]> de.git.xonotic.org Git - xonotic/darkplaces.git/commitdiff
Overhauled r_shadow_bouncegrid, it performs much faster, makes use of as many threads...
authorhavoc <havoc@d7cf8633-e32d-0410-b094-e92efae38249>
Fri, 17 Jan 2020 06:22:09 +0000 (06:22 +0000)
committerhavoc <havoc@d7cf8633-e32d-0410-b094-e92efae38249>
Fri, 17 Jan 2020 06:22:09 +0000 (06:22 +0000)
r_shadow_bouncegrid now does refractive index calculations on water surfaces, it's not very obvious yet because the traceline doesn't return texture information (if it did, there would be cool caustics effects in dynamic mode).

Implemented TaskQueue system for simple threaded tasks (fibers but less featureful), this is used by r_shadow_bouncegrid but will be used by other parts of the engine in future, also added Thread_Atomic* functions.

git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@12493 d7cf8633-e32d-0410-b094-e92efae38249

host.c
model_brush.c
model_shared.c
model_shared.h
r_shadow.c
r_shadow.h
sys_shared.c
thread.h
thread_sdl.c

diff --git a/host.c b/host.c
index a83d072c36d1ee6ba37ca2ad8393558591294bbe..61e84bc2e196a0b81c280416ef2950dc0635a320 100644 (file)
--- a/host.c
+++ b/host.c
@@ -745,6 +745,8 @@ void Host_Main(void)
                if (host_framerate.value < 0.00001 && host_framerate.value != 0)
                        Cvar_SetValue("host_framerate", 0);
 
+               TaskQueue_Frame(false);
+
                // keep the random time dependent, but not when playing demos/benchmarking
                if(!*sv_random_seed.string && !cls.demoplayback)
                        rand();
index cf2fed0cf3df6e7a7bd454697871118ce5a2cb5b..79609fb13d2cc1c2667f969d63ad5459828c4dc2 100644 (file)
@@ -57,6 +57,7 @@ cvar_t mod_q3shader_default_offsetmapping_scale = {CVAR_SAVE, "mod_q3shader_defa
 cvar_t mod_q3shader_default_offsetmapping_bias = {CVAR_SAVE, "mod_q3shader_default_offsetmapping_bias", "0", "default bias used for offsetmapping"};
 cvar_t mod_q3shader_default_polygonfactor = {0, "mod_q3shader_default_polygonfactor", "0", "biases depth values of 'polygonoffset' shaders to prevent z-fighting artifacts"};
 cvar_t mod_q3shader_default_polygonoffset = {0, "mod_q3shader_default_polygonoffset", "-2", "biases depth values of 'polygonoffset' shaders to prevent z-fighting artifacts"};
+cvar_t mod_q3shader_default_refractive_index = { 0, "mod_q3shader_default_refractive_index", "1.33", "angle of refraction specified as n to apply when a photon is refracted, example values are: 1.0003 = air, water = 1.333, crown glass = 1.517, flint glass = 1.655, diamond = 2.417" };
 cvar_t mod_q3shader_force_addalpha = {0, "mod_q3shader_force_addalpha", "0", "treat GL_ONE GL_ONE (or add) blendfunc as GL_SRC_ALPHA GL_ONE for compatibility with older DarkPlaces releases"};
 cvar_t mod_q3shader_force_terrain_alphaflag = {0, "mod_q3shader_force_terrain_alphaflag", "0", "for multilayered terrain shaders force TEXF_ALPHA flag on both layers"};
 
@@ -105,6 +106,7 @@ void Mod_BrushInit(void)
        Cvar_RegisterVariable(&mod_q3shader_default_offsetmapping_bias);
        Cvar_RegisterVariable(&mod_q3shader_default_polygonfactor);
        Cvar_RegisterVariable(&mod_q3shader_default_polygonoffset);
+       Cvar_RegisterVariable(&mod_q3shader_default_refractive_index);
        Cvar_RegisterVariable(&mod_q3shader_force_addalpha);
        Cvar_RegisterVariable(&mod_q3shader_force_terrain_alphaflag);
        Cvar_RegisterVariable(&mod_q1bsp_polygoncollisions);
@@ -1745,6 +1747,7 @@ static void Mod_Q1BSP_LoadTextures(sizebuf_t *sb)
                // clear water settings
                tx->reflectmin = 0;
                tx->reflectmax = 1;
+               tx->refractive_index = mod_q3shader_default_refractive_index.value;
                tx->refractfactor = 1;
                Vector4Set(tx->refractcolor4f, 1, 1, 1, 1);
                tx->reflectfactor = 1;
index ae6cc3b78aff238088b8631a9eb59095fbba16c3..931fc660281e38cf34047c32f4b670744e3b8841 100644 (file)
@@ -1431,14 +1431,6 @@ static void Q3Shader_AddToHash (q3shaderinfo_t* shader)
        memcpy (&entry->shader, shader, sizeof (q3shaderinfo_t));
 }
 
-extern cvar_t mod_noshader_default_offsetmapping;
-extern cvar_t mod_q3shader_default_offsetmapping;
-extern cvar_t mod_q3shader_default_offsetmapping_scale;
-extern cvar_t mod_q3shader_default_offsetmapping_bias;
-extern cvar_t mod_q3shader_default_polygonoffset;
-extern cvar_t mod_q3shader_default_polygonfactor;
-extern cvar_t mod_q3shader_force_addalpha;
-extern cvar_t mod_q3shader_force_terrain_alphaflag;
 void Mod_LoadQ3Shaders(void)
 {
        int j;
@@ -2473,6 +2465,7 @@ nothing                GL_ZERO GL_ONE
                texture->specularscalemod = shader->specularscalemod;
                texture->specularpowermod = shader->specularpowermod;
                texture->rtlightambient = shader->rtlightambient;
+               texture->refractive_index = mod_q3shader_default_refractive_index.value;
                if (shader->dpreflectcube[0])
                        texture->reflectcubetexture = R_GetCubemap(shader->dpreflectcube);
 
index cf64465f42b4f0d8b12248bb6d622115e4e28d4b..b21bbf4b65eaf7ac692e40802d9e9ce160e49a19 100644 (file)
@@ -632,6 +632,7 @@ typedef struct texture_s
        vec4_t reflectcolor4f; // color tint of reflection (including alpha factor)
        float r_water_wateralpha; // additional wateralpha to apply when r_water is active
        float r_water_waterscroll[2]; // scale and speed
+       float refractive_index; // used by r_shadow_bouncegrid for bending photons for refracted light
        int camera_entity; // entity number for use by cameras
 
        // offsetmapping
@@ -1096,6 +1097,16 @@ extern unsigned char *mod_base;
 // texture fullbrights
 extern cvar_t r_fullbrights;
 
+extern cvar_t mod_noshader_default_offsetmapping;
+extern cvar_t mod_q3shader_default_offsetmapping;
+extern cvar_t mod_q3shader_default_offsetmapping_scale;
+extern cvar_t mod_q3shader_default_offsetmapping_bias;
+extern cvar_t mod_q3shader_default_polygonoffset;
+extern cvar_t mod_q3shader_default_polygonfactor;
+extern cvar_t mod_q3shader_default_refractive_index;
+extern cvar_t mod_q3shader_force_addalpha;
+extern cvar_t mod_q3shader_force_terrain_alphaflag;
+
 void Mod_Init (void);
 void Mod_Reload (void);
 dp_model_t *Mod_LoadModel(dp_model_t *mod, qboolean crash, qboolean checkdisk);
index c293ac78e7e81b75ec93cb259e8fa6dee212791f..9bd1bad43a6c991a1a07351c68905e435f7cf99f 100644 (file)
@@ -195,7 +195,7 @@ cvar_t r_shadow_culllights_trace_delay = {CVAR_SAVE, "r_shadow_culllights_trace_
 cvar_t r_shadow_bouncegrid = {CVAR_SAVE, "r_shadow_bouncegrid", "0", "perform particle tracing for indirect lighting (Global Illumination / radiosity) using a 3D texture covering the scene, only active on levels with realtime lights active (r_shadow_realtime_world is usually required for these)"};
 cvar_t r_shadow_bouncegrid_blur = {CVAR_SAVE, "r_shadow_bouncegrid_blur", "0", "apply a 1-radius blur on bouncegrid to denoise it and deal with boundary issues with surfaces"};
 cvar_t r_shadow_bouncegrid_dynamic_bounceminimumintensity = { CVAR_SAVE, "r_shadow_bouncegrid_dynamic_bounceminimumintensity", "0.05", "stop bouncing once intensity drops below this fraction of the original particle color" };
-cvar_t r_shadow_bouncegrid_dynamic_culllightpaths = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_culllightpaths", "1", "skip accumulating light in the bouncegrid texture where the light paths are out of view (dynamic mode only)"};
+cvar_t r_shadow_bouncegrid_dynamic_culllightpaths = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_culllightpaths", "0", "skip accumulating light in the bouncegrid texture where the light paths are out of view (dynamic mode only)"};
 cvar_t r_shadow_bouncegrid_dynamic_directionalshading = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_directionalshading", "1", "use diffuse shading rather than ambient, 3D texture becomes 8x as many pixels to hold the additional data"};
 cvar_t r_shadow_bouncegrid_dynamic_dlightparticlemultiplier = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_dlightparticlemultiplier", "1", "if set to a high value like 16 this can make dlights look great, but 0 is recommended for performance reasons"};
 cvar_t r_shadow_bouncegrid_dynamic_hitmodels = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_hitmodels", "0", "enables hitting character model geometry (SLOW)"};
@@ -210,10 +210,10 @@ cvar_t r_shadow_bouncegrid_dynamic_y = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_
 cvar_t r_shadow_bouncegrid_dynamic_z = {CVAR_SAVE, "r_shadow_bouncegrid_dynamic_z", "32", "maximum texture size of bouncegrid on Z axis"};
 cvar_t r_shadow_bouncegrid_floatcolors = {CVAR_SAVE, "r_shadow_bouncegrid_floatcolors", "1", "upload texture as RGBA16F (or RGBA32F when set to 2) rather than RGBA8 format - this gives more dynamic range and accuracy"};
 cvar_t r_shadow_bouncegrid_includedirectlighting = {CVAR_SAVE, "r_shadow_bouncegrid_includedirectlighting", "0", "allows direct lighting to be recorded, not just indirect (gives an effect somewhat like r_shadow_realtime_world_lightmaps)"};
-cvar_t r_shadow_bouncegrid_intensity = {CVAR_SAVE, "r_shadow_bouncegrid_intensity", "1", "overall brightness of bouncegrid texture"};
+cvar_t r_shadow_bouncegrid_intensity = {CVAR_SAVE, "r_shadow_bouncegrid_intensity", "4", "overall brightness of bouncegrid texture"};
 cvar_t r_shadow_bouncegrid_lightpathsize = {CVAR_SAVE, "r_shadow_bouncegrid_lightpathsize", "64", "radius (in game units) of the light path for accumulation of light in the bouncegrid texture"};
 cvar_t r_shadow_bouncegrid_normalizevectors = { CVAR_SAVE, "r_shadow_bouncegrid_normalizevectors", "1", "normalize random vectors (otherwise their length can vary, which dims the lighting further from the light)" };
-cvar_t r_shadow_bouncegrid_particlebounceintensity = {CVAR_SAVE, "r_shadow_bouncegrid_particlebounceintensity", "2", "amount of energy carried over after each bounce, this is a multiplier of texture color and the result is clamped to 1 or less, to prevent adding energy on each bounce"};
+cvar_t r_shadow_bouncegrid_particlebounceintensity = {CVAR_SAVE, "r_shadow_bouncegrid_particlebounceintensity", "4", "amount of energy carried over after each bounce, this is a multiplier of texture color and the result is clamped to 1 or less, to prevent adding energy on each bounce"};
 cvar_t r_shadow_bouncegrid_particleintensity = {CVAR_SAVE, "r_shadow_bouncegrid_particleintensity", "1", "brightness of particles contributing to bouncegrid texture"};
 cvar_t r_shadow_bouncegrid_rng_seed = { CVAR_SAVE, "r_shadow_bouncegrid_rng_seed", "0", "0+ = use this number as RNG seed, -1 = use time instead for disco-like craziness in dynamic mode" };
 cvar_t r_shadow_bouncegrid_rng_type = { CVAR_SAVE, "r_shadow_bouncegrid_rng_type", "0", "0 = Lehmer 128bit RNG (slow but high quality), 1 = lhcheeserand 32bit RNG (quick)" };
@@ -223,8 +223,10 @@ cvar_t r_shadow_bouncegrid_static_directionalshading = {CVAR_SAVE, "r_shadow_bou
 cvar_t r_shadow_bouncegrid_static_lightradiusscale = {CVAR_SAVE, "r_shadow_bouncegrid_static_lightradiusscale", "5", "particles stop at this fraction of light radius (can be more than 1) when in static mode"};
 cvar_t r_shadow_bouncegrid_static_maxbounce = {CVAR_SAVE, "r_shadow_bouncegrid_static_maxbounce", "5", "maximum number of bounces for a particle (minimum is 0) in static mode"};
 cvar_t r_shadow_bouncegrid_static_maxphotons = {CVAR_SAVE, "r_shadow_bouncegrid_static_maxphotons", "250000", "upper bound on photons in static mode"};
-cvar_t r_shadow_bouncegrid_static_quality = { CVAR_SAVE, "r_shadow_bouncegrid_static_quality", "16", "amount of photons that should be fired (this is multiplied by spacing ^ 2 to make it adaptive with spacing changes)" };
+cvar_t r_shadow_bouncegrid_static_quality = {CVAR_SAVE, "r_shadow_bouncegrid_static_quality", "16", "amount of photons that should be fired (this is multiplied by spacing ^ 2 to make it adaptive with spacing changes)"};
 cvar_t r_shadow_bouncegrid_static_spacing = {CVAR_SAVE, "r_shadow_bouncegrid_static_spacing", "64", "unit size of bouncegrid pixel when in static mode"};
+cvar_t r_shadow_bouncegrid_subsamples = {CVAR_SAVE, "r_shadow_bouncegrid_subsamples", "1", "when generating the texture, sample this many points along each dimension (multisampling uses more compute but not more memory bandwidth)"};
+cvar_t r_shadow_bouncegrid_threaded = {CVAR_SAVE, "r_shadow_bouncegrid_threaded", "1", "enables use of taskqueue_maxthreads to perform the traces and slice rendering of bouncegrid"};
 cvar_t r_coronas = {CVAR_SAVE, "r_coronas", "0", "brightness of corona flare effects around certain lights, 0 disables corona effects"};
 cvar_t r_coronas_occlusionsizescale = {CVAR_SAVE, "r_coronas_occlusionsizescale", "0.1", "size of light source for corona occlusion checksum the proportion of hidden pixels controls corona intensity"};
 cvar_t r_coronas_occlusionquery = {CVAR_SAVE, "r_coronas_occlusionquery", "0", "fades coronas according to visibility"};
@@ -491,8 +493,9 @@ static void r_shadow_shutdown(void)
        if (r_shadow_bouncegrid_state.blurpixels[1]) Mem_Free(r_shadow_bouncegrid_state.blurpixels[1]); r_shadow_bouncegrid_state.blurpixels[1] = NULL;
        if (r_shadow_bouncegrid_state.u8pixels) Mem_Free(r_shadow_bouncegrid_state.u8pixels); r_shadow_bouncegrid_state.u8pixels = NULL;
        if (r_shadow_bouncegrid_state.fp16pixels) Mem_Free(r_shadow_bouncegrid_state.fp16pixels); r_shadow_bouncegrid_state.fp16pixels = NULL;
-       if (r_shadow_bouncegrid_state.splatpaths) Mem_Free(r_shadow_bouncegrid_state.splatpaths); r_shadow_bouncegrid_state.splatpaths = NULL;
-       r_shadow_bouncegrid_state.maxsplatpaths = 0;
+       if (r_shadow_bouncegrid_state.photons) Mem_Free(r_shadow_bouncegrid_state.photons); r_shadow_bouncegrid_state.photons = NULL;
+       if (r_shadow_bouncegrid_state.photons_tasks) Mem_Free(r_shadow_bouncegrid_state.photons_tasks); r_shadow_bouncegrid_state.photons_tasks = NULL;
+       if (r_shadow_bouncegrid_state.slices_tasks) Mem_Free(r_shadow_bouncegrid_state.slices_tasks); r_shadow_bouncegrid_state.slices_tasks = NULL;
        memset(&r_shadow_bouncegrid_state, 0, sizeof(r_shadow_bouncegrid_state));
        r_shadow_attenuationgradienttexture = NULL;
        R_FreeTexturePool(&r_shadow_texturepool);
@@ -564,9 +567,9 @@ static void r_shadow_newmap(void)
        if (r_shadow_bouncegrid_state.blurpixels[1]) { Mem_Free(r_shadow_bouncegrid_state.blurpixels[1]); r_shadow_bouncegrid_state.blurpixels[1] = NULL; }
        if (r_shadow_bouncegrid_state.u8pixels) { Mem_Free(r_shadow_bouncegrid_state.u8pixels); r_shadow_bouncegrid_state.u8pixels = NULL; }
        if (r_shadow_bouncegrid_state.fp16pixels) { Mem_Free(r_shadow_bouncegrid_state.fp16pixels); r_shadow_bouncegrid_state.fp16pixels = NULL; }
-       if (r_shadow_bouncegrid_state.splatpaths) { Mem_Free(r_shadow_bouncegrid_state.splatpaths); r_shadow_bouncegrid_state.splatpaths = NULL; }
-
-       r_shadow_bouncegrid_state.maxsplatpaths = 0;
+       if (r_shadow_bouncegrid_state.photons) Mem_Free(r_shadow_bouncegrid_state.photons); r_shadow_bouncegrid_state.photons = NULL;
+       if (r_shadow_bouncegrid_state.photons_tasks) Mem_Free(r_shadow_bouncegrid_state.photons_tasks); r_shadow_bouncegrid_state.photons_tasks = NULL;
+       if (r_shadow_bouncegrid_state.slices_tasks) Mem_Free(r_shadow_bouncegrid_state.slices_tasks); r_shadow_bouncegrid_state.slices_tasks = NULL;
 
