From: havoc Date: Mon, 28 Nov 2016 03:56:16 +0000 (+0000) Subject: Fix severe bugs in mul128 implementation which was breaking the Lehmer RNG. X-Git-Tag: xonotic-v0.8.5~88^2~21 X-Git-Url: http://de.git.xonotic.org/?p=xonotic%2Fdarkplaces.git;a=commitdiff_plain;h=bbb2827707bfbc2d2a7d3304473cafadffac5da5 Fix severe bugs in mul128 implementation which was breaking the Lehmer RNG. Added gcc specific implementation of mul128 using __int128 intrinsic type. Added unit tests for mul128 implementation. Added r_shadow_bouncegrid_normalizevectors cvar, default on, this slightly brightens typical scenes, some more affected than others. Changed the behavior of r_shadow_bouncegrid_dynamic_stablerandom to also affect static mode, changed it to use the light position as the RNG seed, and changed the meaning of stablerandom 1 to be lhcheeserandom again as it is fast and of sufficient quality, the Lehmer RNG is available as stablerandom 2 but it has a framerate hit in dynamic mode that I can't ignore. git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@12301 d7cf8633-e32d-0410-b094-e92efae38249 --- diff --git a/mathlib.c b/mathlib.c index b2285768..119f8bf7 100644 --- a/mathlib.c +++ b/mathlib.c @@ -836,6 +836,7 @@ float RadiusFromBoundsAndOrigin (const vec3_t mins, const vec3_t maxs, const vec return sqrt(max(m1[0], m2[0]) + max(m1[1], m2[1]) + max(m1[2], m2[2])); } +static void Math_RandomSeed_UnitTests(void); void Mathlib_Init(void) { int a; @@ -844,6 +845,8 @@ void Mathlib_Init(void) ixtable[0] = 0; for (a = 1;a < 4096;a++) ixtable[a] = 1.0f / a; + + Math_RandomSeed_UnitTests(); } #include "matrixlib.h" @@ -902,30 +905,129 @@ int LoopingFrameNumberFromDouble(double t, int loopframes) static unsigned int mul_Lecuyer[4] = { 0x12e15e35, 0xb500f16e, 0x2e714eb2, 0xb37916a5 }; -static void mul128(unsigned int a[], unsigned int b[], unsigned int dest[4]) +static void mul128(const unsigned int a[], const unsigned int b[], unsigned int dest[4]) { +#ifdef __GNUC__ + unsigned __int128 ia = (a[0] << 96) + (a[1] << 64) + (a[2] << 32) + (a[3]); + unsigned __int128 ib = (b[0] << 96) + (b[1] << 64) + (b[2] << 32) + (b[3]); + unsigned __int128 id = ia * ib; + dest[0] = (id >> 96) & 0xffffffff; + dest[1] = (id >> 64) & 0xffffffff; + dest[2] = (id >> 32) & 0xffffffff; + dest[3] = (id) & 0xffffffff; +#else unsigned long long t[4]; - t[0] = a[0] * b[0]; - t[1] = a[1] * b[1]; - t[2] = a[2] * b[2]; - t[3] = a[3] * b[3]; - - // this is complicated because C doesn't have a way to make use of the - // cpu status carry flag, so we do it all in reverse order from what - // would otherwise make sense, and have to make multiple passes... - t[3] += t[2] >> 32; t[2] &= 0xffffffff; - t[2] += t[1] >> 32; t[1] &= 0xffffffff; - t[1] += t[0] >> 32; t[0] &= 0xffffffff; - t[3] += t[2] >> 32; t[2] &= 0xffffffff; - t[2] += t[1] >> 32; t[1] &= 0xffffffff; - - t[3] += t[2] >> 32; t[2] &= 0xffffffff; + // this multiply chain is relatively straightforward - a[] is repeatedly + // added with shifts based on b[] and the results stored into uint64, + // but due to C limitations (no access to carry flag) we have to resolve + // carries in a really lame way which wastes a fair number of ops + // (repeatedly iterating MSB to LSB, rather than LSB to MSB with carry), + // an alternative would be to use 16bit multiplies and resolve carries + // only at the end, but that would be twice as many