1 | #include "../shaderenv.h"
|
---|
2 | #include "common.h"
|
---|
3 |
|
---|
4 | ////////////////////
|
---|
5 | // Screen Spaced Ambient Occlusion shader
|
---|
6 | // based on shader of Alexander Kusternig
|
---|
7 |
|
---|
8 |
|
---|
9 | #define USE_EYESPACE_DEPTH 1
|
---|
10 |
|
---|
11 |
|
---|
12 | struct fragment
|
---|
13 | {
|
---|
14 | float2 texCoord: TEXCOORD0;
|
---|
15 | float3 view: TEXCOORD1;
|
---|
16 | };
|
---|
17 |
|
---|
18 |
|
---|
19 | struct pixel2
|
---|
20 | {
|
---|
21 | float4 illum_col: COLOR0;
|
---|
22 | };
|
---|
23 |
|
---|
24 |
|
---|
25 | // this function is inspired from the paper of shamulgaan in order
|
---|
26 | // to get a physical expression for the occlusion culling
|
---|
27 | inline float occlusionPower(float radius, float dist)
|
---|
28 | {
|
---|
29 | return 6.283185307179586476925286766559f * (1.0f - cos(asin(radius / dist)));
|
---|
30 | }
|
---|
31 |
|
---|
32 |
|
---|
33 |
|
---|
34 | // reconstruct world space position
|
---|
35 | inline float3 ReconstructSamplePos(float eyeSpaceDepth,
|
---|
36 | float2 texcoord,
|
---|
37 | float3 bl, float3 br, float3 tl, float3 tr)
|
---|
38 | {
|
---|
39 | float3 viewVec = Interpol(texcoord, bl, br, tl, tr);
|
---|
40 | float3 samplePos = -viewVec * eyeSpaceDepth;
|
---|
41 |
|
---|
42 | return samplePos;
|
---|
43 | }
|
---|
44 |
|
---|
45 |
|
---|
46 | float ComputeConvergence(uniform sampler2D tex, float2 texCoord, float2 res)
|
---|
47 | {
|
---|
48 | // get the minimum convergence by exactly sampling the 4 surrounding
|
---|
49 | // texels in the old texture, otherwise flickering because convergence
|
---|
50 | // will be interpolated when upsampling and filter size does not match!
|
---|
51 |
|
---|
52 | const float2 invRes = float2(1.0f / res.x, 1.0f / res.y);
|
---|
53 |
|
---|
54 | // get position exactly between texel centers
|
---|
55 | float2 center = (floor(texCoord * res) + float2(.5f)) * texCoord;
|
---|
56 | //center.x = (floor(texCoord.x * res.x - .5f) + 1.0f) / res.x;
|
---|
57 | //center.y = (floor(texCoord.y * res.y - .5f) + 1.0f) / res.y;
|
---|
58 | //center.y = (floor(texCoord.y * res.y) + .5f) * yOffs;
|
---|
59 |
|
---|
60 | /*texelCenterConv.x = tex2Dlod(tex, float4(center + float2( xoffs, yoffs), 0, 0)).y;
|
---|
61 | texelCenterConv.y = tex2Dlod(tex, float4(center + float2( xoffs, -yoffs), 0, 0)).y;
|
---|
62 | texelCenterConv.z = tex2Dlod(tex, float4(center + float2(-xoffs, -yoffs), 0, 0)).y;
|
---|
63 | texelCenterConv.w = tex2Dlod(tex, float4(center + float2(-xoffs, yoffs), 0, 0)).y;
|
---|
64 |
|
---|
65 | const float m1 = min(texelCenterConv.x, texelCenterConv.y);
|
---|
66 | const float m2 = min(texelCenterConv.z, texelCenterConv.w);
|
---|
67 |
|
---|
68 | const float convergence = min(m1, m2);*/
|
---|
69 |
|
---|
70 | //const float convergence = tex2Dlod(tex, float4(center, 0, 0)).y;
|
---|
71 | const float convergence = tex2Dlod(tex, float4(texCoord, 0, 0)).y;
|
---|
72 |
|
---|
73 | return convergence;
|
---|
74 | }
|
---|
75 |
|
---|
76 | /** This shader computes the reprojection and stores
|
---|
77 | the ssao value of the old pixel as well as the
|
---|
78 | weight of the pixel in the new frame.