        if (r_shadow_bouncegrid_state.texture)    { R_FreeTexture(r_shadow_bouncegrid_state.texture);r_shadow_bouncegrid_state.texture = NULL; }
        if (r_shadow_lightcorona)                 { R_SkinFrame_MarkUsed(r_shadow_lightcorona); }
@@ -672,6 +675,8 @@ void R_Shadow_Init(void)
        Cvar_RegisterVariable(&r_shadow_bouncegrid_static_maxphotons);
        Cvar_RegisterVariable(&r_shadow_bouncegrid_static_quality);
        Cvar_RegisterVariable(&r_shadow_bouncegrid_static_spacing);
+       Cvar_RegisterVariable(&r_shadow_bouncegrid_subsamples);
+       Cvar_RegisterVariable(&r_shadow_bouncegrid_threaded);
        Cvar_RegisterVariable(&r_coronas);
        Cvar_RegisterVariable(&r_coronas_occlusionsizescale);
        Cvar_RegisterVariable(&r_coronas_occlusionquery);
@@ -1654,94 +1659,6 @@ void R_Shadow_RenderMode_DrawDeferredLight(qboolean shadowmapping)
        R_Mesh_Draw(0, 8, 0, 12, NULL, NULL, 0, bboxelements, NULL, 0);
 }
 
-#define MAXBOUNCEGRIDSPLATSIZE 7
-#define MAXBOUNCEGRIDSPLATSIZE1 (MAXBOUNCEGRIDSPLATSIZE+1)
-
-// these are temporary data per-frame, to be cache friendly the texture is
-// generated in slices (on Z), rendering each slice one after another and each
-// row in the slice one after another to be more cache friendly than randomly
-// seeking around a large memory space.  Each slice keeps track of a linked list
-// of splat paths that are relevant to it.  Also these are canonically flipped
-// to ensure that start[2] <= end[2].
-typedef struct r_shadow_bouncegrid_splatpath_s
-{
-       int nextpathonslice;
-       int slicerange[2];
-       vec3_t slicecenter;
-       vec3_t start;
-       vec3_t end;
-       vec3_t splatcolor;
-       vec3_t splatdir;
-       vec_t splatintensity;
-       vec_t splatsize_current;
-       vec_t splatsize_perstep;
-}
-r_shadow_bouncegrid_splatpath_t;
-
-static void R_Shadow_BounceGrid_AddSplatPath(vec3_t originalstart, vec3_t originalend, vec3_t color, vec_t distancetraveled)
-{
-       vec3_t start;
-       vec3_t end;
-       vec3_t originaldir;
-       r_shadow_bouncegrid_splatpath_t *path;
-
-       // cull paths that fail R_CullBox in dynamic mode
-       if (!r_shadow_bouncegrid_state.settings.staticmode
-        && r_shadow_bouncegrid_dynamic_culllightpaths.integer)
-       {
-               vec3_t cullmins, cullmaxs;
-               cullmins[0] = min(originalstart[0], originalend[0]) - r_shadow_bouncegrid_state.settings.spacing[0] - r_shadow_bouncegrid_state.settings.lightpathsize;
-               cullmins[1] = min(originalstart[1], originalend[1]) - r_shadow_bouncegrid_state.settings.spacing[1] - r_shadow_bouncegrid_state.settings.lightpathsize;
-               cullmins[2] = min(originalstart[2], originalend[2]) - r_shadow_bouncegrid_state.settings.spacing[2] - r_shadow_bouncegrid_state.settings.lightpathsize;
-               cullmaxs[0] = max(originalstart[0], originalend[0]) + r_shadow_bouncegrid_state.settings.spacing[0] + r_shadow_bouncegrid_state.settings.lightpathsize;
-               cullmaxs[1] = max(originalstart[1], originalend[1]) + r_shadow_bouncegrid_state.settings.spacing[1] + r_shadow_bouncegrid_state.settings.lightpathsize;
-               cullmaxs[2] = max(originalstart[2], originalend[2]) + r_shadow_bouncegrid_state.settings.spacing[2] + r_shadow_bouncegrid_state.settings.lightpathsize;
-               if (R_CullBox(cullmins, cullmaxs))
-                       return;
-       }
-
-       // if the light path is going upward, reverse it - we always draw down.
-       if (originalend[2] < originalstart[2])
-       {
-               VectorCopy(originalend, start);
-               VectorCopy(originalstart, end);
-       }
-       else
-       {
-               VectorCopy(originalstart, start);
-               VectorCopy(originalend, end);
-       }
-
-       // transform to texture pixels
-       start[0] = (start[0] - r_shadow_bouncegrid_state.mins[0]);
-       start[1] = (start[1] - r_shadow_bouncegrid_state.mins[1]);
-       start[2] = (start[2] - r_shadow_bouncegrid_state.mins[2]);
-       end[0] = (end[0] - r_shadow_bouncegrid_state.mins[0]);
-       end[1] = (end[1] - r_shadow_bouncegrid_state.mins[1]);
-       end[2] = (end[2] - r_shadow_bouncegrid_state.mins[2]);
-
-       // check if we need to grow the splatpaths array
-       if (r_shadow_bouncegrid_state.maxsplatpaths <= r_shadow_bouncegrid_state.numsplatpaths)
-       {
-               // double the limit, this will persist from frame to frame so we don't
-               // make the same mistake each time
-               r_shadow_bouncegrid_state.maxsplatpaths *= 2;
-               if (r_shadow_bouncegrid_state.maxsplatpaths < 16384)
-                       r_shadow_bouncegrid_state.maxsplatpaths = 16384;
-               r_shadow_bouncegrid_state.splatpaths = (r_shadow_bouncegrid_splatpath_t *)Mem_Realloc(r_main_mempool, r_shadow_bouncegrid_state.splatpaths, sizeof(r_shadow_bouncegrid_splatpath_t) * r_shadow_bouncegrid_state.maxsplatpaths);
-       }
-
-       VectorSubtract(originalstart, originalend, originaldir);
-       VectorNormalize(originaldir);
-
-       path = r_shadow_bouncegrid_state.splatpaths + r_shadow_bouncegrid_state.numsplatpaths++;
-       VectorCopy(start, path->start);
-       VectorCopy(end, path->end);
-       VectorCopy(color, path->splatcolor);
-       VectorCopy(originaldir, path->splatdir);
-       path->splatintensity = VectorLength(color);
-}
-
 static qboolean R_Shadow_BounceGrid_CheckEnable(int flag)
 {
        qboolean enable = r_shadow_bouncegrid_state.capable && r_shadow_bouncegrid.integer != 0 && r_refdef.scene.worldmodel;
@@ -1809,6 +1726,7 @@ static void R_Shadow_BounceGrid_GenerateSettings(r_shadow_bouncegrid_settings_t
        settings->bounceminimumintensity2       = bounceminimumintensity * bounceminimumintensity;
        settings->bounceminimumintensity2       = bounceminimumintensity * bounceminimumintensity;
        settings->normalizevectors              = r_shadow_bouncegrid_normalizevectors.integer != 0;
+       settings->subsamples                    = bound(1, r_shadow_bouncegrid_subsamples.integer, 4);
 
        // bound the values for sanity
        settings->maxphotons = bound(1, settings->maxphotons, 25000000);
@@ -1959,9 +1877,10 @@ static void R_Shadow_BounceGrid_UpdateSpacing(void)
                if (r_shadow_bouncegrid_state.blurpixels[1]) { Mem_Free(r_shadow_bouncegrid_state.blurpixels[1]); r_shadow_bouncegrid_state.blurpixels[1] = NULL; }
                if (r_shadow_bouncegrid_state.u8pixels) { Mem_Free(r_shadow_bouncegrid_state.u8pixels); r_shadow_bouncegrid_state.u8pixels = NULL; }
                if (r_shadow_bouncegrid_state.fp16pixels) { Mem_Free(r_shadow_bouncegrid_state.fp16pixels); r_shadow_bouncegrid_state.fp16pixels = NULL; }
-               if (r_shadow_bouncegrid_state.splatpaths) { Mem_Free(r_shadow_bouncegrid_state.splatpaths); r_shadow_bouncegrid_state.splatpaths = NULL; }
+               if (r_shadow_bouncegrid_state.photons) { Mem_Free(r_shadow_bouncegrid_state.photons); r_shadow_bouncegrid_state.photons = NULL; }
+               if (r_shadow_bouncegrid_state.photons_tasks) { Mem_Free(r_shadow_bouncegrid_state.photons_tasks); r_shadow_bouncegrid_state.photons_tasks = NULL; }
+               if (r_shadow_bouncegrid_state.slices_tasks) Mem_Free(r_shadow_bouncegrid_state.slices_tasks); r_shadow_bouncegrid_state.slices_tasks = NULL;
 
-               r_shadow_bouncegrid_state.maxsplatpaths = 0;
                r_shadow_bouncegrid_state.numpixels = numpixels;
        }
 
@@ -1977,13 +1896,37 @@ static void R_Shadow_BounceGrid_UpdateSpacing(void)
        Matrix4x4_FromArrayFloatD3D(&r_shadow_bouncegrid_state.matrix, m);
 }
 