multiplies... + // + // note: >> 32 is a function call in win32 MSVS2015 debug builds. + t[0] = (unsigned long long)a[0] * b[3]; + t[1] = (unsigned long long)a[1] * b[3]; + t[2] = (unsigned long long)a[2] * b[3]; + t[3] = (unsigned long long)a[3] * b[3]; + t[0] += t[1] >> 32; + t[1] &= 0xffffffff; + t[1] += t[2] >> 32; + t[2] &= 0xffffffff; + t[2] += t[3] >> 32; + + t[0] += t[1] >> 32; + t[1] &= 0xffffffff; + t[1] += t[2] >> 32; + t[2] &= 0xffffffff; + + t[0] += t[1] >> 32; + t[1] &= 0xffffffff; + + t[0] += (unsigned long long)a[1] * b[2]; + t[1] += (unsigned long long)a[2] * b[2]; + t[2] += (unsigned long long)a[3] * b[2]; + t[0] += t[1] >> 32; + t[1] &= 0xffffffff; + t[1] += t[2] >> 32; + + t[0] += t[1] >> 32; + t[1] &= 0xffffffff; + + t[0] += (unsigned long long)a[2] * b[1]; + t[1] += (unsigned long long)a[3] * b[1]; + t[0] += t[1] >> 32; + + t[0] += (unsigned long long)a[3] * b[0]; dest[0] = t[0] & 0xffffffff; dest[1] = t[1] & 0xffffffff; dest[2] = t[2] & 0xffffffff; dest[3] = t[3] & 0xffffffff; +#endif +} + +void testmul128(unsigned int a0, unsigned int a1, unsigned int a2, unsigned int a3, unsigned int b0, unsigned int b1, unsigned int b2, unsigned int b3, unsigned int x0, unsigned int x1, unsigned int x2, unsigned int x3) +{ + unsigned int a[4]; + unsigned int b[4]; + unsigned int expected[4]; + unsigned int result[4]; + a[0] = a0; + a[1] = a1; + a[2] = a2; + a[3] = a3; + b[0] = b0; + b[1] = b1; + b[2] = b2; + b[3] = b3; + expected[0] = x0; + expected[1] = x1; + expected[2] = x2; + expected[3] = x3; + mul128(a, b, result); + if (result[0] != expected[0] + || result[1] != expected[1] + || result[2] != expected[2] + || result[3] != expected[3]) + Con_Printf("testmul128(\na = %08x %08x %08x %08x,\nb = %08x %08x %08x %08x,\nx = %08x %08x %08x %08x) instead computed\nc = %08x %08x %08x %08x\n", a[0], a[1], a[2], a[3], b[0], b[1], b[2], b[3], expected[0], expected[1], expected[2], expected[3], result[0], result[1], result[2], result[3]); +} + +void Math_RandomSeed_UnitTests(void) +{ + testmul128( + 0x00000000, 0x00000000, 0x00000000, 0x00000001, + 0x00000000, 0x00000000, 0x00000000, 0x00000001, + 0x00000000, 0x00000000, 0x00000000, 0x00000001); + testmul128( + 0x00000000, 0x00000000, 0x00000000, 0x00000001, + 0x00000000, 0x00000000, 0x00000001, 0x00000000, + 0x00000000, 0x00000000, 0x00000001, 0x00000000); + testmul128( + 0x00000000, 0x00000000, 0x00000001, 0x00000000, + 0x00000000, 0x00000000, 0x00000000, 0x00000001, + 0x00000000, 0x00000000, 0x00000001, 0x00000000); + testmul128( + 0x00000000, 0x00000000, 0x00000000, 0x00000001, + 0x00000001, 0x00000001, 0x00000001, 0x00000001, + 0x00000001, 0x00000001, 0x00000001, 0x00000001); + testmul128( + 0x00000000, 0x00000000, 0x00000000, 0x00000002, + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, + 0xffffffff, 0xffffffff, 0xffffffff, 0xfffffffe); + testmul128( + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, + 0x00000000, 0x00000000, 0x00000000, 0x00000002, + 0xffffffff, 0xffffffff, 0xffffffff, 0xfffffffe); + testmul128( + 0x00000000, 0x00000000, 0xffffffff, 0xffffffff, + 0x00000000, 0x00000000, 0x00000002, 0x00000000, + 0x00000001, 0xffffffff, 0xfffffffe, 0x00000000); + testmul128( + 0x00000000, 0x00000000, 0x00000002, 0x00000000, + 0x00000000, 0x00000000, 0xffffffff, 0xffffffff, + 0x00000001, 0xffffffff, 0xfffffffe, 0x00000000); } void Math_RandomSeed_Reset(randomseed_t *r) @@ -936,13 +1038,12 @@ void Math_RandomSeed_Reset(randomseed_t *r) r->s[3] = 0; } -void Math_RandomSeed_FromInt(randomseed_t *r, unsigned int n) +void Math_RandomSeed_FromInts(randomseed_t *r, unsigned int s0, unsigned int s1, unsigned int s2, unsigned int s3) { - // if the entire s[] is zero the algorithm would break completely, so make sure it isn't zero by putting a 1 here - r->s[0] = 1; - r->s[1] = 0; - r->s[2] = 0; - r->s[3] = n; + r->s[0] = s0; + r->s[1] = s1; + r->s[2] = s2; + r->s[3] = s3 | 1; // the Lehmer RNG requires that the seed be odd } unsigned long long Math_rand64(randomseed_t *r) diff --git a/mathlib.h b/mathlib.h index 472d5759..0acb9500 100644 --- a/mathlib.h +++ b/mathlib.h @@ -311,7 +311,7 @@ typedef struct randomseed_s randomseed_t; void Math_RandomSeed_Reset(randomseed_t *r); -void Math_RandomSeed_FromInt(randomseed_t *r, unsigned int n); +void Math_RandomSeed_FromInts(randomseed_t *r, unsigned int s0, unsigned int s1, unsigned int s2, unsigned int s3); unsigned long long Math_rand64(randomseed_t *r); float Math_randomf(randomseed_t *r); float Math_crandomf(randomseed_t *r); diff --git a/r_shadow.c b/r_shadow.c index caef494e..0d05ac1e 100644 --- a/r_shadow.c +++ b/r_shadow.c @@ -359,6 +359,7 @@ cvar_t r_shadow_bouncegrid_particlebounceintensity = {CVAR_SAVE, "r_shadow_bounc cvar_t r_shadow_bouncegrid_particleintensity = {CVAR_SAVE, "r_shadow_bouncegrid_particleintensity", "0.25", "brightness of particles contributing to bouncegrid texture"}; cvar_t r_shadow_bouncegrid_sortlightpaths = {CVAR_SAVE, "r_shadow_bouncegrid_sortlightpaths", "1", "sort light paths before accumulating them into the bouncegrid texture, this reduces cpu cache misses"}; cvar_t r_shadow_bouncegrid_lightpathsize = {CVAR_SAVE, "r_shadow_bouncegrid_lightpathsize", "1", "width 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_static = {CVAR_SAVE, "r_shadow_bouncegrid_static", "1", "use static radiosity solution (high quality) rather than dynamic (splotchy)"}; cvar_t r_shadow_bouncegrid_static_bounceminimumintensity = { CVAR_SAVE, "r_shadow_bouncegrid_static_bounceminimumintensity", "0.01", "stop bouncing once intensity drops below this fraction of the original particle color" }; cvar_t r_shadow_bouncegrid_static_directionalshading = {CVAR_SAVE, "r_shadow_bouncegrid_static_directionalshading", "1", "whether to use directionalshading when in static mode"}; @@ -823,6 +824,7 @@ void R_Shadow_Init(void) Cvar_RegisterVariable(&r_shadow_bouncegrid_includedirectlighting); Cvar_RegisterVariable(&r_shadow_bouncegrid_intensity); Cvar_RegisterVariable(&r_shadow_bouncegrid_lightpathsize); + Cvar_RegisterVariable(&r_shadow_bouncegrid_normalizevectors); Cvar_RegisterVariable(&r_shadow_bouncegrid_particlebounceintensity); Cvar_RegisterVariable(&r_shadow_bouncegrid_particleintensity); Cvar_RegisterVariable(&r_shadow_bouncegrid_sortlightpaths); @@ -2575,8 +2577,9 @@ static void R_Shadow_BounceGrid_GenerateSettings(r_shadow_bouncegrid_settings_t settings->spacing[0] = spacing; settings->spacing[1] = spacing; settings->spacing[2] = spacing; - settings->stablerandom = s ? 