|
---|
79 | */
|
---|
80 | inline float3 Reproject(float4 worldPos,
|
---|
81 | float eyeSpaceDepth,
|
---|
82 | float2 texcoord0,
|
---|
83 | float3 oldEyePos,
|
---|
84 | sampler2D oldTex,
|
---|
85 | float4x4 oldModelViewProj,
|
---|
86 | sampler2D colors,
|
---|
87 | float3 projPos,
|
---|
88 | float invW,
|
---|
89 | float3 oldbl,
|
---|
90 | float3 oldbr,
|
---|
91 | float3 oldtl,
|
---|
92 | float3 oldtr,
|
---|
93 | float3 diffVec
|
---|
94 | )
|
---|
95 | {
|
---|
96 | // compute position from old frame for dynamic objects + translational portion
|
---|
97 | const float3 translatedPos = diffVec - oldEyePos + worldPos.xyz;
|
---|
98 |
|
---|
99 |
|
---|
100 | /////////////////
|
---|
101 | //-- reproject into old frame and calculate texture position of sample in old frame
|
---|
102 |
|
---|
103 | // note: the old model view matrix only holds the view orientation part
|
---|
104 | float4 backProjPos = mul(oldModelViewProj, float4(translatedPos, 1.0f));
|
---|
105 | backProjPos /= backProjPos.w;
|
---|
106 |
|
---|
107 | // fit from unit cube into 0 .. 1
|
---|
108 | const float2 oldTexCoords = backProjPos.xy * 0.5f + 0.5f;
|
---|
109 | // retrieve the sample from the last frame
|
---|
110 | const float4 oldPixel = tex2Dlod(oldTex, float4(oldTexCoords, .0f, .0f));
|
---|
111 |
|
---|
112 | // the ssao value in the old frame
|
---|
113 | const float ssao = oldPixel.x;
|
---|
114 |
|
---|
115 | // calculate eye space position of sample in old frame
|
---|
116 | const float oldEyeSpaceDepth = oldPixel.w;
|
---|
117 |
|
---|
118 | // vector from eye pos to old sample
|
---|
119 | const float3 viewVec = Interpol(oldTexCoords, oldbl, oldbr, oldtl, oldtr);
|
---|
120 | const float invLen = 1.0f / length(viewVec);
|
---|
121 | const float projectedEyeSpaceDepth = invLen * length(translatedPos);
|
---|
122 | //const float projectedEyeSpaceDepth = length(translatedPos);
|
---|
123 |
|
---|
124 | const float depthDif = abs(1.0f - oldEyeSpaceDepth / projectedEyeSpaceDepth);
|
---|
125 |
|
---|
126 | // the weight of the accumulated samples from the previous frames
|
---|
127 | float w;
|
---|
128 | float idx;
|
---|
129 |
|
---|
130 |
|
---|
131 | //////////////
|
---|
132 | //-- reuse old value only if it was still valid in the old frame
|
---|
133 |
|
---|
134 | if (1
|
---|
135 | && (oldTexCoords.x > 0) && (oldTexCoords.x < 1.0f)
|
---|
136 | && (oldTexCoords.y > 0) && (oldTexCoords.y < 1.0f)
|
---|
137 | && (depthDif <= MIN_DEPTH_DIFF)
|
---|
138 | )
|
---|
139 | {
|
---|
140 | // pixel valid => retrieve the convergence weight
|
---|
141 | /*float w1 = tex2Dlod(oldTex, float4(oldTexCoords + float2(0.5f / 1024.0f, 0), .0f, .0f)).y;
|
---|
142 | float w2 = tex2Dlod(oldTex, float4(oldTexCoords - float2(0.5f / 1024.0f, 0), .0f, .0f)).y;
|
---|
143 | float w3 = tex2Dlod(oldTex, float4(oldTexCoords + float2(0, 0.5f / 768.0f), .0f, .0f)).y;
|
---|
144 | float w4 = tex2Dlod(oldTex, float4(oldTexCoords - float2(0, 0.5f / 768.0f), .0f, .0f)).y;
|
---|
145 |
|
---|
146 | w = min(min(w1, w2), min(w3, w4));*/
|
---|
147 |
|
---|
148 | //w = ComputeConvergence(oldTex, oldTexCoords, float2(1024.0f, 768.0f));
|
---|
149 | w = oldPixel.y;
|
---|
150 | idx = floor(oldPixel.z);
|
---|
151 | //idx = oldPixel.z;
|
---|
152 | }
|
---|
153 | else
|
---|
154 | {
|
---|
155 | w = .0f;
|
---|
156 | idx = .0f;
|
---|
157 | }
|
---|
158 |
|
---|
159 | return float3(ssao, w, idx);
|
---|
160 | }
|
---|
161 |
|
---|
162 |
|
---|
163 | /** The ssao shader returning the an intensity value between 0 and 1.
|
---|
164 | This version of the ssao shader uses the dotproduct between
|
---|
165 | pixel-to-sample direction and sample normal as weight.
|
---|
166 |
|
---|
167 | The algorithm works like the following:
|
---|
168 | 1) Check in a circular area around the current position.
|
---|
169 | 2) Shoot vectors to the positions there, and check the angle to these positions.
|
---|
170 | 3) Summing up these angles gives an estimation of the occlusion at the current position.
|
---|
171 | */
|
---|
172 | float3 ssao2(fragment IN,
|
---|
173 | sampler2D colors,
|
---|
174 | sampler2D noiseTex,
|
---|
175 | sampler2D samples,
|
---|
176 | float3 normal,
|
---|
177 | float3 centerPosition,
|
---|
178 | float radius,
|
---|
179 | float3 bl,
|
---|
180 | float3 br,
|
---|
181 | float3 tl,
|
---|
182 | float3 tr,
|
---|
183 | float3 viewDir,
|
---|
184 | float convergence,
|
---|
185 | float sampleIntensity,
|
---|
186 | bool isMovingObject,
|
---|
187 | sampler2D normalTex,
|
---|
188 | float idx
|
---|
189 | )
|
---|
190 | {
|
---|
191 | float total_ao = .0f;
|
---|
192 | float validSamples = .0f;
|
---|
193 | float numSamples = .0f;
|
---|
194 |
|
---|
195 | for (int i = 0; i < NUM_SAMPLES; ++ i)
|
---|
196 | {
|
---|
197 | float2 offset;
|
---|
198 |
|
---|
199 | const float2 ssaoOffset =
|
---|
200 | tex2Dlod(samples, float4((0.5f + i + idx) / NUM_PRECOMPUTED_SAMPLES, 0.5f, .0f, .0f)).xy;
|
---|
201 |
|
---|
202 | ////////////////////
|
---|
203 | //-- add random noise: reflect around random normal vector
|
---|
204 | //-- (affects performance for some reason!)