+static float R_Shadow_BounceGrid_RefractiveIndexAtPoint(vec3_t point)
+{
+       // check material at shadoworigin to see what the initial refractive index should be
+       int hitsupercontentsmask = SUPERCONTENTS_SOLID | SUPERCONTENTS_LIQUIDSMASK;
+       int skipsupercontentsmask = 0;
+       int skipmaterialflagsmask = MATERIALFLAG_CUSTOMBLEND;
+       trace_t trace = CL_TracePoint(point, r_shadow_bouncegrid_state.settings.staticmode ? MOVE_WORLDONLY : (r_shadow_bouncegrid_state.settings.hitmodels ? MOVE_HITMODEL : MOVE_NOMONSTERS), NULL, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask, true, false, NULL, true);
+       if (trace.starttexture && (trace.starttexture->currentmaterialflags & (MATERIALFLAG_REFRACTION | MATERIALFLAG_WATERSHADER)))
+               return trace.starttexture->refractive_index;
+       else if (trace.startsupercontents & SUPERCONTENTS_LIQUIDSMASK)
+               return 1.333f; // water
+       else
+               return 1.0003f; // air
+}
+
 // enumerate world rtlights and sum the overall amount of light in the world,
 // from that we can calculate a scaling factor to fairly distribute photons
 // to all the lights
 //
 // this modifies rtlight->photoncolor and rtlight->photons
-static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *settings, unsigned int range, unsigned int range1, unsigned int range2, int flag)
+static void R_Shadow_BounceGrid_AssignPhotons_Task(taskqueue_task_t *t)
 {
+       // get the range of light numbers we'll be looping over:
+       // range = static lights
+       // range1 = dynamic lights (optional)
+       // range2 = range + range1
+       unsigned int range = (unsigned int)Mem_ExpandableArray_IndexRange(&r_shadow_worldlightsarray); // checked
+       unsigned int range1 = r_shadow_bouncegrid_state.settings.staticmode ? 0 : r_refdef.scene.numlights;
+       unsigned int range2 = range + range1;
+       int flag = r_refdef.scene.rtworld ? LIGHTFLAG_REALTIMEMODE : LIGHTFLAG_NORMALMODE;
+
        float normalphotonscaling;
        float photonscaling;
        float photonintensity;
@@ -1997,7 +1940,17 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
        unsigned int lightindex;
        dlight_t *light;
        rtlight_t *rtlight;
-       normalphotonscaling = 1.0f / max(0.0000001f, settings->energyperphoton);
+       int shootparticles;
+       int shotparticles;
+       float bounceminimumintensity2;
+       float startrefractiveindex;
+       unsigned int seed;
+       randomseed_t randomseed;
+       vec3_t baseshotcolor;
+
+       t->started = 1;
+
+       normalphotonscaling = 1.0f / max(0.0000001f, r_shadow_bouncegrid_state.settings.energyperphoton);
        for (lightindex = 0;lightindex < range2;lightindex++)
        {
                if (lightindex < range)
@@ -2013,7 +1966,7 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
                        rtlight->bouncegrid_effectiveradius = 0;
                        if (!(light->flags & flag))
                                continue;
-                       if (settings->staticmode)
+                       if (r_shadow_bouncegrid_state.settings.staticmode)
                        {
                                // when static, we skip styled lights because they tend to change...
                                if (rtlight->style > 0 && r_shadow_bouncegrid.integer != 2)
@@ -2032,7 +1985,7 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
                        rtlight->bouncegrid_effectiveradius = 0;
                }
                // draw only visible lights (major speedup)
-               radius = rtlight->radius * settings->lightradiusscale;
+               radius = rtlight->radius * r_shadow_bouncegrid_state.settings.lightradiusscale;
                cullmins[0] = rtlight->shadoworigin[0] - radius;
                cullmins[1] = rtlight->shadoworigin[1] - radius;
                cullmins[2] = rtlight->shadoworigin[2] - radius;
@@ -2040,7 +1993,7 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
                cullmaxs[1] = rtlight->shadoworigin[1] + radius;
                cullmaxs[2] = rtlight->shadoworigin[2] + radius;
                w = r_shadow_lightintensityscale.value * (rtlight->ambientscale + rtlight->diffusescale + rtlight->specularscale);
-               if (!settings->staticmode)
+               if (!r_shadow_bouncegrid_state.settings.staticmode)
                {
                        // skip if the expanded light box does not touch any visible leafs
                        if (r_refdef.scene.worldmodel
@@ -2082,10 +2035,10 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
                s = rtlight->radius;
                lightintensity = VectorLength(rtlight->color) * (rtlight->ambientscale + rtlight->diffusescale + rtlight->specularscale);
                if (lightindex >= range)
-                       lightintensity *= settings->dlightparticlemultiplier;
+                       lightintensity *= r_shadow_bouncegrid_state.settings.dlightparticlemultiplier;
                rtlight->bouncegrid_photons = lightintensity * s * s * normalphotonscaling;
                photoncount += rtlight->bouncegrid_photons;
-               VectorScale(rtlight->bouncegrid_photoncolor, settings->particleintensity * settings->energyperphoton, rtlight->bouncegrid_photoncolor);
+               VectorScale(rtlight->bouncegrid_photoncolor, r_shadow_bouncegrid_state.settings.particleintensity * r_shadow_bouncegrid_state.settings.energyperphoton, rtlight->bouncegrid_photoncolor);
                // if the lightstyle happens to be off right now, we can skip actually
                // firing the photons, but we did have to count them in the total.
                //if (VectorLength2(rtlight->photoncolor) == 0.0f)
@@ -2097,9 +2050,9 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
        // like doing that in the typical case
        photonscaling = 1.0f;
        photonintensity = 1.0f;
-       if (photoncount > settings->maxphotons)
+       if (photoncount > r_shadow_bouncegrid_state.settings.maxphotons)
        {
-               photonscaling = settings->maxphotons / photoncount;
+               photonscaling = r_shadow_bouncegrid_state.settings.maxphotons / photoncount;
                photonintensity = 1.0f / photonscaling;
        }
 
@@ -2118,155 +2071,257 @@ static void R_Shadow_BounceGrid_AssignPhotons(r_shadow_bouncegrid_settings_t *se
                rtlight->bouncegrid_photons *= photonscaling;
                VectorScale(rtlight->bouncegrid_photoncolor, photonintensity, rtlight->bouncegrid_photoncolor);
        }
-}
 
-static void R_Shadow_BounceGrid_ClearPixels(void)
-{
-       // clear the highpixels array we'll be accumulating into
-       if (r_shadow_bouncegrid_state.blurpixels[0] == NULL)
-               r_shadow_bouncegrid_state.blurpixels[0] = (float *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
-       if (r_shadow_bouncegrid_state.settings.blur && r_shadow_bouncegrid_state.blurpixels[1] == NULL)
-               r_shadow_bouncegrid_state.blurpixels[1] = (float *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
-       r_shadow_bouncegrid_state.highpixels_index = 0;
-       r_shadow_bouncegrid_state.highpixels = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
-       memset(r_shadow_bouncegrid_state.highpixels, 0, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
+       // compute a seed for the unstable random modes
+       Math_RandomSeed_FromInts(&randomseed, 0, 0, 0, realtime * 1000.0);
+       seed = realtime * 1000.0;
+
+       for (lightindex = 0; lightindex < range2; lightindex++)
+       {
+               if (lightindex < range)
+               {
+                       light = (dlight_t *)Mem_ExpandableArray_RecordAtIndex(&r_shadow_worldlightsarray, lightindex);
+                       if (!light)
+                               continue;
+                       rtlight = &light->rtlight;
+               }
+               else
+                       rtlight = r_refdef.scene.lights[lightindex - range];
+               // note that this code used to keep track of residual photons and
+               // distribute them evenly to achieve exactly a desired photon count,
+               // but that caused unwanted flickering in dynamic mode
+               shootparticles = (int)floor(rtlight->bouncegrid_photons);
+               // skip if we won't be shooting any photons
+               if (!shootparticles)
+                       continue;
+               radius = rtlight->radius * r_shadow_bouncegrid_state.settings.lightradiusscale;
+               //s = settings.particleintensity / shootparticles;
+               //VectorScale(rtlight->bouncegrid_photoncolor, s, baseshotcolor);
+               VectorCopy(rtlight->bouncegrid_photoncolor, baseshotcolor);
+               if (VectorLength2(baseshotcolor) <= 0.0f)
+                       continue;
+               r_refdef.stats[r_stat_bouncegrid_lights]++;
+               r_refdef.stats[r_stat_bouncegrid_particles] += shootparticles;
+               // we stop caring about bounces once the brightness goes below this fraction of the original intensity
+               bounceminimumintensity2 = VectorLength(baseshotcolor) * r_shadow_bouncegrid_state.settings.bounceminimumintensity2;
+
+               // check material at shadoworigin to see what the initial refractive index should be
+               startrefractiveindex = R_Shadow_BounceGrid_RefractiveIndexAtPoint(rtlight->shadoworigin);
+
+               // for seeded random we start the RNG with the position of the light
+               if (r_shadow_bouncegrid_state.settings.rng_seed >= 0)
+               {
+                       union
+                       {
+                               unsigned int i[4];
+                               float f[4];
+                       }
+                       u;
+                       u.f[0] = rtlight->shadoworigin[0];
+                       u.f[1] = rtlight->shadoworigin[1];
+                       u.f[2] = rtlight->shadoworigin[2];
+                       u.f[3] = 1;
+                       switch (r_shadow_bouncegrid_state.settings.rng_type)
+                       {
+                       default:
+                       case 0:
+                               // we have to shift the seed provided by the user because the result must be odd
+                               Math_RandomSeed_FromInts(&randomseed, u.i[0], u.i[1], u.i[2], u.i[3] ^ (r_shadow_bouncegrid_state.settings.rng_seed << 1));
+                               break;
+                       case 1:
+                               seed = u.i[0] ^ u.i[1] ^ u.i[2] ^ u.i[3] ^ r_shadow_bouncegrid_state.settings.rng_seed;
+                               break;
+                       }
+               }
+
+               for (shotparticles = 0; shotparticles < shootparticles && r_shadow_bouncegrid_state.numphotons < r_shadow_bouncegrid_state.settings.maxphotons; shotparticles++)
+               {
+                       r_shadow_bouncegrid_photon_t *p = r_shadow_bouncegrid_state.photons + r_shadow_bouncegrid_state.numphotons++;
+                       VectorCopy(baseshotcolor, p->color);
+                       VectorCopy(rtlight->shadoworigin, p->start);
+                       switch (r_shadow_bouncegrid_state.settings.rng_type)
+                       {
+                       default:
+                       case 0:
+                               // figure out a random direction for the initial photon to go
+                               VectorLehmerRandom(&randomseed, p->end);
+                               break;
+                       case 1:
+                               // figure out a random direction for the initial photon to go
+                               VectorCheeseRandom(seed, p->end);
+                               break;
+                       }
+
+                       // we want a uniform distribution spherically, not merely within the sphere
+                       if (r_shadow_bouncegrid_state.settings.normalizevectors)
+                               VectorNormalize(p->end);
+
+                       VectorMA(p->start, radius, p->end, p->end);
+                       p->bounceminimumintensity2 = bounceminimumintensity2;
+                       p->startrefractiveindex = startrefractiveindex;
+                       p->numpaths = 0;
+               }
+       }
+
+       t->done = 1;
 }
 
-static void R_Shadow_BounceGrid_PerformSplats(void)
+static void R_Shadow_BounceGrid_Slice(int zi)
 {
-       r_shadow_bouncegrid_splatpath_t *splatpaths = r_shadow_bouncegrid_state.splatpaths;
-       r_shadow_bouncegrid_splatpath_t *splatpath;
        float *highpixels = r_shadow_bouncegrid_state.highpixels;
-       int numsplatpaths = r_shadow_bouncegrid_state.numsplatpaths;
-       int pathindex;
-       int xi, yi, zi; // pixel increments
-       float xf, yf, zf; // pixel centers
-       float splatcolor[32] = { 0 };
+       int xi, yi; // pixel increments
+       float color[32] = { 0 };
        float radius = r_shadow_bouncegrid_state.settings.lightpathsize;
        float iradius = 1.0f / radius;
+       int slicemins[3], slicemaxs[3];
        int resolution[3];
        int pixelsperband = r_shadow_bouncegrid_state.pixelsperband;
        int pixelbands = r_shadow_bouncegrid_state.pixelbands;
+       int photonindex;
+       int samples = r_shadow_bouncegrid_state.settings.subsamples;
+       float isamples = 1.0f / samples;
+       float samplescolorscale = isamples * isamples * isamples;
 
        // we use these a lot, so get a local copy
        VectorCopy(r_shadow_bouncegrid_state.resolution, resolution);
 
