Skip to content

Commit cb528ef

Browse files
CrissplCrisspl
committed
rgb19e7 fixed (no more black/green flickering)
1 parent d3090af commit cb528ef

File tree

3 files changed

+12
-192
lines changed

3 files changed

+12
-192
lines changed

‎examples_tests/22.RaytracedAO/main.cpp

Lines changed: 1 addition & 186 deletions
Original file line numberDiff line numberDiff line change
@@ -18,192 +18,8 @@
1818
using namespace nbl;
1919
using namespace core;
2020

21-
//#define TEST_FORMAT
22-
23-
#ifdef TEST_FORMAT
24-
struct Test_RGB19E7
25-
{
26-
static constexpr const char* name = "rgb19e7";
27-
28-
static uint64_t acc_;
29-
static core::vectorSIMDf fetch()
30-
{
31-
auto rgb = core::rgb19e7_to_rgb32f(acc_);
32-
return core::vectorSIMDf(rgb.x, rgb.y, rgb.z);
33-
}
34-
static void store(const core::vectorSIMDf& x)
35-
{
36-
acc_ = core::rgb32f_to_rgb19e7(x.pointer);
37-
}
38-
};
39-
uint64_t Test_RGB19E7::acc_ = core::rgb32f_to_rgb19e7(0.f, 0.f, 0.f);
40-
struct Test_Expected
41-
{
42-
static constexpr const char* name = "expected";
43-
44-
static core::vectorSIMDf acc_;
45-
static core::vectorSIMDf fetch()
46-
{
47-
return acc_;
48-
}
49-
static void store(const core::vectorSIMDf& x)
50-
{
51-
acc_ = x;
52-
}
53-
};
54-
core::vectorSIMDf Test_Expected::acc_ = core::vectorSIMDf(0.f);
55-
void print(const core::vectorSIMDf v)
56-
{
57-
std::cout << v.x << ", " << v.y << ", " << v.z << std::endl;
58-
}
59-
60-
// raygen=0, resolve=1
61-
template <typename Test>
62-
auto test_()
63-
{
64-
const core::vectorSIMDf em(0.f, 0.5f, 0.25f);
65-
const core::vectorSIMDf dir(0.0f);
66-
for (int i = 0; i <= 30; ++i)
67-
{
68-
std::cout << "i = " << i << std::endl;
69-
float rcp = 1.f / float(i);
70-
if (rcp < 1.f)
71-
{
72-
auto acc = Test::fetch();
73-
acc += core::vectorSIMDf(em / float(i - 1));
74-
Test::store(acc);
75-
}
76-
else Test::store(em);
77-
//std::cout << "step 1: "; print(Test::fetch());
78-
79-
if (rcp < 1.f)
80-
{
81-
auto acc = Test::fetch();
82-
acc += (dir - acc) * rcp;
83-
Test::store(acc);
84-
}
85-
else
86-
{
87-
auto acc = Test::fetch();
88-
acc += dir;
89-
Test::store(acc);
90-
}
91-
std::cout << "step 2: "; print(Test::fetch());
92-
}
93-
94-
return Test::fetch();
95-
}
96-
template <typename T>
97-
auto test()
98-
{
99-
const auto result = test_<T>();
100-
std::cout << std::left << std::setw(15) << T::name << " result = " << result.x << ", " << result.y << ", " << result.z << ";\n";
101-
return result;
102-
}
103-
10421
int main()
10522
{
106-
test<Test_RGB19E7>();
107-
/*auto error = core::abs(test<Test_Expected>() - test<Test_RGB19E7>());
108-
std::cout << std::setprecision(15) << error.x << ", " << error.y << ", " << error.z << "\n";*/
109-
return 0;
110-
}
111-
112-
#else
113-
114-
struct uvec2
115-
{
116-
uvec2() : x{ 0,0 } {}
117-
uvec2(uint32_t xx, uint32_t yy) : x{ xx,yy } {}
118-
uvec2(uint64_t xx) : u64(xx) {}
119-
120-
operator uint64_t() const { return *reinterpret_cast<const uint64_t*>(x); }
121-
122-
uint32_t& operator[](int a) { return x[a]; }
123-
private:
124-
union {
125-
uint32_t x[2];