|
---|
205 |
|
---|
206 | if (!USE_OPTIMIZATION ||
|
---|
207 | (convergence < SSAO_CONVERGENCE_THRESHOLD))
|
---|
208 | {
|
---|
209 | float2 mynoise = tex2Dlod(noiseTex, float4(IN.texCoord * 4.0f, 0, 0)).xy;
|
---|
210 | //offset = myreflect(samples[i], mynoise);
|
---|
211 | //offset = myrotate(samples[i], mynoise.x);
|
---|
212 | offset = myrotate(ssaoOffset, mynoise.x);
|
---|
213 | }
|
---|
214 | else
|
---|
215 | {
|
---|
216 | offset = ssaoOffset;
|
---|
217 | }
|
---|
218 |
|
---|
219 | // weight with projected coordinate to reach similar kernel size for near and far
|
---|
220 | const float2 texcoord = IN.texCoord.xy + offset * radius;
|
---|
221 |
|
---|
222 | const float4 sampleColor = tex2Dlod(colors, float4(texcoord, .0f, .0f));
|
---|
223 | const float3 samplePos = ReconstructSamplePos(sampleColor.w, texcoord, bl, br, tl, tr);
|
---|
224 |
|
---|
225 |
|
---|
226 | ////////////////
|
---|
227 | //-- compute contribution of sample using the direction and angle
|
---|
228 |
|
---|
229 | float3 dirSample = samplePos - centerPosition;
|
---|
230 |
|
---|
231 | const float minDist = 1e-6f;
|
---|
232 | const float delta = 1e-3f;
|
---|
233 |
|
---|
234 | const float lengthToSample = length(dirSample);
|
---|
235 | const float sampleWeight = 1.0f / (lengthToSample + delta);
|
---|
236 |
|
---|
237 | dirSample /= max(lengthToSample, minDist); // normalize
|
---|
238 |
|
---|
239 |
|
---|
240 | // angle between current normal and direction to sample controls AO intensity.
|
---|
241 | const float cosAngle = dot(dirSample, normal);
|
---|
242 |
|
---|
243 | // the normal of the current sample
|
---|
244 | const float3 sampleNormal = normalize(tex2Dlod(normalTex, float4(texcoord, 0, 0)).xyz);
|
---|
245 |
|
---|
246 | // angle between current normal and direction to sample controls AO intensity.
|
---|
247 | //const float cosAngle2 = dot(-dirSample, sampleNormal);
|
---|
248 | const float cosAngle2 = .5f + dot(sampleNormal, -normal) * .5f;
|
---|
249 |
|
---|
250 | dirSample *= minDist;
|
---|
251 | const float aoContrib = sampleIntensity * sampleWeight;
|
---|
252 |
|
---|
253 | //const float aoContrib = (1.0f > lengthToSample) ? occlusionPower(9e-2f, DISTANCE_SCALE + lengthToSample): .0f;
|
---|
254 | //total_ao += max(cosAngle, .0f) * max(cosAngle2, .0f) * aoContrib;
|
---|
255 | total_ao += max(cosAngle, .0f) * cosAngle2 * aoContrib;
|
---|
256 |
|
---|
257 | ++ numSamples;
|
---|
258 |
|
---|
259 | // check if the samples have been valid in the last frame
|
---|
260 | // only mark sample as invalid if in the last / current frame
|
---|
261 | // they possibly have any influence on the ao
|
---|
262 |
|
---|
263 | const float changeFactor = sampleColor.y;
|
---|
264 | const float pixelValid = sampleColor.x;
|
---|
265 |
|
---|
266 | // hack:
|
---|
267 | // we check if the sample could have been near enough
|
---|
268 | // to the current pixel or if the angle is small enough
|
---|
269 | // to have any influence in the current or last frame
|
---|
270 | #if 1
|
---|
271 | const float tooFarAway = step(0.5f, lengthToSample - changeFactor);
|
---|
272 | const float partlyResetThres = 1.0f;
|
---|
273 |
|
---|
274 | if (pixelValid <= partlyResetThres)
|
---|
275 | validSamples = max(validSamples, pixelValid * (1.0f - tooFarAway) * step(-0.1f, cosAngle));
|
---|
276 | else
|
---|
277 | validSamples = max(validSamples, pixelValid);
|
---|
278 | #endif
|
---|
279 |
|
---|
280 | #ifdef USE_GTX
|
---|
281 | // we can bail out early and use a minimal #samples)
|
---|
282 | // if some conditions are met as long as the hardware supports it
|
---|
283 | if (numSamples >= MIN_SAMPLES)
|
---|
284 | {
|
---|
285 | //break;
|
---|
286 | // if the pixel belongs to a static object and all the samples stay valid in the current frame
|
---|
287 | if (!isMovingObject && (validSamples < 1.0f) && (convergence > NUM_SAMPLES)) break;
|
---|
288 | // if the pixel belongs to a dynamic object but the #accumulated samples for this pixel is sufficiently high
|
---|
289 | // (=> there was no discontinuity recently)
|
---|
290 | //else if (isMovingObject && (convergence > SSAO_CONVERGENCE_THRESHOLD)) break;
|
---|
291 | else if (isMovingObject && (convergence > NUM_SAMPLES * 5)) break;
|
---|
292 | }
|
---|
293 | #endif
|
---|
294 | }
|
---|
295 |
|
---|
296 | // "normalize" ao contribution
|
---|
297 | total_ao /= numSamples;
|
---|
298 |
|
---|
299 | #if 1
|
---|
300 | // if surface normal perpenticular to view dir, approx. half of the samples will not count
|
---|
301 | // => compensate for this (on the other hand, projected sampling area could be larger!)