-       for (pathindex = 0, splatpath = splatpaths; pathindex < numsplatpaths; pathindex++, splatpath++)
+       for (photonindex = 0; photonindex < r_shadow_bouncegrid_state.numphotons; photonindex++)
        {
-               splatpath->slicerange[0] = (int)floor((min(splatpath->start[2], splatpath->end[2]) - radius) * r_shadow_bouncegrid_state.ispacing[2] - 0.5f);
-               splatpath->slicerange[1] = (int)floor((max(splatpath->start[2], splatpath->end[2]) + radius) * r_shadow_bouncegrid_state.ispacing[2] - 0.5f + 1.0f);
-       }
-
-       // we keep a 1 pixel border around the whole texture to make sure that GL_CLAMP_TO_EDGE filtering doesn't create streaks if the texture is smaller than the visible scene (instead it goes black out there, which isn't ideal either)
-       for (zi = 1, zf = (zi + 0.5f) * r_shadow_bouncegrid_state.spacing[2]; zi < resolution[2] - 1; zi++, zf += r_shadow_bouncegrid_state.spacing[2])
-       {
-               int slicefirstpathindex = -1;
-               for (pathindex = 0, splatpath = splatpaths; pathindex < numsplatpaths; pathindex++, splatpath++)
+               r_shadow_bouncegrid_photon_t *photon = r_shadow_bouncegrid_state.photons + photonindex;
+               int pathindex;
+               for (pathindex = 0; pathindex < photon->numpaths; pathindex++)
                {
-                       if (zi >= splatpath->slicerange[0] && zi < splatpath->slicerange[1])
+                       r_shadow_bouncegrid_photon_path_t *path = photon->paths + pathindex;
+                       float pathstart[3], pathend[3], pathmins[3], pathmaxs[3], pathdelta[3], pathdir[3], pathlength2, pathilength;
+
+                       VectorSubtract(path->start, r_shadow_bouncegrid_state.mins, pathstart);
+                       VectorSubtract(path->end, r_shadow_bouncegrid_state.mins, pathend);
+
+                       pathmins[2] = min(pathstart[2], pathend[2]);
+                       slicemins[2] = (int)floor((pathmins[2] - radius) * r_shadow_bouncegrid_state.ispacing[2]);
+                       pathmaxs[2] = max(pathstart[2], pathend[2]);
+                       slicemaxs[2] = (int)floor((pathmaxs[2] + radius) * r_shadow_bouncegrid_state.ispacing[2] + 1);
+
+                       // skip if the path doesn't touch this slice
+                       if (zi < slicemins[2] || zi >= slicemaxs[2])
+                               continue;
+
+                       pathmins[0] = min(pathstart[0], pathend[0]);
+                       slicemins[0] = (int)floor((pathmins[0] - radius) * r_shadow_bouncegrid_state.ispacing[0]);
+                       slicemins[0] = max(slicemins[0], 1);
+                       pathmaxs[0] = max(pathstart[0], pathend[0]);
+                       slicemaxs[0] = (int)floor((pathmaxs[0] + radius) * r_shadow_bouncegrid_state.ispacing[0]);
+                       slicemaxs[0] = min(slicemaxs[0], resolution[0] - 1);
+
+                       pathmins[1] = min(pathstart[1], pathend[1]);
+                       slicemins[1] = (int)floor((pathmins[1] - radius) * r_shadow_bouncegrid_state.ispacing[1] + 1);
+                       slicemins[1] = max(slicemins[1], 1);
+                       pathmaxs[1] = max(pathstart[1], pathend[1]);
+                       slicemaxs[1] = (int)floor((pathmaxs[1] + radius) * r_shadow_bouncegrid_state.ispacing[1] + 1);
+                       slicemaxs[1] = min(slicemaxs[1], resolution[1] - 1);
+
+                       // skip if the path is out of bounds on X or Y
+                       if (slicemins[0] >= slicemaxs[0] || slicemins[1] >= slicemaxs[1])
+                               continue;
+
+                       // calculate second order spherical harmonics values (average, slopeX, slopeY, slopeZ)
+                       // accumulate average shotcolor
+                       VectorSubtract(pathend, pathstart, pathdelta);
+                       pathlength2 = VectorLength2(pathdelta);
+                       pathilength = pathlength2 > 0.0f ? 1.0f / sqrt(pathlength2) : 0.0f;
+                       VectorScale(pathdelta, pathilength, pathdir);
+                       // the color is scaled by the number of subsamples
+                       color[0] = path->color[0] * samplescolorscale;
+                       color[1] = path->color[1] * samplescolorscale;
+                       color[2] = path->color[2] * samplescolorscale;
+                       color[3] = 0.0f;
+                       if (pixelbands > 1)
                        {
-                               if (zf < splatpath->start[2])
-                                       VectorCopy(splatpath->start, splatpath->slicecenter);
-                               else if (zf > splatpath->end[2])
-                                       VectorCopy(splatpath->end, splatpath->slicecenter);
-                               else
-                               {
-                                       float lerp = (zf - splatpath->start[2]) / (splatpath->end[2] - splatpath->start[2]);
-                                       splatpath->slicecenter[2] = zf;
-                                       splatpath->slicecenter[0] = splatpath->start[0] + lerp * (splatpath->end[0] - splatpath->start[0]);
-                                       splatpath->slicecenter[1] = splatpath->start[1] + lerp * (splatpath->end[1] - splatpath->start[1]);
-                               }
-                               float distz = (splatpath->slicecenter[2] - zf) * iradius;
-                               if (distz * distz < 1.0f)
-                               {
-                                       splatpath->nextpathonslice = slicefirstpathindex;
-                                       slicefirstpathindex = pathindex;
-                               }
+                               // store bentnormal in case the shader has a use for it,
+                               // bentnormal is an intensity-weighted average of the directions,
+                               // and will be normalized on conversion to texture pixels.
+                               float intensity = VectorLength(color);
+                               color[4] = pathdir[0] * intensity;
+                               color[5] = pathdir[1] * intensity;
+                               color[6] = pathdir[2] * intensity;
+                               color[7] = intensity;
+                               // for each color component (R, G, B) calculate the amount that a
+                               // direction contributes
+                               color[8] = color[0] * max(0.0f, pathdir[0]);
+                               color[9] = color[0] * max(0.0f, pathdir[1]);
+                               color[10] = color[0] * max(0.0f, pathdir[2]);
+                               color[11] = 0.0f;
+                               color[12] = color[1] * max(0.0f, pathdir[0]);
+                               color[13] = color[1] * max(0.0f, pathdir[1]);
+                               color[14] = color[1] * max(0.0f, pathdir[2]);
+                               color[15] = 0.0f;
+                               color[16] = color[2] * max(0.0f, pathdir[0]);
+                               color[17] = color[2] * max(0.0f, pathdir[1]);
+                               color[18] = color[2] * max(0.0f, pathdir[2]);
+                               color[19] = 0.0f;
+                               // and do the same for negative directions
+                               color[20] = color[0] * max(0.0f, -pathdir[0]);
+                               color[21] = color[0] * max(0.0f, -pathdir[1]);
+                               color[22] = color[0] * max(0.0f, -pathdir[2]);
+                               color[23] = 0.0f;
+                               color[24] = color[1] * max(0.0f, -pathdir[0]);
+                               color[25] = color[1] * max(0.0f, -pathdir[1]);
+                               color[26] = color[1] * max(0.0f, -pathdir[2]);
+                               color[27] = 0.0f;
+                               color[28] = color[2] * max(0.0f, -pathdir[0]);
+                               color[29] = color[2] * max(0.0f, -pathdir[1]);
+                               color[30] = color[2] * max(0.0f, -pathdir[2]);
+                               color[31] = 0.0f;
                        }
-               }
-               for (yi = 1, yf = (yi + 0.5f) * r_shadow_bouncegrid_state.spacing[1]; yi < resolution[1] - 1; yi++, yf += r_shadow_bouncegrid_state.spacing[1])
-               {
-                       for (pathindex = slicefirstpathindex; pathindex >= 0; pathindex = splatpaths[pathindex].nextpathonslice)
+
+                       for (yi = slicemins[1]; yi < slicemaxs[1]; yi++)
                        {
-                               splatpath = splatpaths + pathindex;
-                               float disty = (splatpath->slicecenter[1] - yf) * iradius;
-                               float distz = (splatpath->slicecenter[2] - zf) * iradius;
-                               float distyz = disty * disty + distz * distz;
-                               if (distyz < 1.0f)
+                               for (xi = slicemins[0]; xi < slicemaxs[0]; xi++)
                                {
-                                       int xstart = (int)floor((splatpath->slicecenter[0] - radius) * r_shadow_bouncegrid_state.ispacing[0] - 0.5f);
-                                       int xend = (int)floor((splatpath->slicecenter[0] + radius) * r_shadow_bouncegrid_state.ispacing[0] - 0.5f + 1.0f);
-                                       float dir[3];
-                                       xstart = max(1, xstart);
-                                       xend = min(resolution[0] - 2, xend);
-                                       if (xstart >= xend)
-                                               continue;
-                                       // calculate second order spherical harmonics values (average, slopeX, slopeY, slopeZ)
-                                       // accumulate average shotcolor
-                                       VectorCopy(splatpath->splatdir, dir);
-                                       splatcolor[0] = splatpath->splatcolor[0];
-                                       splatcolor[1] = splatpath->splatcolor[1];
-                                       splatcolor[2] = splatpath->splatcolor[2];
-                                       splatcolor[3] = 0.0f;
-                                       if (pixelbands > 1)
-                                       {
-                                               // store bentnormal in case the shader has a use for it,
-                                               // bentnormal is an intensity-weighted average of the directions,
-                                               // and will be normalized on conversion to texture pixels.
-                                               splatcolor[4] = dir[0] * splatpath->splatintensity;
-                                               splatcolor[5] = dir[1] * splatpath->splatintensity;
-                                               splatcolor[6] = dir[2] * splatpath->splatintensity;
-                                               splatcolor[7] = splatpath->splatintensity;
-                                               // for each color component (R, G, B) calculate the amount that a
-                                               // direction contributes
-                                               splatcolor[8] = splatcolor[0] * max(0.0f, dir[0]);
-                                               splatcolor[9] = splatcolor[0] * max(0.0f, dir[1]);
-                                               splatcolor[10] = splatcolor[0] * max(0.0f, dir[2]);
-                                               splatcolor[11] = 0.0f;
-                                               splatcolor[12] = splatcolor[1] * max(0.0f, dir[0]);
-                                               splatcolor[13] = splatcolor[1] * max(0.0f, dir[1]);
-                                               splatcolor[14] = splatcolor[1] * max(0.0f, dir[2]);
-                                               splatcolor[15] = 0.0f;
-                                               splatcolor[16] = splatcolor[2] * max(0.0f, dir[0]);
-                                               splatcolor[17] = splatcolor[2] * max(0.0f, dir[1]);
-                                               splatcolor[18] = splatcolor[2] * max(0.0f, dir[2]);
-                                               splatcolor[19] = 0.0f;
-                                               // and do the same for negative directions
-                                               splatcolor[20] = splatcolor[0] * max(0.0f, -dir[0]);
-                                               splatcolor[21] = splatcolor[0] * max(0.0f, -dir[1]);
-                                               splatcolor[22] = splatcolor[0] * max(0.0f, -dir[2]);
-                                               splatcolor[23] = 0.0f;
-                                               splatcolor[24] = splatcolor[1] * max(0.0f, -dir[0]);
-                                               splatcolor[25] = splatcolor[1] * max(0.0f, -dir[1]);
-                                               splatcolor[26] = splatcolor[1] * max(0.0f, -dir[2]);
-                                               splatcolor[27] = 0.0f;
-                                               splatcolor[28] = splatcolor[2] * max(0.0f, -dir[0]);
-                                               splatcolor[29] = splatcolor[2] * max(0.0f, -dir[1]);
-                                               splatcolor[30] = splatcolor[2] * max(0.0f, -dir[2]);
-                                               splatcolor[31] = 0.0f;
-                                       }
-                                       for (xi = xstart, xf = (xi + 0.5f) * r_shadow_bouncegrid_state.spacing[0]; xi < xend; xi++, xf += r_shadow_bouncegrid_state.spacing[0])
+                                       float sample[3], diff[3], nearest[3], along, distance2;
+                                       float *p = highpixels + 4 * ((zi * resolution[1] + yi) * resolution[0] + xi);
+                                       int xs, ys, zs;
+                                       // loop over the subsamples
+                                       for (zs = 0; zs < samples; zs++)
                                        {
-                                               float distx = (splatpath->slicecenter[0] - xf) * iradius;
-                                               float distxyz = (distx * distx + distyz);
-                                               if (distxyz < 1.0f)
+                                               sample[2] = (zi + (zs + 0.5f) * isamples) * r_shadow_bouncegrid_state.spacing[2];
+                                               for (ys = 0; ys < samples; ys++)
                                                {
-                                                       // contribute some color to this pixel, across all bands
-                                                       float w = 1.0f - 1.0f * sqrt(distxyz);
-                                                       int band;
-                                                       float *p = highpixels + 4 * ((zi * resolution[1] + yi) * resolution[0] + xi);
-                                                       w = min(w, 1.0f);
-                                                       if (pixelbands > 1)
-                                                       {
-                                                               // small optimization for alpha - only splatcolor[7] is non-zero, so skip the rest of the alpha elements.
-                                                               p[pixelsperband * 4 + 3] += splatcolor[7] * w;
-                                                       }
-                                                       for (band = 0; band < pixelbands; band++, p += pixelsperband * 4)
+                                                       sample[1] = (yi + (ys + 0.5f) * isamples) * r_shadow_bouncegrid_state.spacing[1];
+                                                       for (xs = 0; xs < samples; xs++)
                                                        {
-                                                               // add to the pixel color (RGB only - see above)
-                                                               p[0] += splatcolor[band * 4 + 0] * w;
-                                                               p[1] += splatcolor[band * 4 + 1] * w;
-                                                               p[2] += splatcolor[band * 4 + 2] * w;
+                                                               sample[0] = (xi + (xs + 0.5f) * isamples) * r_shadow_bouncegrid_state.spacing[0];
+
+                                                               // measure distance from subsample to line segment and see if it is within radius
+                                                               along = DotProduct(sample, pathdir) * pathilength;
+                                                               if (along <= 0)
+                                                                       VectorCopy(pathstart, nearest);
+                                                               else if (along >= 1)
+                                                                       VectorCopy(pathend, nearest);
+                                                               else
+                                                                       VectorLerp(pathstart, along, pathend, nearest);
+                                                               VectorSubtract(sample, nearest, diff);
+                                                               VectorScale(diff, iradius, diff);
+                                                               distance2 = VectorLength2(diff);
+                                                               if (distance2 < 1.0f)
+                                                               {
+                                                                       // contribute some color to this pixel, across all bands
+                                                                       float w = 1.0f - sqrt(distance2);
+                                                                       int band;
+                                                                       w *= w;
+                                                                       if (pixelbands > 1)
+                                                                       {
+                                                                               // small optimization for alpha - only color[7] is non-zero, so skip the rest of the alpha elements.
+                                                                               p[pixelsperband * 4 + 3] += color[7] * w;
+                                                                       }
+                                                                       for (band = 0; band < pixelbands; band++)
+                                                                       {
+                                                                               // add to the pixel color (RGB only - see above)
+                                                                               p[band * pixelsperband * 4 + 0] += color[band * 4 + 0] * w;
+                                                                               p[band * pixelsperband * 4 + 1] += color[band * 4 + 1] * w;
+                                                                               p[band * pixelsperband * 4 + 2] += color[band * 4 + 2] * w;
+                                                                       }
+                                                               }
                                                        }
                                                }
                                        }
@@ -2276,6 +2331,32 @@ static void R_Shadow_BounceGrid_PerformSplats(void)
        }
 }
 