126-
uint64_t u64;
127-
};
128-
};
129-
130-
core::rgb32f decode(uvec2 x)
131-
{
132-
int _exp = int((bitfieldExtract(x[1], 25, 7) - 63u) - 19u);
133-
float scale = exp2(float(_exp));
134-
rgb32f v;
135-
v.x = float(bitfieldExtract(x[0], 0, 19));
136-
v.y = float(bitfieldInsert(bitfieldExtract(x[0], 19, 13), bitfieldExtract(x[1], 0, 6), 13, 6));
137-
v.z = float(bitfieldExtract(x[1], 6, 19));
138-
v.x *= scale;
139-
v.y *= scale;
140-
v.z *= scale;
141-
return v;
142-
}
143-
uvec2 encode(rgb32f col)
144-
{
145-
rgb32f clamped;
146-
clamped.x = clamp(col.x, 0.f, 35184304979968.0f);
147-
clamped.y = clamp(col.y, 0.f, 35184304979968.0f);
148-
clamped.z = clamp(col.z, 0.f, 35184304979968.0f);
149-
float maxrgb = max(max(clamped.x, clamped.y), clamped.z);
150-
int f32_exp = ((floatBitsToInt(maxrgb) >> 23) & 255) - 127;
151-
int shared_exp = clamp(f32_exp, -64, 64) + 1;
152-
uint32_t mantissas[3];// = uvec3((clamped * exp2(float(19 - shared_exp))) + vec3(0.5));
153-
float scale = ::exp2(19 - shared_exp);
154-
uint32_t maxm = maxrgb*scale + 0.5f;
155-
bool need = (maxm == float(1u<<19));
156-
scale = need ? (scale*0.5f) : scale;
157-
shared_exp = need ? (shared_exp+1) : shared_exp;
158-
mantissas[0] = uint32_t((clamped.x * scale) + 0.5f);
159-
mantissas[1] = uint32_t((clamped.y * scale) + 0.5f);
160-
mantissas[2] = uint32_t((clamped.z * scale) + 0.5f);
161-
uvec2 encoded;
162-
encoded[0] = bitfieldInsert(mantissas[0], mantissas[1], 19, 13);
163-
encoded[1] = bitfieldInsert(mantissas[1] >> 13, mantissas[2], 6, 19) | ((shared_exp + 63) << 25);
164-
return encoded;
165-
}
166-
167-
168-
169-
int main()
170-
{
171-
float rgb32_[3]{ 0.f, 0.333333333f, 0.f };
172-
rgb32f rgb32; memcpy(&rgb32.x, rgb32_, 12);
173-
const uvec2 rgb19e7_frame109_input = encode(rgb32); //uvec2(2863136768, 2080374826);
174-
175-
const float rcp = 0.14286f;
176-
177-
/*uint32_t rgb_[2]{ 2863136768, 2080374826 }; // frame 108
178-
//uint32_t rgb_[2]{ 0, 2080374784 }; // frame 109 step 1
179-
uint64_t* rgb = reinterpret_cast<uint64_t*>(rgb_);
180-
auto _19e7 = core::rgb19e7_to_rgb32f(*rgb);
181-
float t = 1.f;
182-
t /= 6.f;
183-
_19e7.y += t;
184-
*rgb = core::rgb32f_to_rgb19e7(&_19e7.x); // step 1 store (end)
185-
// step 2 start
186-
_19e7 = core::rgb19e7_to_rgb32f(*rgb);
187-
_19e7.y += (0.f - _19e7.y) * rcp;
188-
*rgb = core::rgb32f_to_rgb19e7(&_19e7.x);
189-
_19e7 = core::rgb19e7_to_rgb32f(*rgb);
190-
printf("");
191-
*/
192-
uvec2 rgb19e7 = rgb19e7_frame109_input;
193-
auto acc = decode(rgb19e7);
194-
float t = 1.f;
195-
t /= 6.f;
196-
acc.y += t;
197-
rgb19e7 = encode(acc); // step 1 store (end)
198-
// step 2
199-
acc = decode(rgb19e7);
200-
acc.y += (0.f - acc.y) * rcp;
201-
rgb19e7 = encode(acc);
202-
// put into fbo
203-
rgb32f fbo = decode(rgb19e7);
204-
printf("");
205-
206-
20723
// create device with full flexibility over creation parameters
20824
// you can add more parameters if desired, check nbl::SIrrlichtCreationParameters
20925
nbl::SIrrlichtCreationParameters params;
@@ -497,5 +313,4 @@ int main()
497313
renderer = nullptr;
498314