|
---|
302 | const float viewCorrection = 1.0f + VIEW_CORRECTION_SCALE * max(dot(viewDir, normal), 0.0f);
|
---|
303 | total_ao *= viewCorrection;
|
---|
304 | #endif
|
---|
305 |
|
---|
306 | //return float3(total_ao, validSamples, numSamples);
|
---|
307 | return float3(min(1.0f, total_ao), validSamples, numSamples);
|
---|
308 | }
|
---|
309 |
|
---|
310 | //#define TRYOUT
|
---|
311 |
|
---|
312 | #ifdef TRYOUT
|
---|
313 |
|
---|
314 | /** The ssao shader returning the an intensity value between 0 and 1.
|
---|
315 | This version of the ssao shader uses the dotproduct between
|
---|
316 | pixel-to-sample direction and sample normal as weight.
|
---|
317 |
|
---|
318 | The algorithm works like the following:
|
---|
319 | 1) Check in a circular area around the current position.
|
---|
320 | 2) Shoot vectors to the positions there, and check the angle to these positions.
|
---|
321 | 3) Summing up these angles gives an estimation of the occlusion at the current position.
|
---|
322 | */
|
---|
323 | float3 ssao(fragment IN,
|
---|
324 | sampler2D colors,
|
---|
325 | sampler2D noiseTex,
|
---|
326 | sampler2D samples,
|
---|
327 | float3 normal,
|
---|
328 | float3 centerPosition,
|
---|
329 | float radius,
|
---|
330 | float3 bl,
|
---|
331 | float3 br,
|
---|
332 | float3 tl,
|
---|
333 | float3 tr,
|
---|
334 | float3 viewDir,
|
---|
335 | float convergence,
|
---|
336 | float sampleIntensity,
|
---|
337 | bool isMovingObject,
|
---|
338 | float oldIdx
|
---|
339 | )
|
---|
340 | {
|
---|
341 | float total_ao = .0f;
|
---|
342 | float validSamples = .0f;
|
---|
343 | float numSamples = .0f;
|
---|
344 |
|
---|
345 | for (int i = 0; i < NUM_SAMPLES; ++ i)
|
---|
346 | {
|
---|
347 | float2 offset;
|
---|
348 |
|
---|
349 | const float2 ssaoOffset =
|
---|
350 | tex2Dlod(samples, float4((0.5f + i + oldIdx) / NUM_PRECOMPUTED_SAMPLES, 0.5f, .0f, .0f)).xy;
|
---|
351 |
|
---|
352 |
|
---|
353 | ////////////////////
|
---|
354 | //-- add random noise: reflect around random normal vector
|
---|
355 | //-- (affects performance for some reason!)
|
---|
356 |
|
---|
357 | if (!USE_OPTIMIZATION ||
|
---|
358 | (convergence < SSAO_CONVERGENCE_THRESHOLD))
|
---|
359 | {
|
---|
360 | float2 mynoise = tex2Dlod(noiseTex, float4(IN.texCoord * 4.0f, 0, 0)).xy;
|
---|
361 | //offset = myreflect(samples[i], mynoise);
|
---|
362 | //offset = myrotate(samples[i], mynoise.x);
|
---|
363 | offset = myrotate(ssaoOffset, mynoise.x);
|
---|
364 | }
|
---|
365 | else
|
---|
366 | {
|
---|
367 | offset = ssaoOffset;
|
---|
368 | }
|
---|
369 |
|
---|
370 |
|
---|
371 | // weight with projected coordinate to reach similar kernel size for near and far
|
---|
372 | const float2 texcoord = IN.texCoord.xy + offset * radius;
|
---|
373 |
|
---|
374 | const float4 sampleColor = tex2Dlod(colors, float4(texcoord, .0f, .0f));
|
---|
375 | const float3 samplePos = ReconstructSamplePos(sampleColor.w, texcoord, bl, br, tl, tr);
|
---|
376 |
|
---|
377 |
|
---|
378 | ////////////////
|
---|
379 | //-- compute contribution of sample using the direction and angle
|
---|
380 |
|
---|
381 | float3 dirSample = samplePos - centerPosition;
|
---|
382 |
|
---|
383 | const float minDist = 1e-6f;
|
---|
384 | const float eps = 1e-3f;
|
---|
385 |
|
---|
386 | const float lengthToSample = length(dirSample);
|
---|
387 | const float sampleWeight = 1.0f / max(lengthToSample, eps);
|
---|
388 |
|
---|
389 | dirSample /= max(length(dirSample), minDist); // normalize
|
---|
390 |
|
---|
391 | // angle between current normal and direction to sample controls AO intensity.
