0f); +# endif + if(kernel_data.film.pass_denoising_clean) { + float3 noisy, clean; + path_radiance_split_denoising(kg, L, &noisy, &clean); + kernel_write_pass_float3_variance( + buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_COLOR, + noisy); + kernel_write_pass_float3_unaligned( + buffer + kernel_data.film.pass_denoising_clean, + clean); + } + else { + kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_COLOR, + ensure_finite3(L_sum)); + } + + kernel_write_pass_float3_variance( + buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_NORMAL, + L->denoising_normal); + kernel_write_pass_float3_variance( + buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_ALBEDO, + L->denoising_albedo); + kernel_write_pass_float_variance( + buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_DEPTH, + L->denoising_depth); + } +#endif /* __DENOISING_FEATURES__ */ + + +#ifdef __KERNEL_DEBUG__ + kernel_write_debug_passes(kg, buffer, L); #endif } |