+static void R_Shadow_BounceGrid_Slice_Task(taskqueue_task_t *t)
+{
+       t->started = 1;
+       R_Shadow_BounceGrid_Slice((int)t->i[0]);
+       t->done = 1;
+}
+
+static void R_Shadow_BounceGrid_EnqueueSlices_Task(taskqueue_task_t *t)
+{
+       int i, slices;
+       // we need to wait for the texture clear to finish before we start adding light to it
+       if (r_shadow_bouncegrid_state.cleartex_task.done == 0)
+       {
+               TaskQueue_Yield(t);
+               return;
+       }
+       t->started = 1;
+       slices = r_shadow_bouncegrid_state.resolution[2] - 2;
+       for (i = 0; i < slices; i++)
+               TaskQueue_Setup(r_shadow_bouncegrid_state.slices_tasks + i, NULL, R_Shadow_BounceGrid_Slice_Task, i + 1, 0, NULL, NULL);
+       TaskQueue_Enqueue(slices, r_shadow_bouncegrid_state.slices_tasks);
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.slices_done_task, NULL, TaskQueue_Task_CheckTasksDone, slices, 0, r_shadow_bouncegrid_state.slices_tasks, 0);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.slices_done_task);
+       t->done = 1;
+}
+
 static void R_Shadow_BounceGrid_BlurPixelsInDirection(const float *inpixels, float *outpixels, int off)
 {
        const float *inpixel;
@@ -2308,31 +2389,32 @@ static void R_Shadow_BounceGrid_BlurPixelsInDirection(const float *inpixels, flo
        }
 }
 
-static void R_Shadow_BounceGrid_BlurPixels(void)
+static void R_Shadow_BounceGrid_BlurPixels_Task(taskqueue_task_t *t)
 {
        float *pixels[4];
        unsigned int resolution[3];
+       t->started = 1;
+       if (r_shadow_bouncegrid_state.settings.blur)
+       {
+               VectorCopy(r_shadow_bouncegrid_state.resolution, resolution);
 
-       if (!r_shadow_bouncegrid_state.settings.blur)
-               return;
-       
-       VectorCopy(r_shadow_bouncegrid_state.resolution, resolution);
-
-       pixels[0] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
-       pixels[1] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index ^ 1];
-       pixels[2] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
-       pixels[3] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index ^ 1];
+               pixels[0] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
+               pixels[1] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index ^ 1];
+               pixels[2] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
+               pixels[3] = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index ^ 1];
 
-       // blur on X
-       R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[0], pixels[1], 4);
-       // blur on Y
-       R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[1], pixels[2], resolution[0] * 4);
-       // blur on Z
-       R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[2], pixels[3], resolution[0] * resolution[1] * 4);
+               // blur on X
+               R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[0], pixels[1], 4);
+               // blur on Y
+               R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[1], pixels[2], resolution[0] * 4);
+               // blur on Z
+               R_Shadow_BounceGrid_BlurPixelsInDirection(pixels[2], pixels[3], resolution[0] * resolution[1] * 4);
 
-       // toggle the state, highpixels now points to pixels[3] result
-       r_shadow_bouncegrid_state.highpixels_index ^= 1;
-       r_shadow_bouncegrid_state.highpixels = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
+               // toggle the state, highpixels now points to pixels[3] result
+               r_shadow_bouncegrid_state.highpixels_index ^= 1;
+               r_shadow_bouncegrid_state.highpixels = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
+       }
+       t->done = 1;
 }
 
 static void R_Shadow_BounceGrid_ConvertPixelsAndUpload(void)
@@ -2509,192 +2591,178 @@ static void R_Shadow_BounceGrid_ConvertPixelsAndUpload(void)
        r_shadow_bouncegrid_state.lastupdatetime = realtime;
 }
 
-static void R_Shadow_BounceGrid_TracePhotons(r_shadow_bouncegrid_settings_t settings, unsigned int range, unsigned int range1, unsigned int range2, int flag)
+void R_Shadow_BounceGrid_ClearTex_Task(taskqueue_task_t *t)
 {
-       dlight_t *light;
-       int bouncecount;
-       int hitsupercontentsmask;
-       int skipsupercontentsmask;
-       int skipmaterialflagsmask;
-       int maxbounce;
-       int shootparticles;
-       int shotparticles;
-       float bounceminimumintensity2;
-       trace_t cliptrace;
-       //trace_t cliptrace2;
-       //trace_t cliptrace3;
-       unsigned int lightindex;
-       unsigned int seed;
-       randomseed_t randomseed;
-       vec3_t shotcolor;
-       vec3_t baseshotcolor;
-       vec3_t surfcolor;
-       vec3_t clipend;
-       vec3_t clipstart;
-       vec3_t clipdiff;
-       vec_t radius;
-       vec_t distancetraveled;
-       vec_t s;
-       rtlight_t *rtlight;
-
-       // compute a seed for the unstable random modes
-       Math_RandomSeed_FromInts(&randomseed, 0, 0, 0, realtime * 1000.0);
-       seed = realtime * 1000.0;
-
-       r_shadow_bouncegrid_state.numsplatpaths = 0;
+       t->started = 1;
+       memset(r_shadow_bouncegrid_state.highpixels, 0, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
+       t->done = 1;
+}
 
+static void R_Shadow_BounceGrid_TracePhotons_Shot(r_shadow_bouncegrid_photon_t *p, int remainingbounces, vec3_t shotstart, vec3_t shotend, vec3_t shotcolor, float bounceminimumintensity2, float previousrefractiveindex)
+{
+       int hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask;
+       vec3_t shothit;
+       vec3_t surfacenormal;
+       vec3_t reflectstart, reflectend, reflectcolor;
+       vec3_t refractstart, refractend, refractcolor;
+       vec_t s;
+       float reflectamount = 1.0f;
+       trace_t cliptrace;
        // figure out what we want to interact with
-       if (settings.hitmodels)
-               hitsupercontentsmask = SUPERCONTENTS_SOLID;// | SUPERCONTENTS_LIQUIDSMASK;
-       else
-               hitsupercontentsmask = SUPERCONTENTS_SOLID;// | SUPERCONTENTS_LIQUIDSMASK;
+       hitsupercontentsmask = SUPERCONTENTS_SOLID | SUPERCONTENTS_LIQUIDSMASK;
        skipsupercontentsmask = 0;
-       skipmaterialflagsmask = MATERIALFLAGMASK_TRANSLUCENT;
-       maxbounce = settings.maxbounce;
-
-       for (lightindex = 0;lightindex < range2;lightindex++)
+       skipmaterialflagsmask = MATERIALFLAG_CUSTOMBLEND;
+       //r_refdef.scene.worldmodel->TraceLineAgainstSurfaces(r_refdef.scene.worldmodel, NULL, NULL, &cliptrace, clipstart, clipend, hitsupercontentsmask);
+       //r_refdef.scene.worldmodel->TraceLine(r_refdef.scene.worldmodel, NULL, NULL, &cliptrace2, clipstart, clipend, hitsupercontentsmask);
+       if (r_shadow_bouncegrid_state.settings.staticmode || r_shadow_bouncegrid_state.settings.rng_seed < 0 || r_shadow_bouncegrid_threaded.integer)
        {
-               if (lightindex < range)
+               // static mode fires a LOT of rays but none of them are identical, so they are not cached
+               // non-stable random in dynamic mode also never reuses a direction, so there's no reason to cache it
+               cliptrace = CL_TraceLine(shotstart, shotend, r_shadow_bouncegrid_state.settings.staticmode ? MOVE_WORLDONLY : (r_shadow_bouncegrid_state.settings.hitmodels ? MOVE_HITMODEL : MOVE_NOMONSTERS), NULL, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask, collision_extendmovelength.value, true, false, NULL, true, true);
+       }
+       else
+       {
+               // dynamic mode fires many rays and most will match the cache from the previous frame
+               cliptrace = CL_Cache_TraceLineSurfaces(shotstart, shotend, r_shadow_bouncegrid_state.settings.staticmode ? MOVE_WORLDONLY : (r_shadow_bouncegrid_state.settings.hitmodels ? MOVE_HITMODEL : MOVE_NOMONSTERS), hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask);
+       }
+       VectorCopy(cliptrace.endpos, shothit);
+       if ((remainingbounces == r_shadow_bouncegrid_state.settings.maxbounce || r_shadow_bouncegrid_state.settings.includedirectlighting) && p->numpaths < PHOTON_MAX_PATHS)
+       {
+               qboolean notculled = true;
+               // cull paths that fail R_CullBox in dynamic mode
+               if (!r_shadow_bouncegrid_state.settings.staticmode
+                       && r_shadow_bouncegrid_dynamic_culllightpaths.integer)
                {
-                       light = (dlight_t *) Mem_ExpandableArray_RecordAtIndex(&r_shadow_worldlightsarray, lightindex);
-                       if (!light)
-                               continue;
-                       rtlight = &light->rtlight;
+                       vec3_t cullmins, cullmaxs;
+                       cullmins[0] = min(shotstart[0], shothit[0]) - r_shadow_bouncegrid_state.settings.spacing[0] - r_shadow_bouncegrid_state.settings.lightpathsize;
+                       cullmins[1] = min(shotstart[1], shothit[1]) - r_shadow_bouncegrid_state.settings.spacing[1] - r_shadow_bouncegrid_state.settings.lightpathsize;
+                       cullmins[2] = min(shotstart[2], shothit[2]) - r_shadow_bouncegrid_state.settings.spacing[2] - r_shadow_bouncegrid_state.settings.lightpathsize;
+                       cullmaxs[0] = max(shotstart[0], shothit[0]) + r_shadow_bouncegrid_state.settings.spacing[0] + r_shadow_bouncegrid_state.settings.lightpathsize;
+                       cullmaxs[1] = max(shotstart[1], shothit[1]) + r_shadow_bouncegrid_state.settings.spacing[1] + r_shadow_bouncegrid_state.settings.lightpathsize;
+                       cullmaxs[2] = max(shotstart[2], shothit[2]) + r_shadow_bouncegrid_state.settings.spacing[2] + r_shadow_bouncegrid_state.settings.lightpathsize;
+                       if (R_CullBox(cullmins, cullmaxs))
+                               notculled = false;
                }
-               else
-                       rtlight = r_refdef.scene.lights[lightindex - range];
-               // note that this code used to keep track of residual photons and
-               // distribute them evenly to achieve exactly a desired photon count,
-               // but that caused unwanted flickering in dynamic mode
-               shootparticles = (int)floor(rtlight->bouncegrid_photons);
-               // skip if we won't be shooting any photons
-               if (!shootparticles)
-                       continue;
-               radius = rtlight->radius * settings.lightradiusscale;
-               //s = settings.particleintensity / shootparticles;
-               //VectorScale(rtlight->bouncegrid_photoncolor, s, baseshotcolor);
-               VectorCopy(rtlight->bouncegrid_photoncolor, baseshotcolor);
-               if (VectorLength2(baseshotcolor) <= 0.0f)
-                       continue;
-               r_refdef.stats[r_stat_bouncegrid_lights]++;
-               r_refdef.stats[r_stat_bouncegrid_particles] += shootparticles;
-               // we stop caring about bounces once the brightness goes below this fraction of the original intensity
-               bounceminimumintensity2 = VectorLength(baseshotcolor) * settings.bounceminimumintensity2;
-
-               // for seeded random we start the RNG with the position of the light
-               if (settings.rng_seed >= 0)
+               if (notculled)
                {
-                       union
+                       r_shadow_bouncegrid_photon_path_t *path = p->paths + p->numpaths++;
+                       VectorCopy(shotstart, path->start);
+                       VectorCopy(shothit, path->end);
+                       VectorCopy(shotcolor, path->color);
+               }
+       }
+       if (cliptrace.fraction < 1.0f && remainingbounces > 0)
+       {
+               // scale down shot color by bounce intensity and texture color (or 50% if no texture reported)
+               // also clamp the resulting color to never add energy, even if the user requests extreme values
+               VectorCopy(cliptrace.plane.normal, surfacenormal);
+               VectorSet(reflectcolor, 0.5f, 0.5f, 0.5f);
+               VectorClear(refractcolor);
+               // FIXME: we need to determine the exact triangle, vertex color and texcoords and texture color and texture normal for the impacted point
+               if (cliptrace.hittexture)
+               {
+                       if (cliptrace.hittexture->currentskinframe)
+                               VectorCopy(cliptrace.hittexture->currentskinframe->avgcolor, reflectcolor);
+                       if (cliptrace.hittexture->currentalpha < 1.0f && (cliptrace.hittexture->currentmaterialflags & (MATERIALFLAG_ALPHA | MATERIALFLAG_ALPHATEST)))
                        {
-                               unsigned int i[4];
-                               float f[4];
+                               reflectamount *= cliptrace.hittexture->currentalpha;
+                               if (cliptrace.hittexture->currentskinframe)
+                                       reflectamount *= cliptrace.hittexture->currentskinframe->avgcolor[3];
                        }
-                       u;
-                       u.f[0] = rtlight->shadoworigin[0];
-                       u.f[1] = rtlight->shadoworigin[1];
-                       u.f[2] = rtlight->shadoworigin[2];
-                       u.f[3] = 1;
-                       switch (settings.rng_type)
+                       if (cliptrace.hittexture->currentmaterialflags & MATERIALFLAG_WATERSHADER)
                        {
-                       default:
-                       case 0:
-                               // we have to shift the seed provided by the user because the result must be odd
-                               Math_RandomSeed_FromInts(&randomseed, u.i[0], u.i[1], u.i[2], u.i[3] ^ (settings.rng_seed << 1));
-                               break;
-                       case 1:
-                               seed = u.i[0] ^ u.i[1] ^ u.i[2] ^ u.i[3] ^ settings.rng_seed;
-                               break;
+                               float Fresnel;
+                               vec3_t lightdir;
+                               //reflectchance = pow(min(1.0f, 1.0f - cliptrace.
+                               VectorSubtract(shotstart, shotend, lightdir);
+                               VectorNormalize(lightdir);
+                               Fresnel = min(1.0f, 1.0f - DotProduct(lightdir, surfacenormal));
+                               Fresnel = Fresnel * Fresnel * (cliptrace.hittexture->reflectmax - cliptrace.hittexture->reflectmin) + cliptrace.hittexture->reflectmin;
+                               reflectamount *= Fresnel;
+                               VectorCopy(cliptrace.hittexture->refractcolor4f, refractcolor);
                        }
+                       if (cliptrace.hittexture->currentmaterialflags & MATERIALFLAG_REFRACTION)
+                               VectorCopy(cliptrace.hittexture->refractcolor4f, refractcolor);
+                       // make sure we do not gain energy even if surface colors are out of bounds
+                       reflectcolor[0] = min(reflectcolor[0], 1.0f);
+                       reflectcolor[1] = min(reflectcolor[1], 1.0f);
+                       reflectcolor[2] = min(reflectcolor[2], 1.0f);
+                       refractcolor[0] = min(refractcolor[0], 1.0f);
+                       refractcolor[1] = min(refractcolor[1], 1.0f);
+                       refractcolor[2] = min(refractcolor[2], 1.0f);
                }
+               // reflected and refracted shots
+               VectorScale(reflectcolor, r_shadow_bouncegrid_state.settings.particlebounceintensity * reflectamount, reflectcolor);
+               VectorScale(refractcolor, (1.0f - reflectamount), refractcolor);
+               VectorMultiply(reflectcolor, shotcolor, reflectcolor);
+               VectorMultiply(refractcolor, shotcolor, refractcolor);
 