|
---|
392 | const float cosAngle = dot(dirSample, normal);
|
---|
393 |
|
---|
394 | //const float aoContrib = sampleIntensity / sqrLen;
|
---|
395 | const float aoContrib = sampleIntensity * sampleWeight;
|
---|
396 | //const float aoContrib = (1.0f > lengthToSample) ? occlusionPower(9e-2f, DISTANCE_SCALE + lengthToSample): .0f;
|
---|
397 |
|
---|
398 | total_ao += max(cosAngle, .0f) * aoContrib;
|
---|
399 |
|
---|
400 | ++ numSamples;
|
---|
401 |
|
---|
402 | #ifdef PERFORMANCE_TEST
|
---|
403 | // check if the samples have been valid in the last frame
|
---|
404 | // only mark sample as invalid if in the last / current frame
|
---|
405 | // they possibly have any influence on the AO
|
---|
406 |
|
---|
407 | const float changeFactor = sampleColor.y;
|
---|
408 | const float pixelValid = sampleColor.x;
|
---|
409 |
|
---|
410 | // hack:
|
---|
411 | // we check if the sample could have been near enough to the current pixel
|
---|
412 | // or if the angle is small enough
|
---|
413 | // to have any influence in the current or last frame
|
---|
414 |
|
---|
415 | #if 1
|
---|
416 | const float partlyResetThres = 1.0f;
|
---|
417 |
|
---|
418 | const float tooFarAway = step(.5f, lengthToSample - changeFactor);
|
---|
419 | if (0)//pixelValid <= partlyResetThres)
|
---|
420 | validSamples = max(validSamples, pixelValid * (1.0f - tooFarAway) * step(-0.1f, cosAngle));
|
---|
421 | else
|
---|
422 | validSamples = max(validSamples, pixelValid);
|
---|
423 | #endif
|
---|
424 |
|
---|
425 | #ifdef USE_GTX
|
---|
426 | // we can bail out early and use a minimal #samples)
|
---|
427 | // if some conditions are met as long as the hardware supports it
|
---|
428 | if (numSamples >= MIN_SAMPLES)
|
---|
429 | {
|
---|
430 | //break;
|
---|
431 | // if the pixel belongs to a static object and all the samples stay valid in the current frame
|
---|
432 | if (!isMovingObject && (validSamples < 1.0f) && (convergence > NUM_SAMPLES)) break;
|
---|
433 | // if the pixel belongs to a dynamic object but the #accumulated samples for this pixel is sufficiently high
|
---|
434 | // (=> there was no discontinuity recently)
|
---|
435 | //else if (isMovingObject && (convergence > SSAO_CONVERGENCE_THRESHOLD)) break;
|
---|
436 | else if (isMovingObject && (convergence > NUM_SAMPLES * 5)) break;
|
---|
437 | }
|
---|
438 | #endif
|
---|
439 |
|
---|
440 | #endif // PERFORMANCE_TEST
|
---|
441 | }
|
---|
442 |
|
---|
443 | // "normalize" ao contribution
|
---|
444 | total_ao /= numSamples;
|
---|
445 |
|
---|
446 | #if 1
|
---|
447 | // if surface normal perpenticular to view dir, approx. half of the samples will not count
|
---|
448 | // => compensate for this (on the other hand, projected sampling area could be larger!)
|
---|
449 | const float viewCorrection = 1.0f + VIEW_CORRECTION_SCALE * max(dot(viewDir, normal), 0.0f);
|
---|
450 | total_ao *= viewCorrection;
|
---|
451 | #endif
|
---|
452 |
|
---|
453 | //return float3(total_ao, validSamples, numSamples);
|
---|
454 | return float3(min(1.0f, total_ao), validSamples, numSamples);
|
---|
455 | }
|
---|
456 |
|
---|
457 | #else
|
---|
458 |
|
---|
459 | float3 ssao(fragment IN,
|
---|
460 | sampler2D colors,
|
---|
461 | sampler2D noiseTex,
|
---|
462 | sampler2D samples,
|
---|
463 | float3 normal,
|
---|
464 | float3 centerPosition,
|
---|
465 | float radius,
|
---|
466 | float3 bl,
|
---|
467 | float3 br,
|
---|
468 | float3 tl,
|
---|
469 | float3 tr,
|
---|
470 | float3 viewDir,
|
---|
471 | float convergence,
|
---|
472 | float sampleIntensity,
|
---|
473 | bool isMovingObject,
|
---|
474 | float oldIdx,
|
---|
475 | sampler2D attribsTex,
|
---|
476 | float3 oldPos
|
---|
477 | )
|
---|
478 | {
|
---|
479 | float total_ao = .0f;
|
---|
480 | float validSamples = .0f;
|
---|
481 | float numSamples = .0f;
|
---|
482 |
|
---|
483 | //float3 diffVec = tex2Dlod(attribsTex, float4(IN.texCoord, 0, 0)).xyz;
|
---|
484 |
|
---|
485 | for (int i = 0; i < NUM_SAMPLES; ++ i)
|
---|
486 | {
|
---|
487 | float2 offset;
|
---|
488 |
|
---|
489 | const float2 ssaoOffset =
|
---|
490 | tex2Dlod(samples, float4((0.5f + i + oldIdx) / NUM_PRECOMPUTED_SAMPLES, .5f, .0f, .0f)).xy;
|
---|
491 |
|
---|
492 |
|
---|
493 | ////////////////////
|
---|
494 | //-- add random noise: reflect around random normal vector
|
---|
495 | //-- (affects performance for some reason!)