1 : r_shadow_bouncegrid_dynamic_stablerandom.integer; + settings->stablerandom = r_shadow_bouncegrid_dynamic_stablerandom.integer; settings->bounceminimumintensity2 = bounceminimumintensity * bounceminimumintensity; + settings->normalizevectors = r_shadow_bouncegrid_normalizevectors.integer != 0; // bound the values for sanity settings->maxphotons = bound(1, settings->maxphotons, 25000000); @@ -3273,7 +3276,7 @@ static void R_Shadow_BounceGrid_TracePhotons(r_shadow_bouncegrid_settings_t sett //trace_t cliptrace2; //trace_t cliptrace3; unsigned int lightindex; - unsigned int seed = (unsigned int)(realtime * 1000.0f); + unsigned int seed; randomseed_t randomseed; vec3_t shotcolor; vec3_t baseshotcolor; @@ -3284,8 +3287,18 @@ static void R_Shadow_BounceGrid_TracePhotons(r_shadow_bouncegrid_settings_t sett vec_t radius; vec_t s; rtlight_t *rtlight; + union + { + unsigned int s[4]; + double d; + } + rseed; - Math_RandomSeed_FromInt(&randomseed, seed); + // compute a seed for the unstable random modes + memset(&rseed, 0, sizeof(rseed)); + rseed.d = realtime; + Math_RandomSeed_FromInts(&randomseed, rseed.s[0], rseed.s[1], rseed.s[2], rseed.s[3]); + seed = rseed.s[0] ^ rseed.s[1] ^ rseed.s[2] ^ rseed.s[3]; r_shadow_bouncegrid_state.numsplatpaths = 0; @@ -3325,21 +3338,33 @@ static void R_Shadow_BounceGrid_TracePhotons(r_shadow_bouncegrid_settings_t sett 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; - switch (settings.stablerandom) + + // for stablerandom we start the RNG with the position of the light + if (settings.stablerandom > 0) { - default: - break; - case 1: - Math_RandomSeed_FromInt(&randomseed, lightindex * 11937); - // prime the random number generator a bit - Math_crandomf(&randomseed); - break; - case 2: - seed = lightindex * 11937; - // prime the random number generator a bit - lhcheeserand(seed); - break; + 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 (settings.stablerandom) + { + default: + break; + case 1: + seed = u.i[0] ^ u.i[1] ^ u.i[2] ^ u.i[3]; + break; + case 2: + Math_RandomSeed_FromInts(&randomseed, u.i[0], u.i[1], u.i[2], u.i[3]); + break; + } } + for (shotparticles = 0;shotparticles < shootparticles;shotparticles++) { VectorCopy(baseshotcolor, shotcolor); @@ -3358,25 +3383,30 @@ static void R_Shadow_BounceGrid_TracePhotons(r_shadow_bouncegrid_settings_t sett break; case -1: case 1: - VectorLehmerRandom(&randomseed, clipend); + VectorCheeseRandom(seed, clipend); if (settings.bounceanglediffuse) { // we want random to be stable, so we still have to do all the random we would have done for (bouncecount = 0; bouncecount < maxbounce; bouncecount++) - VectorLehmerRandom(&randomseed, bouncerandom[bouncecount]); + VectorCheeseRandom(seed, bouncerandom[bouncecount]); } break; case -2: case 2: - VectorCheeseRandom(seed, clipend); + VectorLehmerRandom(&randomseed, clipend); if (settings.bounceanglediffuse) { // we want random to be stable, so we still have to do all the random we would have done for (bouncecount = 0; bouncecount < maxbounce; bouncecount++) - VectorCheeseRandom(seed, bouncerandom[bouncecount]); + VectorLehmerRandom(&randomseed, bouncerandom[bouncecount]); } break; } + + // we want a uniform distribution spherically, not merely within the sphere + if (settings.normalizevectors) + VectorNormalize(clipend); + VectorMA(clipstart, radius, clipend, clipend); for (bouncecount = 0;;bouncecount++) { diff --git a/r_shadow.h b/r_shadow.h index 10f8e1e0..fc36f337 100644 --- a/r_shadow.h +++ b/r_shadow.h @@ -45,6 +45,7 @@ typedef struct r_shadow_bouncegrid_settings_s qboolean directionalshading; qboolean includedirectlighting; qboolean blur; + qboolean normalizevectors; int floatcolors; float dlightparticlemultiplier; qboolean hitmodels;