-               for (shotparticles = 0;shotparticles < shootparticles;shotparticles++)
+               if (VectorLength2(reflectcolor) >= bounceminimumintensity2)
                {
-                       VectorCopy(baseshotcolor, shotcolor);
-                       VectorCopy(rtlight->shadoworigin, clipstart);
-                       switch (settings.rng_type)
-                       {
-                       default:
-                       case 0:
-                               VectorLehmerRandom(&randomseed, clipend);
-                               break;
-                       case 1:
-                               VectorCheeseRandom(seed, clipend);
-                               break;
-                       }
-
-                       // we want a uniform distribution spherically, not merely within the sphere
-                       if (settings.normalizevectors)
-                               VectorNormalize(clipend);
+                       // reflect the remaining portion of the line across plane normal
+                       VectorSubtract(shotend, shothit, reflectend);
+                       VectorReflect(reflectend, 1.0, surfacenormal, reflectend);
+                       // calculate the new line start and end
+                       VectorCopy(shothit, reflectstart);
+                       VectorAdd(reflectstart, reflectend, reflectend);
+                       R_Shadow_BounceGrid_TracePhotons_Shot(p, remainingbounces - 1, reflectstart, reflectend, reflectcolor, bounceminimumintensity2, previousrefractiveindex);
+               }
 
-                       VectorMA(clipstart, radius, clipend, clipend);
-                       distancetraveled = 0.0f;
-                       for (bouncecount = 0;;bouncecount++)
-                       {
-                               r_refdef.stats[r_stat_bouncegrid_traces]++;
-                               rtlight->bouncegrid_traces++;
-                               //r_refdef.scene.worldmodel->TraceLineAgainstSurfaces(r_refdef.scene.worldmodel, NULL, NULL, &cliptrace, clipstart, clipend, hitsupercontentsmask);
-                               //r_refdef.scene.worldmodel->TraceLine(r_refdef.scene.worldmodel, NULL, NULL, &cliptrace2, clipstart, clipend, hitsupercontentsmask);
-                               if (settings.staticmode || settings.rng_seed < 0)
-                               {
-                                       // static mode fires a LOT of rays but none of them are identical, so they are not cached
-                                       // non-stable random in dynamic mode also never reuses a direction, so there's no reason to cache it
-                                       cliptrace = CL_TraceLine(clipstart, clipend, settings.staticmode ? MOVE_WORLDONLY : (settings.hitmodels ? MOVE_HITMODEL : MOVE_NOMONSTERS), NULL, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask, collision_extendmovelength.value, true, false, NULL, true, true);
-                               }
-                               else
-                               {
-                                       // dynamic mode fires many rays and most will match the cache from the previous frame
-                                       cliptrace = CL_Cache_TraceLineSurfaces(clipstart, clipend, settings.staticmode ? MOVE_WORLDONLY : (settings.hitmodels ? MOVE_HITMODEL : MOVE_NOMONSTERS), hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask);
-                               }
-                               if (bouncecount > 0 || settings.includedirectlighting)
-                               {
-                                       vec3_t hitpos;
-                                       VectorCopy(cliptrace.endpos, hitpos);
-                                       R_Shadow_BounceGrid_AddSplatPath(clipstart, hitpos, shotcolor, distancetraveled);
-                               }
-                               distancetraveled += VectorDistance(clipstart, cliptrace.endpos);
-                               s = VectorDistance(rtlight->shadoworigin, cliptrace.endpos);
-                               if (rtlight->bouncegrid_effectiveradius < s)
-                                       rtlight->bouncegrid_effectiveradius = s;
-                               if (cliptrace.fraction >= 1.0f)
-                                       break;
-                               r_refdef.stats[r_stat_bouncegrid_hits]++;
-                               rtlight->bouncegrid_hits++;
-                               if (bouncecount >= maxbounce)
-                                       break;
-                               // scale down shot color by bounce intensity and texture color (or 50% if no texture reported)
-                               // also clamp the resulting color to never add energy, even if the user requests extreme values
-                               if (cliptrace.hittexture && cliptrace.hittexture->currentskinframe)
-                                       VectorCopy(cliptrace.hittexture->currentskinframe->avgcolor, surfcolor);
-                               else
-                                       VectorSet(surfcolor, 0.5f, 0.5f, 0.5f);
-                               VectorScale(surfcolor, settings.particlebounceintensity, surfcolor);
-                               surfcolor[0] = min(surfcolor[0], 1.0f);
-                               surfcolor[1] = min(surfcolor[1], 1.0f);
-                               surfcolor[2] = min(surfcolor[2], 1.0f);
-                               VectorMultiply(shotcolor, surfcolor, shotcolor);
-                               if (VectorLength2(shotcolor) <= bounceminimumintensity2)
-                                       break;
-                               r_refdef.stats[r_stat_bouncegrid_bounces]++;
-                               // reflect the remaining portion of the line across plane normal
-                               VectorSubtract(clipend, cliptrace.endpos, clipdiff);
-                               VectorReflect(clipdiff, 1.0, cliptrace.plane.normal, clipend);
-                               // calculate the new line start and end
-                               VectorCopy(cliptrace.endpos, clipstart);
-                               VectorAdd(clipstart, clipend, clipend);
-                       }
+               if (VectorLength2(refractcolor) >= bounceminimumintensity2)
+               {
+                       // Check what refractive index is on the other side
+                       float refractiveindex;
+                       VectorMA(shothit, 0.0625f, cliptrace.plane.normal, refractstart);
+                       refractiveindex = R_Shadow_BounceGrid_RefractiveIndexAtPoint(refractstart);
+                       // reflect the remaining portion of the line across plane normal
+                       VectorSubtract(shotend, shothit, refractend);
+                       s = refractiveindex / previousrefractiveindex;
+                       VectorReflect(refractend, -1.0f / s, surfacenormal, refractend);
+                       // we also need to reflect the start to the other side of the plane so it doesn't just hit the same surface again
+                       // calculate the new line start and end
+                       VectorMA(shothit, 0.0625f, cliptrace.plane.normal, refractstart);
+                       VectorAdd(refractstart, refractend, refractend);
+                       R_Shadow_BounceGrid_TracePhotons_Shot(p, remainingbounces - 1, refractstart, refractend, refractcolor, bounceminimumintensity2, refractiveindex);
                }
        }
 }
 
+static void R_Shadow_BounceGrid_TracePhotons_ShotTask(taskqueue_task_t *t)
+{
+       r_shadow_bouncegrid_photon_t *p = (r_shadow_bouncegrid_photon_t *)t->p[0];
+       t->started = 1;
+       R_Shadow_BounceGrid_TracePhotons_Shot(p, r_shadow_bouncegrid_state.settings.maxbounce, p->start, p->end, p->color, p->bounceminimumintensity2, p->startrefractiveindex);
+       t->done = 1;
+}
+
+static void R_Shadow_BounceGrid_EnqueuePhotons_Task(taskqueue_task_t *t)
+{
+       int i;
+       t->started = 1;
+       for (i = 0; i < r_shadow_bouncegrid_state.numphotons; i++)
+               TaskQueue_Setup(r_shadow_bouncegrid_state.photons_tasks + i, NULL, R_Shadow_BounceGrid_TracePhotons_ShotTask, 0, 0, r_shadow_bouncegrid_state.photons + i, NULL);
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.photons_done_task, NULL, TaskQueue_Task_CheckTasksDone, r_shadow_bouncegrid_state.numphotons, 0, r_shadow_bouncegrid_state.photons_tasks, NULL);
+       if (r_shadow_bouncegrid_threaded.integer)
+       {
+               TaskQueue_Enqueue(r_shadow_bouncegrid_state.numphotons, r_shadow_bouncegrid_state.photons_tasks);
+               TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.photons_done_task);
+       }
+       else
+       {
+               // when not threaded we still have to report task status
+               for (i = 0; i < r_shadow_bouncegrid_state.numphotons; i++)
+                       r_shadow_bouncegrid_state.photons_tasks[i].func(r_shadow_bouncegrid_state.photons_tasks + i);
+               r_shadow_bouncegrid_state.photons_done_task.done = 1;
+       }
+       t->done = 1;
+}
+
 void R_Shadow_UpdateBounceGridTexture(void)
 {
        int flag = r_refdef.scene.rtworld ? LIGHTFLAG_REALTIMEMODE : LIGHTFLAG_NORMALMODE;
        r_shadow_bouncegrid_settings_t settings;
        qboolean enable = false;
        qboolean settingschanged;
-       unsigned int range; // number of world lights
-       unsigned int range1; // number of dynamic lights (or zero if disabled)
-       unsigned int range2; // range+range1
 
        enable = R_Shadow_BounceGrid_CheckEnable(flag);
        
@@ -2719,9 +2787,11 @@ void R_Shadow_UpdateBounceGridTexture(void)
                if (r_shadow_bouncegrid_state.blurpixels[1]) Mem_Free(r_shadow_bouncegrid_state.blurpixels[1]); r_shadow_bouncegrid_state.blurpixels[1] = NULL;
                if (r_shadow_bouncegrid_state.u8pixels) Mem_Free(r_shadow_bouncegrid_state.u8pixels); r_shadow_bouncegrid_state.u8pixels = NULL;
                if (r_shadow_bouncegrid_state.fp16pixels) Mem_Free(r_shadow_bouncegrid_state.fp16pixels); r_shadow_bouncegrid_state.fp16pixels = NULL;
-               if (r_shadow_bouncegrid_state.splatpaths) Mem_Free(r_shadow_bouncegrid_state.splatpaths); r_shadow_bouncegrid_state.splatpaths = NULL;
-               r_shadow_bouncegrid_state.maxsplatpaths = 0;
+               if (r_shadow_bouncegrid_state.photons) Mem_Free(r_shadow_bouncegrid_state.photons); r_shadow_bouncegrid_state.photons = NULL;
+               if (r_shadow_bouncegrid_state.photons_tasks) Mem_Free(r_shadow_bouncegrid_state.photons_tasks); r_shadow_bouncegrid_state.photons_tasks = NULL;
+               if (r_shadow_bouncegrid_state.slices_tasks) Mem_Free(r_shadow_bouncegrid_state.slices_tasks); r_shadow_bouncegrid_state.slices_tasks = NULL;
                r_shadow_bouncegrid_state.numpixels = 0;
+               r_shadow_bouncegrid_state.numphotons = 0;
                r_shadow_bouncegrid_state.directional = false;
 
                if (!enable)
@@ -2737,37 +2807,62 @@ void R_Shadow_UpdateBounceGridTexture(void)
 
        R_Shadow_BounceGrid_UpdateSpacing();
 