|
---|
496 |
|
---|
497 | if (!USE_OPTIMIZATION ||
|
---|
498 | (convergence < SSAO_CONVERGENCE_THRESHOLD))
|
---|
499 | {
|
---|
500 | float2 mynoise = tex2Dlod(noiseTex, float4(IN.texCoord * 4.0f, 0, 0)).xy;
|
---|
501 | //offset = myreflect(samples[i], mynoise);
|
---|
502 | //offset = myrotate(samples[i], mynoise.x);
|
---|
503 | offset = myrotate(ssaoOffset, mynoise.x);
|
---|
504 | }
|
---|
505 | else
|
---|
506 | {
|
---|
507 | offset = ssaoOffset;
|
---|
508 | }
|
---|
509 |
|
---|
510 |
|
---|
511 | // weight with projected coordinate to reach similar kernel size for near and far
|
---|
512 | const float2 texcoord = IN.texCoord.xy + offset * radius;
|
---|
513 |
|
---|
514 | const float4 sampleColor = tex2Dlod(colors, float4(texcoord, .0f, .0f));
|
---|
515 | const float3 oldSamplePos = sampleColor.xyz;
|
---|
516 | const float3 samplePos = ReconstructSamplePos(sampleColor.w, texcoord, bl, br, tl, tr);
|
---|
517 |
|
---|
518 |
|
---|
519 |
|
---|
520 | ////////////////
|
---|
521 | //-- compute contribution of sample using the direction and angle
|
---|
522 |
|
---|
523 | float3 dirSample = samplePos - centerPosition;
|
---|
524 |
|
---|
525 | const float minDist = 1e-6f;
|
---|
526 | const float eps = 1e-3f;
|
---|
527 |
|
---|
528 | const float lengthToSample = length(dirSample);
|
---|
529 | const float sampleWeight = 1.0f / max(lengthToSample, eps);
|
---|
530 |
|
---|
531 | dirSample /= max(length(dirSample), minDist); // normalize
|
---|
532 |
|
---|
533 | // angle between current normal and direction to sample controls AO intensity.
|
---|
534 | const float cosAngle = dot(dirSample, normal);
|
---|
535 |
|
---|
536 | //const float aoContrib = sampleIntensity / sqrLen;
|
---|
537 | const float aoContrib = sampleIntensity * sampleWeight;
|
---|
538 | //const float aoContrib = (1.0f > lengthToSample) ? occlusionPower(9e-2f, DISTANCE_SCALE + lengthToSample): .0f;
|
---|
539 |
|
---|
540 | total_ao += max(cosAngle, .0f) * aoContrib;
|
---|
541 | ++ numSamples;
|
---|
542 |
|
---|
543 |
|
---|
544 | //#ifdef PERFORMANCE_TEST
|
---|
545 | #if 1
|
---|
546 | // check if the samples have been valid in the last frame
|
---|
547 | // only mark sample as invalid if in the last / current frame
|
---|
548 | // they possibly have any influence on the ao
|
---|
549 |
|
---|
550 | const float3 oldVector = oldSamplePos - oldPos;
|
---|
551 | const float oldDistance = length(oldVector);
|
---|
552 | const float cosAngle2 = dot(oldVector, normal);
|
---|
553 |
|
---|
554 | if ((cosAngle >= 0) || (cosAngle2 >= 0))
|
---|
555 | {
|
---|
556 | const float oldDistance = length(oldSamplePos - oldPos);
|
---|
557 | const float distanceDiff = abs(oldDistance - lengthToSample);
|
---|
558 |
|
---|
559 | float pixelValid = (distanceDiff > 1e-3f) ? 100.0f : 0;
|
---|
560 |
|
---|
561 | validSamples = max(validSamples, pixelValid);
|
---|
562 | }
|
---|
563 | #ifdef USE_GTX
|
---|
564 | // we can bail out early and use a minimal #samples)
|
---|
565 | // if some conditions are met as long as the hardware supports it
|
---|
566 | if (numSamples >= MIN_SAMPLES)
|
---|
567 | {
|
---|
568 | //break;
|
---|
569 | // if the pixel belongs to a static object and all the samples stay valid in the current frame
|
---|
570 | if (!isMovingObject && (validSamples < 1.0f) && (convergence > NUM_SAMPLES)) break;
|
---|
571 | // if the pixel belongs to a dynamic object but the #accumulated samples for this pixel is sufficiently high
|
---|
572 | // (=> there was no discontinuity recently)
|
---|
573 | //else if (isMovingObject && (convergence > SSAO_CONVERGENCE_THRESHOLD)) break;
|
---|
574 | else if (isMovingObject && (convergence > NUM_SAMPLES * 5)) break;
|
---|
575 | }
|
---|
576 | #endif
|
---|
577 |
|
---|
578 | #endif // PERFORMANCE_TEST
|
---|
579 | }
|
---|
580 |
|
---|
581 | // "normalize" ao contribution
|
---|
582 | total_ao /= numSamples;
|
---|
583 |
|
---|
584 | #if 1
|
---|
585 | // if surface normal perpenticular to view dir, approx. half of the samples will not count
|
---|
586 | // => compensate for this (on the other hand, projected sampling area could be larger!)