499315
return 0;
500-
}
501-
#endif
316+
}

‎include/nbl/builtin/glsl/format/constants.glsl

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -8,9 +8,9 @@
88
#define nbl_glsl_RGB19E7_EXP_BIAS 63
99
#define nbl_glsl_MAX_RGB19E7_EXP (nbl_glsl_RGB19E7_EXP_BIAS+1)
1010

11-
#define MAX_RGB19E7_MANTISSA_VALUES (0x1<<nbl_glsl_RGB19E7_MANTISSA_BITS)
12-
#define MAX_RGB19E7_MANTISSA (MAX_RGB19E7_MANTISSA_VALUES-1)
13-
#define nbl_glsl_MAX_RGB19E7 float(MAX_RGB19E7_MANTISSA)/float(MAX_RGB19E7_MANTISSA_VALUES)*exp2(float(nbl_glsl_MAX_RGB19E7_EXP-nbl_glsl_RGB19E7_MANTISSA_BITS))
11+
#define nbl_glsl_MAX_RGB19E7_MANTISSA_VALUES (0x1<<nbl_glsl_RGB19E7_MANTISSA_BITS)
12+
#define nbl_glsl_MAX_RGB19E7_MANTISSA (nbl_glsl_MAX_RGB19E7_MANTISSA_VALUES-1)
13+
#define nbl_glsl_MAX_RGB19E7 float(nbl_glsl_MAX_RGB19E7_MANTISSA)/float(nbl_glsl_MAX_RGB19E7_MANTISSA_VALUES)*exp2(float(nbl_glsl_MAX_RGB19E7_EXP-nbl_glsl_RGB19E7_MANTISSA_BITS))
1414

1515
#define nbl_glsl_RGB19E7_COMPONENT_INDICES ivec4(0,0,1,1)
1616
#define nbl_glsl_RGB19E7_COMPONENT_BITOFFSETS ivec4(0,nbl_glsl_RGB19E7_MANTISSA_BITS,(2*nbl_glsl_RGB19E7_MANTISSA_BITS)&31,(3*nbl_glsl_RGB19E7_MANTISSA_BITS)&31)

‎include/nbl/builtin/glsl/format/encode.glsl

Lines changed: 8 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -9,9 +9,14 @@ uvec2 nbl_glsl_encodeRGB19E7(in vec3 col)
99
const float maxrgb = max(max(clamped.r,clamped.g),clamped.b);
1010

1111
const int f32_exp = ((floatBitsToInt(maxrgb)>>23) & 0xff) - 127;
12-
const int shared_exp = clamp(f32_exp,-nbl_glsl_RGB19E7_EXP_BIAS-1,nbl_glsl_MAX_RGB19E7_EXP) + 1;
13-
14-
const uvec3 mantissas = uvec3(clamped*exp2(nbl_glsl_RGB19E7_MANTISSA_BITS-shared_exp) + 0.5);
12+
int shared_exp = clamp(f32_exp,-nbl_glsl_RGB19E7_EXP_BIAS-1,nbl_glsl_MAX_RGB19E7_EXP) + 1;
13+
14+
float scale = exp2(nbl_glsl_RGB19E7_MANTISSA_BITS - shared_exp);
15+
const uint maxm = uint(maxrgb*scale + 0.5);
16+
const bool need = (maxm == nbl_glsl_MAX_RGB19E7_MANTISSA_VALUES);
17+
scale = need ? 0.5*scale : scale;
18+
shared_exp = need ? (shared_exp+1) : shared_exp;
19+
const uvec3 mantissas = uvec3(clamped*scale + 0.5);
1520

1621
uvec2 encoded;
1722
encoded.x = bitfieldInsert(mantissas.x,mantissas.y,nbl_glsl_RGB19E7_COMPONENT_BITOFFSETS[1],nbl_glsl_RGB19E7_G_COMPONENT_SPLIT);

0 commit comments

Comments
 (0)