-       // get the range of light numbers we'll be looping over:
-       // range = static lights
-       // range1 = dynamic lights (optional)
-       // range2 = range + range1
-       range = (unsigned int)Mem_ExpandableArray_IndexRange(&r_shadow_worldlightsarray); // checked
-       range1 = settings.staticmode ? 0 : r_refdef.scene.numlights;
-       range2 = range + range1;
-
-       // calculate weighting factors for distributing photons among the lights
-       R_Shadow_BounceGrid_AssignPhotons(&settings, range, range1, range2, flag);
-       R_TimeReport("bg_assignphotons");
+       // allocate the highpixels array we'll be accumulating light into
+       if (r_shadow_bouncegrid_state.blurpixels[0] == NULL)
+               r_shadow_bouncegrid_state.blurpixels[0] = (float *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
+       if (r_shadow_bouncegrid_state.settings.blur && r_shadow_bouncegrid_state.blurpixels[1] == NULL)
+               r_shadow_bouncegrid_state.blurpixels[1] = (float *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.numpixels * sizeof(float[4]));
+       r_shadow_bouncegrid_state.highpixels_index = 0;
+       r_shadow_bouncegrid_state.highpixels = r_shadow_bouncegrid_state.blurpixels[r_shadow_bouncegrid_state.highpixels_index];
 
-       // trace the photons from lights and accumulate illumination
-       R_Shadow_BounceGrid_TracePhotons(settings, range, range1, range2, flag);
-       R_TimeReport("bg_tracephotons");
+       // set up the tracking of photon data
+       if (r_shadow_bouncegrid_state.photons == NULL)
+               r_shadow_bouncegrid_state.photons = (r_shadow_bouncegrid_photon_t *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.settings.maxphotons * sizeof(r_shadow_bouncegrid_photon_t));
+       if (r_shadow_bouncegrid_state.photons_tasks == NULL)
+               r_shadow_bouncegrid_state.photons_tasks = (taskqueue_task_t *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.settings.maxphotons * sizeof(taskqueue_task_t));
+       r_shadow_bouncegrid_state.numphotons = 0;
+
+       // set up the tracking of slice tasks
+       if (r_shadow_bouncegrid_state.slices_tasks == NULL)
+               r_shadow_bouncegrid_state.slices_tasks = (taskqueue_task_t *)Mem_Alloc(r_main_mempool, r_shadow_bouncegrid_state.resolution[2] * sizeof(taskqueue_task_t));
+
+       memset(&r_shadow_bouncegrid_state.cleartex_task, 0, sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.assignphotons_task, 0, sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.enqueuephotons_task, 0, sizeof(taskqueue_task_t));
+       memset(r_shadow_bouncegrid_state.photons_tasks, 0, r_shadow_bouncegrid_state.settings.maxphotons * sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.photons_done_task, 0, sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.enqueue_slices_task, 0, sizeof(taskqueue_task_t));
+       memset(r_shadow_bouncegrid_state.slices_tasks, 0, r_shadow_bouncegrid_state.resolution[2] * sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.slices_done_task, 0, sizeof(taskqueue_task_t));
+       memset(&r_shadow_bouncegrid_state.blurpixels_task, 0, sizeof(taskqueue_task_t));
 
        // clear the texture
-       R_Shadow_BounceGrid_ClearPixels();
-       R_TimeReport("bg_cleartex");
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.cleartex_task, NULL, R_Shadow_BounceGrid_ClearTex_Task, 0, 0, NULL, NULL);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.cleartex_task);
 
-       // accumulate the light splatting into texture
-       R_Shadow_BounceGrid_PerformSplats();
-       R_TimeReport("bg_lighttex");
+       // calculate weighting factors for distributing photons among the lights
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.assignphotons_task, NULL, R_Shadow_BounceGrid_AssignPhotons_Task, 0, 0, NULL, NULL);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.assignphotons_task);
+
+       // enqueue tasks to trace the photons from lights
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.enqueuephotons_task, &r_shadow_bouncegrid_state.assignphotons_task, R_Shadow_BounceGrid_EnqueuePhotons_Task, 0, 0, NULL, NULL);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.enqueuephotons_task);
+
+       // accumulate the light paths into texture
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.enqueue_slices_task, &r_shadow_bouncegrid_state.photons_done_task, R_Shadow_BounceGrid_EnqueueSlices_Task, 0, 0, NULL, NULL);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.enqueue_slices_task);
 
        // apply a mild blur filter to the texture
-       R_Shadow_BounceGrid_BlurPixels();
-       R_TimeReport("bg_blurtex");
+       TaskQueue_Setup(&r_shadow_bouncegrid_state.blurpixels_task, &r_shadow_bouncegrid_state.slices_done_task, R_Shadow_BounceGrid_BlurPixels_Task, 0, 0, NULL, NULL);
+       TaskQueue_Enqueue(1, &r_shadow_bouncegrid_state.blurpixels_task);
+
+       TaskQueue_WaitForTaskDone(&r_shadow_bouncegrid_state.blurpixels_task);
+       R_TimeReport("bouncegrid_gen");
 
        // convert the pixels to lower precision and upload the texture
+       // this unfortunately has to run on the main thread for OpenGL calls, so we have to block on the previous task...
        R_Shadow_BounceGrid_ConvertPixelsAndUpload();
-       R_TimeReport("bg_uploadtex");
+       R_TimeReport("bouncegrid_tex");
 
        // after we compute the static lighting we don't need to keep the highpixels array around
        if (settings.staticmode)
@@ -2777,8 +2872,9 @@ void R_Shadow_UpdateBounceGridTexture(void)
                if (r_shadow_bouncegrid_state.blurpixels[1]) Mem_Free(r_shadow_bouncegrid_state.blurpixels[1]); r_shadow_bouncegrid_state.blurpixels[1] = NULL;
                if (r_shadow_bouncegrid_state.u8pixels) Mem_Free(r_shadow_bouncegrid_state.u8pixels); r_shadow_bouncegrid_state.u8pixels = NULL;
                if (r_shadow_bouncegrid_state.fp16pixels) Mem_Free(r_shadow_bouncegrid_state.fp16pixels); r_shadow_bouncegrid_state.fp16pixels = NULL;
-               if (r_shadow_bouncegrid_state.splatpaths) Mem_Free(r_shadow_bouncegrid_state.splatpaths); r_shadow_bouncegrid_state.splatpaths = NULL;
-               r_shadow_bouncegrid_state.maxsplatpaths = 0;
+               if (r_shadow_bouncegrid_state.photons) Mem_Free(r_shadow_bouncegrid_state.photons); r_shadow_bouncegrid_state.photons = NULL;
+               if (r_shadow_bouncegrid_state.photons_tasks) Mem_Free(r_shadow_bouncegrid_state.photons_tasks); r_shadow_bouncegrid_state.photons_tasks = NULL;
+               if (r_shadow_bouncegrid_state.slices_tasks) Mem_Free(r_shadow_bouncegrid_state.slices_tasks); r_shadow_bouncegrid_state.slices_tasks = NULL;
        }
 }
 
index e1f00aa0855c094bb6f66aa64975a444694e08a0..778a44c4b07fdfd6b5ca31cda1bfa526e22a1cb1 100644 (file)
@@ -2,6 +2,8 @@
 #ifndef R_SHADOW_H
 #define R_SHADOW_H
 
+#include "thread.h"
+
 #define R_SHADOW_SHADOWMAP_NUMCUBEMAPS 8
 
 extern cvar_t r_shadow_bumpscale_basetexture;
@@ -54,9 +56,35 @@ typedef struct r_shadow_bouncegrid_settings_s
        int rng_type;
        int rng_seed;
        float bounceminimumintensity2;
+       int subsamples;
 }
 r_shadow_bouncegrid_settings_t;
 
+#define PHOTON_MAX_PATHS 11
+
+typedef struct r_shadow_bouncegrid_photon_path_s
+{
+       vec3_t start;
+       vec3_t end;
+       vec3_t color;
+}
+r_shadow_bouncegrid_photon_path_t;
+
+typedef struct r_shadow_bouncegrid_photon_s
+{
+       // parameters for tracing this photon
+       vec3_t start;
+       vec3_t end;
+       float color[3];
+       float bounceminimumintensity2;
+       float startrefractiveindex;
+
+       // results
+       int numpaths;
+       r_shadow_bouncegrid_photon_path_t paths[PHOTON_MAX_PATHS];
+}
+r_shadow_bouncegrid_photon_t;
+
 typedef struct r_shadow_bouncegrid_state_s
 {
        r_shadow_bouncegrid_settings_t settings;
@@ -78,16 +106,27 @@ typedef struct r_shadow_bouncegrid_state_s
        vec3_t mins;
        vec3_t maxs;
        vec3_t size;
-       int maxsplatpaths;
 
        // per-frame data that is very temporary
-       int numsplatpaths;
-       struct r_shadow_bouncegrid_splatpath_s *splatpaths;
        int highpixels_index; // which one is active - this toggles when doing blur
        float *highpixels; // equals blurpixels[highpixels_index]
        float *blurpixels[2];
        unsigned char *u8pixels; // temporary processing buffer when outputting to rgba8 format
        unsigned short *fp16pixels; // temporary processing buffer when outputting to rgba16f format
+                                                               // describe the photons we intend to shoot for threaded dispatch
+       int numphotons; // number of photons to shoot this frame, always <= settings.maxphotons
+       r_shadow_bouncegrid_photon_t *photons; // describes the photons being shot this frame
+
+       // tasks
+       taskqueue_task_t cleartex_task; // clears the highpixels array
+       taskqueue_task_t assignphotons_task; // sets the photon counts on lights, etc
+       taskqueue_task_t enqueuephotons_task; // enqueues tasks to shoot the photons
+       taskqueue_task_t *photons_tasks; // [maxphotons] taskqueue entries to perform the photon shots
+       taskqueue_task_t photons_done_task; // checks that all photon shots are completed
+       taskqueue_task_t enqueue_slices_task; // enqueues slice tasks to render the light accumulation into the texture
+       taskqueue_task_t *slices_tasks; // [resolution[1]] taskqueue entries to perform the light path accumulation into the texture
+       taskqueue_task_t slices_done_task; // checks that light accumulation in the texture is done
+       taskqueue_task_t blurpixels_task; // blurs the highpixels array
 }
 r_shadow_bouncegrid_state_t;
 
index 2747cdd63bfdff7b09477f78d79ae64f599c94da..643d448514fb0f4b2c1d7f9610478a94583bdde3 100644 (file)
@@ -47,6 +47,7 @@ void Sys_Quit (int returnvalue)
        // Unlock mutexes because the quit command may jump directly here, causing a deadlock
        Cbuf_UnlockThreadMutex();
        SV_UnlockThreadMutex();
+       TaskQueue_Frame(true);
 
        if (COM_CheckParm("-profilegameonly"))
                Sys_AllowProfiling(false);
index 7f590a47dd62f355e89be3435adb4da0ea33540d..85e5200140eb73d522a53c2ebcd24f5b8511dfbf 100644 (file)
--- a/thread.h
+++ b/thread.h
@@ -6,6 +6,9 @@
 // use recursive mutex (non-posix) extensions in thread_pthread
 #define THREADRECURSIVE
 
+typedef int Thread_SpinLock;
+typedef struct {int value;} Thread_Atomic;
+
 #define Thread_CreateMutex()              (_Thread_CreateMutex(__FILE__, __LINE__))
 #define Thread_DestroyMutex(m)            (_Thread_DestroyMutex(m, __FILE__, __LINE__))
 #define Thread_LockMutex(m)               (_Thread_LockMutex(m, __FILE__, __LINE__))
 #define Thread_CreateBarrier(count)       (_Thread_CreateBarrier(count, __FILE__, __LINE__))
 #define Thread_DestroyBarrier(barrier)    (_Thread_DestroyBarrier(barrier, __FILE__, __LINE__))
 #define Thread_WaitBarrier(barrier)       (_Thread_WaitBarrier(barrier, __FILE__, __LINE__))
+#define Thread_AtomicGet(a)               (_Thread_AtomicGet(a, __FILE__, __LINE__))
+#define Thread_AtomicSet(a, v)            (_Thread_AtomicSet(a, v, __FILE__, __LINE__))
+#define Thread_AtomicAdd(a, v)            (_Thread_AtomicAdd(a, v, __FILE__, __LINE__))
+#define Thread_AtomicIncRef(a)            (_Thread_AtomicIncRef(a, __FILE__, __LINE__))
+#define Thread_AtomicDecRef(a)            (_Thread_AtomicDecRef(a, __FILE__, __LINE__))
+#define Thread_AtomicTryLock(lock)        (_Thread_AtomicTryLock(lock, __FILE__, __LINE__))
+#define Thread_AtomicLock(lock)           (_Thread_AtomicLock(lock, __FILE__, __LINE__))
+#define Thread_AtomicUnlock(lock)         (_Thread_AtomicUnlock(lock, __FILE__, __LINE__))
 