|
---|
587 | const float viewCorrection = 1.0f + VIEW_CORRECTION_SCALE * max(dot(viewDir, normal), 0.0f);
|
---|
588 | total_ao *= viewCorrection;
|
---|
589 | #endif
|
---|
590 |
|
---|
591 | //return float3(total_ao, validSamples, numSamples);
|
---|
592 | return float3(min(1.0f, total_ao), validSamples, numSamples);
|
---|
593 | }
|
---|
594 | #endif
|
---|
595 |
|
---|
596 |
|
---|
597 | /** The mrt shader for screen space ambient occlusion
|
---|
598 | */
|
---|
599 | pixel2 main(fragment IN,
|
---|
600 | uniform sampler2D colors,
|
---|
601 | uniform sampler2D normals,
|
---|
602 | uniform sampler2D noiseTex,
|
---|
603 | uniform sampler2D samples,
|
---|
604 | uniform sampler2D oldTex,
|
---|
605 | uniform float4x4 modelViewProj,
|
---|
606 | uniform float4x4 oldModelViewProj,
|
---|
607 | uniform float temporalCoherence,
|
---|
608 | uniform float3 bl,
|
---|
609 | uniform float3 br,
|
---|
610 | uniform float3 tl,
|
---|
611 | uniform float3 tr,
|
---|
612 | uniform float3 oldEyePos,
|
---|
613 | uniform float3 oldbl,
|
---|
614 | uniform float3 oldbr,
|
---|
615 | uniform float3 oldtl,
|
---|
616 | uniform float3 oldtr,
|
---|
617 | uniform sampler2D attribsTex,
|
---|
618 | uniform float kernelRadius,
|
---|
619 | uniform float sampleIntensity
|
---|
620 | ,uniform float dummyIdx
|
---|
621 | )
|
---|
622 | {
|
---|
623 | pixel2 OUT;
|
---|
624 |
|
---|
625 | //const float3 normal = normalize(tex2Dlod(normals, float4(IN.texCoord, 0 ,0)).xyz);
|
---|
626 | const float3 normal = tex2Dlod(normals, float4(IN.texCoord, 0 ,0)).xyz;
|
---|
627 |
|
---|
628 | // reconstruct position from the eye space depth
|
---|
629 | const float3 viewDir = IN.view;
|
---|
630 | const float4 mycolor = tex2Dlod(colors, float4(IN.texCoord, 0, 0));
|
---|
631 | const float eyeSpaceDepth = mycolor.w;
|
---|
632 | const float4 eyeSpacePos = float4(-viewDir * eyeSpaceDepth, 1.0f);
|
---|
633 |
|
---|
634 |
|
---|
635 | ////////////////
|
---|
636 | //-- calculcate the current projected posiion (also used for next frame)
|
---|
637 |
|
---|
638 | float4 projPos = mul(modelViewProj, eyeSpacePos);
|
---|
639 | const float invw = 1.0f / projPos.w;
|
---|
640 | projPos *= invw;
|
---|
641 |
|
---|
642 | //const float radiusMult = kernelRadius;
|
---|
643 | //const float radiusMult = 3e-2;
|
---|
644 | const float radiusMult = kernelRadius * invw;
|
---|
645 |
|
---|
646 | #ifdef PERFORMANCE_TEST
|
---|
647 |
|
---|
648 | float3 diffVec = tex2Dlod(attribsTex, float4(IN.texCoord, 0, 0)).xyz;
|
---|
649 |
|
---|
650 | const float sqrMoveSpeed = SqrLen(diffVec);
|
---|
651 | const bool isMovingObject = (sqrMoveSpeed > DYNAMIC_OBJECTS_THRESHOLD);
|
---|
652 |
|
---|
653 |
|
---|
654 | /////////////////
|
---|
655 | //-- compute temporal reprojection
|
---|
656 |
|
---|
657 | float3 temporalVals = Reproject(eyeSpacePos, eyeSpaceDepth, IN.texCoord, oldEyePos,
|
---|
658 | oldTex, oldModelViewProj,
|
---|
659 | colors,
|
---|
660 | projPos.xyz,
|
---|
661 | invw,
|
---|
662 | oldbl, oldbr, oldtl, oldtr,
|
---|
663 | diffVec
|
---|
664 | );
|
---|
665 |
|
---|
666 | const float oldSsao = temporalVals.x;
|
---|
667 |
|
---|
668 | float oldWeight = temporalVals.y;
|
---|
669 | float oldIdx = temporalCoherence > 1 ? temporalVals.z : 0;
|
---|
670 | //float oldIdx = temporalVals.z;
|
---|
671 |
|
---|
672 | #else
|
---|
673 |
|
---|
674 | const float3 diffVec = float3(.0f);
|
---|
675 | const bool isMovingObject = false;
|
---|
676 | const float oldSsao = .0f;
|
---|
677 |
|
---|
678 | float oldWeight = .0f;
|
---|
679 | float oldIdx = .0f;
|
---|
680 |
|
---|
681 | #endif
|
---|
682 |
|
---|
683 | float3 ao;
|
---|
684 |
|
---|
685 | // cull background note: this should be done with the stencil buffer
|
---|
686 | if (eyeSpaceDepth < DEPTH_THRESHOLD)
|
---|
687 | {
|
---|
688 | if (1)
|
---|
689 | {
|
---|
690 | #ifdef TRYOUT
|
---|
691 | ao = ssao(IN, colors, noiseTex, samples, normal, eyeSpacePos.xyz,
|
---|
692 | radiusMult, bl, br, tl, tr, normalize(viewDir),
|
---|
693 | oldWeight, sampleIntensity, isMovingObject, oldIdx);
|
---|
694 | #else
|
---|
695 |
|
---|
696 | ao = ssao(IN, colors, noiseTex, samples,
|
---|
697 | normal, eyeSpacePos.xyz, radiusMult, bl,
|
---|
698 | br, tl, tr, normalize(viewDir),
|
---|
699 | oldWeight, sampleIntensity, isMovingObject, oldIdx,
|
---|
700 | attribsTex, mycolor.xyz);
|
---|
701 | #endif
|
---|
702 | }
|
---|
703 | else
|
---|
704 | {
|
---|
705 | /*ao = ssao2(IN, colors, noiseTex, samples, normal, eyeSpacePos.xyz, radiusMult,
|
---|
706 | bl, br, tl, tr, normalize(viewDir), oldWeight, sampleIntensity,
|
---|
707 | isMovingObject, normals, oldIdx);
|
---|
708 | */
|
---|
709 | }
|
---|
710 | }
|
---|
711 | else
|
---|
712 | {
|
---|
713 | ao = float3(1.0f);
|
---|
714 | }
|
---|
715 |
|
---|
716 |
|
---|
717 | #ifdef PERFORMANCE_TEST
|
---|
718 |
|
---|
719 | ///////////
|
---|
720 | //-- check if we have to reset pixel because one of the sample points was invalid
|
---|
721 | //-- only do this if the current pixel does not belong to a moving object
|
---|
722 |
|
---|
723 | // the weight equals the number of sampled shot in this pass
|
---|
724 | const float newWeight = ao.z;
|
---|
725 | // completely reset the ao in this pixel
|
---|
726 | const float completelyResetThres = 20.0f;
|
---|
727 | // don't fully reset the ao in this pixel, but give low weight to old solution
|
---|
728 | const float partlyResetThres = 1.0f;
|
---|
729 |
|
---|
730 | // don't check for moving objects, otherwise almost no coherence
|
---|
731 | if (1)//!isMovingObject)
|
---|
732 | {
|
---|
733 | if (ao.y > completelyResetThres)
|
---|
734 | {
|
---|
735 | oldWeight = .0f;
|
---|
736 | oldIdx = .0f;
|
---|
737 | }
|
---|
738 | else if (ao.y > partlyResetThres)
|
---|
739 | {
|
---|
740 | const float factor = 4.0f;
|
---|
741 | //if (oldIdx >= factor * newWeight) oldIdx = 0;
|
---|
742 | //oldWeight = min(oldWeight, factor * newWeight);
|
---|
743 | oldWeight = oldIdx = .0f;
|
---|
744 | }
|
---|
745 | }
|
---|
746 |
|
---|
747 |
|
---|
748 | //////////
|
---|
749 | //-- blend ao between old and new samples (and avoid division by zero)
|
---|
750 |
|
---|
751 | OUT.illum_col.x = ao.x * newWeight + oldSsao * oldWeight;
|
---|
752 | OUT.illum_col.x /= (newWeight + oldWeight);
|
---|
753 | //OUT.illum_col.x = clamp(OUT.illum_col.x, 0, 1);
|
---|
754 | //OUT.illum_col.x = ao.y;
|
---|
755 |
|
---|
756 | // the new weight for the next frame
|
---|
757 | const float combinedWeight = clamp(newWeight + oldWeight, .0f, temporalCoherence);
|
---|
758 |
|
---|
759 | OUT.illum_col.y = combinedWeight;
|
---|
760 | OUT.illum_col.z = oldIdx + newWeight; // the new index
|
---|
761 | OUT.illum_col.w = eyeSpaceDepth;
|
---|
762 |
|
---|
763 | #else
|
---|
764 |
|
---|
765 | OUT.illum_col.x = ao.x;
|
---|
766 | OUT.illum_col.w = eyeSpaceDepth;
|
---|
767 |
|
---|
768 | #endif
|
---|
769 |
|
---|
770 | return OUT;
|
---|
771 | } |
---|