 int Thread_Init(void);
 void Thread_Shutdown(void);
@@ -38,5 +49,63 @@ int _Thread_WaitThread(void *thread, int retval, const char *filename, int filel
 void *_Thread_CreateBarrier(unsigned int count, const char *filename, int fileline);
 void _Thread_DestroyBarrier(void *barrier, const char *filename, int fileline);
 void _Thread_WaitBarrier(void *barrier, const char *filename, int fileline);
+int _Thread_AtomicGet(Thread_Atomic *ref, const char *filename, int fileline);
+int _Thread_AtomicSet(Thread_Atomic *ref, int v, const char *filename, int fileline);
+int _Thread_AtomicAdd(Thread_Atomic *ref, int v, const char *filename, int fileline);
+void _Thread_AtomicIncRef(Thread_Atomic *ref, const char *filename, int fileline);
+qboolean _Thread_AtomicDecRef(Thread_Atomic *ref, const char *filename, int fileline);
+qboolean _Thread_AtomicTryLock(Thread_SpinLock *lock, const char *filename, int fileline);
+void _Thread_AtomicLock(Thread_SpinLock *lock, const char *filename, int fileline);
+void _Thread_AtomicUnlock(Thread_SpinLock *lock, const char *filename, int fileline);
+
+typedef struct taskqueue_task_s
+{
+       // doubly linked list
+       struct taskqueue_task_s * volatile prev;
+       struct taskqueue_task_s * volatile next;
+
+       // if not NULL, this task must be done before this one will dequeue (faster than simply Yielding immediately)
+       struct taskqueue_task_s *preceding;
+
+       // see TaskQueue_IsDone() to use proper atomics to poll done status
+       volatile int started;
+       volatile int done;
+
+       // function to call, and parameters for it to use
+       void(*func)(struct taskqueue_task_s *task);
+       void *p[4];
+       size_t i[4];
+
+       // stats:
+       unsigned int yieldcount; // number of times this task has been requeued
+}
+taskqueue_task_t;
+
+// immediately execute any pending tasks if threading is disabled (or if force is true)
+// TRY NOT TO USE THIS IF POSSIBLE - poll task->done instead.
+void TaskQueue_Execute(qboolean force);
+
+// queue the tasks to be executed, or executes them immediately if threading is disabled.
+void TaskQueue_Enqueue(int numtasks, taskqueue_task_t *tasks);
+
+// if the task can not be completed due yet to preconditions, just enqueue it again...
+void TaskQueue_Yield(taskqueue_task_t *t);
+
+// polls for status of task and returns the result immediately - use this instead of checking ->done directly, as this uses atomics
+qboolean TaskQueue_IsDone(taskqueue_task_t *t);
+
+// polls for status of task and waits for it to be done
+void TaskQueue_WaitForTaskDone(taskqueue_task_t *t);
+
+// updates thread count based on the cvar.
+void TaskQueue_Frame(qboolean shutdown);
+
+// convenience function for setting up a task structure.  Does not do the Enqueue, just fills in the struct.
+void TaskQueue_Setup(taskqueue_task_t *t, taskqueue_task_t *preceding, void(*func)(taskqueue_task_t *), size_t i0, size_t i1, void *p0, void *p1);
+
+// general purpose tasks
+// t->i[0] = number of tasks in array
+// t->p[0] = array of taskqueue_task_t to check
+void TaskQueue_Task_CheckTasksDone(taskqueue_task_t *t);
 
 #endif
index 65d14d7218b05d82e1b48d7d8f5682e7a2dedfd3..8431ead2290365e411c064d5e057cbe101e2b1b3 100644 (file)
@@ -3,16 +3,57 @@
 #include "quakedef.h"
 #include "thread.h"
 
+cvar_t taskqueue_maxthreads = {CVAR_SAVE, "taskqueue_maxthreads", "32", "how many threads to use for executing tasks"};
+cvar_t taskqueue_linkedlist = {CVAR_SAVE, "taskqueue_linkedlist", "1", "whether to use a doubly linked list or an array for the FIFO queue"};
+
+typedef struct taskqueue_state_thread_s
+{
+       void *handle;
+}
+taskqueue_state_thread_t;
+
+typedef struct taskqueue_state_s
+{
+       int numthreads;
+       taskqueue_state_thread_t threads[1024];
+
+       // we can enqueue this many tasks before execution of them must proceed
+       int queue_used;
+       int queue_max; // size of queue array
+       taskqueue_task_t **queue_tasks;
+
+       // command 
+       Thread_SpinLock command_lock;
+
+       volatile uint64_t threads_quit;
+
+       // doubly linked list - enqueue pushes to list.prev, dequeue pops from list.next
+       taskqueue_task_t list;
+}
+taskqueue_state_t;
+
+static taskqueue_state_t taskqueue_state;
+
 int Thread_Init(void)
 {
+       Cvar_RegisterVariable(&taskqueue_maxthreads);
+       Cvar_RegisterVariable(&taskqueue_linkedlist);
 #ifdef THREADDISABLE
        Con_Printf("Threading disabled in this build\n");
 #endif
+       // initialize the doubly-linked list header
+       taskqueue_state.list.next = &taskqueue_state.list;
+       taskqueue_state.list.prev = &taskqueue_state.list;
        return 0;
 }
 
 void Thread_Shutdown(void)
 {
+       if (taskqueue_state.numthreads)
+               TaskQueue_Frame(true);
+       if (taskqueue_state.queue_tasks)
+               Mem_Free(taskqueue_state.queue_tasks);
+       taskqueue_state.queue_tasks = NULL;
 }
 
 qboolean Thread_HasThreads(void)
@@ -167,3 +208,256 @@ void _Thread_WaitBarrier(void *barrier, const char *filename, int fileline)
        }
        Thread_UnlockMutex(b->mutex);
 }
+
+int _Thread_AtomicGet(Thread_Atomic *a, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic get at %s:%i\n", a, v, filename, fileline);
+#endif
+       return SDL_AtomicGet((SDL_atomic_t *)a);
+}
+
+int _Thread_AtomicSet(Thread_Atomic *a, int v, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic set %v at %s:%i\n", a, v, filename, fileline);
+#endif
+       return SDL_AtomicSet((SDL_atomic_t *)a, v);
+}
+
+int _Thread_AtomicAdd(Thread_Atomic *a, int v, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic add %v at %s:%i\n", a, v, filename, fileline);
+#endif
+       return SDL_AtomicAdd((SDL_atomic_t *)a, v);
+}
+
+void _Thread_AtomicIncRef(Thread_Atomic *a, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic incref %s:%i\n", lock, filename, fileline);
+#endif
+       SDL_AtomicIncRef((SDL_atomic_t *)a);
+}
+
+qboolean _Thread_AtomicDecRef(Thread_Atomic *a, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic decref %s:%i\n", lock, filename, fileline);
+#endif
+       return SDL_AtomicDecRef((SDL_atomic_t *)a) != SDL_FALSE;
+}
+
+qboolean _Thread_AtomicTryLock(Thread_SpinLock *lock, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic try lock %s:%i\n", lock, filename, fileline);
+#endif
+       return SDL_AtomicTryLock(lock) != SDL_FALSE;
+}
+
+void _Thread_AtomicLock(Thread_SpinLock *lock, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic lock %s:%i\n", lock, filename, fileline);
+#endif
+       SDL_AtomicLock(lock);
+}
+
+void _Thread_AtomicUnlock(Thread_SpinLock *lock, const char *filename, int fileline)
+{
+#ifdef THREADDEBUG
+       Sys_PrintfToTerminal("%p atomic unlock %s:%i\n", lock, filename, fileline);
+#endif
+       SDL_AtomicUnlock(lock);
+}
+
+static taskqueue_task_t *TaskQueue_GetPending(void)
+{
+       taskqueue_task_t *t = NULL;
+       if (taskqueue_state.list.next != &taskqueue_state.list)
+       {
+               // pop from list.next
+               t = taskqueue_state.list.next;
+               t->next->prev = t->prev;
+               t->prev->next = t->next;
+               t->prev = t->next = NULL;
+       }
+       if (t == NULL)
+       {
+               if (taskqueue_state.queue_used > 0)
+               {
+                       t = taskqueue_state.queue_tasks[0];
+                       taskqueue_state.queue_used--;
+                       memmove(taskqueue_state.queue_tasks, taskqueue_state.queue_tasks + 1, taskqueue_state.queue_used * sizeof(taskqueue_task_t *));
+                       taskqueue_state.queue_tasks[taskqueue_state.queue_used] = NULL;
+               }
+       }
+       return t;
+}
+
+static void TaskQueue_ExecuteTask(taskqueue_task_t *t)
+{
+       // see if t is waiting on something
+       if (t->preceding && t->preceding->done == 0)
+               TaskQueue_Yield(t);
+       else
+               t->func(t);
+}
+
+// FIXME: don't use mutex
+// FIXME: this is basically fibers but less featureful - context switching for yield is not implemented
+static int TaskQueue_ThreadFunc(void *d)
+{
+       for (;;)
+       {
+               qboolean quit;
+               taskqueue_task_t *t = NULL;
+               Thread_AtomicLock(&taskqueue_state.command_lock);
+               quit = taskqueue_state.threads_quit != 0;
+               t = TaskQueue_GetPending();
+               Thread_AtomicUnlock(&taskqueue_state.command_lock);
+               if (t)
+                       TaskQueue_ExecuteTask(t);
+               else if (quit)
+                       break;
+       }
+       return 0;
+}
+
+void TaskQueue_Execute(qboolean force)
+{
+       // if we have no threads to run the tasks, just start executing them now
+       if (taskqueue_state.numthreads == 0 || force)
+       {
+               for (;;)
+               {
+                       taskqueue_task_t *t = NULL;
+                       Thread_AtomicLock(&taskqueue_state.command_lock);
+                       t = TaskQueue_GetPending();
+                       Thread_AtomicUnlock(&taskqueue_state.command_lock);
+                       if (!t)
+                               break;
+                       TaskQueue_ExecuteTask(t);
+               }
+       }
+}
+
+void TaskQueue_Enqueue(int numtasks, taskqueue_task_t *tasks)
+{
+       int i;
+       // try not to spinlock for a long time by breaking up large enqueues
+       while (numtasks > 64)
+       {
+               TaskQueue_Enqueue(64, tasks);
+               tasks += 64;
+               numtasks -= 64;
+       }
+       Thread_AtomicLock(&taskqueue_state.command_lock);
+       for (i = 0; i < numtasks; i++)
+       {
+               taskqueue_task_t *t = &tasks[i];
+               if (taskqueue_linkedlist.integer)
+               {
+                       // push to list.prev
+                       t->next = &taskqueue_state.list;
+                       t->prev = taskqueue_state.list.prev;
+                       t->next->prev = t;
+                       t->prev->next = t;
+               }
+               else
+               {
+                       if (taskqueue_state.queue_used >= taskqueue_state.queue_max)
+                       {
+                               taskqueue_state.queue_max *= 2;
+                               if (taskqueue_state.queue_max < 1024)
+                                       taskqueue_state.queue_max = 1024;
+                               taskqueue_state.queue_tasks = (taskqueue_task_t **)Mem_Realloc(cls.permanentmempool, taskqueue_state.queue_tasks, taskqueue_state.queue_max * sizeof(taskqueue_task_t *));
+                       }
+                       taskqueue_state.queue_tasks[taskqueue_state.queue_used++] = t;
+               }
+       }
+       Thread_AtomicUnlock(&taskqueue_state.command_lock);
+}
+
+// if the task can not be completed due yet to preconditions, just enqueue it again...
+void TaskQueue_Yield(taskqueue_task_t *t)
+{
+       t->yieldcount++;
+       TaskQueue_Enqueue(1, t);
+}
+
+void TaskQueue_WaitForTaskDone(taskqueue_task_t *t)
+{
+       qboolean done = false;
+       while (!done)
+       {
+               Thread_AtomicLock(&taskqueue_state.command_lock);
+               done = t->done != 0;
+               Thread_AtomicUnlock(&taskqueue_state.command_lock);
+               // if there are no threads, just execute the tasks immediately
+               if (!done && taskqueue_state.numthreads == 0)
+                       TaskQueue_Execute(true);
+       }
+}
+
+void TaskQueue_Frame(qboolean shutdown)
+{
+       int numthreads = shutdown ? 0 : bound(0, taskqueue_maxthreads.integer, sizeof(taskqueue_state.threads)/sizeof(taskqueue_state.threads[0]));
+#ifdef THREADDISABLE
+       numthreads = 0;
+#endif
+       if (taskqueue_state.numthreads != numthreads)
+       {
+               int i;
+               Thread_AtomicLock(&taskqueue_state.command_lock);
+               taskqueue_state.threads_quit = 1;
+               Thread_AtomicUnlock(&taskqueue_state.command_lock);
+               for (i = 0; i < taskqueue_state.numthreads; i++)
+               {
+                       if (taskqueue_state.threads[i].handle)
+                               Thread_WaitThread(taskqueue_state.threads[i].handle, 0);
+                       taskqueue_state.threads[i].handle = NULL;
+               }
+               Thread_AtomicLock(&taskqueue_state.command_lock);
+               taskqueue_state.threads_quit = 0;
+               Thread_AtomicUnlock(&taskqueue_state.command_lock);
+               taskqueue_state.numthreads = numthreads;
+               for (i = 0; i < taskqueue_state.numthreads; i++)
+                       taskqueue_state.threads[i].handle = Thread_CreateThread(TaskQueue_ThreadFunc, &taskqueue_state.threads[i]);
+               // if there are still pending tasks (e.g. no threads), execute them on main thread now
+               TaskQueue_Execute(true);
+       }
+}
+
+void TaskQueue_Setup(taskqueue_task_t *t, taskqueue_task_t *preceding, void(*func)(taskqueue_task_t *), size_t i0, size_t i1, void *p0, void *p1)
+{
+       memset(t, 0, sizeof(*t));
+       t->preceding = preceding;
+       t->func = func;
+       t->i[0] = i0;
+       t->i[1] = i1;
+       t->p[0] = p0;
+       t->p[1] = p1;
+}
+
+void TaskQueue_Task_CheckTasksDone(taskqueue_task_t *t)
+{
+       size_t numtasks = t->i[0];
+       taskqueue_task_t *tasks = t->p[0];
+       while (numtasks > 0)
+       {
+               // check the last task first as it's usually going to be the last to finish, so we do the least work by checking it first
+               if (!tasks[numtasks - 1].done)
+               {
+                       // update our partial progress, then yield to another pending task.
+                       t->i[0] = numtasks;
+                       TaskQueue_Yield(t);
+                       return;
+               }
+               numtasks--;
+       }
+       t->started = 1;
+       t->done = 1;
+}