This question already has answers here:
How to alpha blend RGBA unsigned byte color fast?
(17 answers)
Closed 8 years ago.
I have two colors and I use this method to do a simple alpha blending:
int Color::blend(int col1, int col2)
{
float a1 = ((col1 & 0x000000FF) / 255.0);
return ((int)((((col1 & 0xFF000000) >> 24) * a1) + (((col2 & 0xFF000000) >> 24) * (1.0 - a1)))) << 24 |
((int)((((col1 & 0x00FF0000) >> 16) * a1) + (((col2 & 0x00FF0000) >> 16) * (1.0 - a1)))) << 16 |
((int)((((col1 & 0x0000FF00) >> 8 ) * a1) + (((col2 & 0x0000FF00) >> 8 ) * (1.0 - a1)))) << 8 | 255;
}
(The colors are in RGBA8888 format)
This works, but i was wondering: is this the fastest way, or is there a more efficient one?
You might be able to eke out a little more performance by representing a1*(2^24) as an integer, doing the arithmetic in integers, then shifting the result down by 24 bits. On modern architectures I doubt it would gain you much, though. If you want better performance, you'll really need to go for SIMD operations.
Oh, one thing: You should express the calculation of a1 as a1 = ((col1 & 0x000000FF) * (1.0 / 255.0)). That'll avoid an expensive FP division. (Compilers won't usually do that on their own, due to the potential loss of precision.)
Related
I am trying to implement a function that blends two colors encoded with RGB565 using Alpha blending
Crgb565 = (1-a)Argb565 + a*Brgb565
Where a is the alpha parameter, and the alpha blending value of 0.0-1.0 is mapped to an unsigned char value on the range 0-32.
we can choose to use a five bit representation for a instead, thus restricting it to the range of 0-31 (effectively mapping to an alpha blending value of 0.0-0.96875).
Following code I am trying to implement, can you please suggest better way wrt less temp variable , memory optimization (number of multiplications and required memory accesses ),Is my logic for alpha bending is correct? I am not getting correct result/expected output, Seems like I am missing something, please review the code, Every suggest is appreciated, have some doubt based on alpha parameter. I have put my doubts in code comment section. Is there any way to shortening the alpha blending equations(division operation)?
=====================================================
unsigned short blend_rgb565(unsigned short A, unsigned short B, unsigned char Alpha)
{
unsigned short res = 0;
// Alpha converted from [0..255] to [0..31] (8 bit to 5 bit)
/* I want the alpha parameter (0-32), do i need to add something in Alpha before right shift?? */
Alpha = Alpha >> 3;
// Split Image A into R, G, B components
/*Do I need to take it as unsigned short or uint8_t also work fine ??*/
unsigned short A_r = A >> 11;
unsigned short A_g = (A >> 5) & ((1u << 6) - 1); // ((1u << 6) - 1) --> 00000000 00111111
unsigned short A_b = A & ((1u << 5) - 1); // ((1u << 5) - 1) --> 00000000 00011111
// Split Image B into R, G, B components
unsigned short B_r = B >> 11;
unsigned short B_g = (B >> 5) & ((1u << 6) - 1);
unsigned short B_b = B & ((1u << 5) - 1);
// Alpha blend components
/*Do I need to use 255(8 bit) instead of 32(5 bit), Why we are dividing by it , I have taken the ref from internet , but need little bit more clarification ??*/
unsigned short uiC_r = (A_r * Alpha + B_r * (32 - Alpha)) / 32;
unsigned short uiC_g = (A_g * Alpha + B_g * (32 - Alpha)) / 32;
unsigned short uiC_b = (A_b * Alpha + B_b * (32 - Alpha)) / 32;
// Pack result
res= (unsigned short) ((uiC_r << 11) | (uiC_g << 5) | uiC_b);
return res;
}
=====================
EDIT:
Adding method 2 ,is this approach is correct ?
Method 2:
// rrrrrggggggbbbbb
#define RB_MASK 63519 // 0b1111100000011111 --> hex :F81F
#define G_MASK 2016 // 0b0000011111100000 --> hex :07E0
#define RB_MUL_MASK 2032608 // 0b111110000001111100000 --> hex :1F03E0
#define G_MUL_MASK 64512 // 0b000001111110000000000 --> hex :FC00
unsigned short blend_rgb565(unsigned short A,unsigned short B,unsigned char Alpha) {
// Alpha converted from [0..255] to [0..31]
Alpha = Alpha >> 3
uint8_t beta = 32 - Alpha;
// so (0..32)*Alpha + (0..32)*beta always in 0..32
return (unsigned short)
(
(
( ( Alpha * (uint32_t)( A & RB_MASK ) + beta * (uint32_t)( B & RB_MASK )) & RB_MUL_MASK )
|
( ( Alpha * ( A & G_MASK ) + beta * ( B & G_MASK )) & G_MUL_MASK )
)
>> 5 // removing the alpha component 5 bit
);
}
It's possible to reduce the multiplies from 6 to 2 if you space out the RGB values into 2 32-bit integers before multiplying:
unsigned short blend_rgb565(unsigned short A, unsigned short B, unsigned char Alpha)
{
unsigned short res = 0;
// Alpha converted from [0..255] to [0..31] (8 bit to 5 bit)
Alpha = Alpha >> 3;
// Alpha = (Alpha + (Alpha >> 5)) >> 3; // map from 0-255 to 0-32 (if Alpha is unsigned short or larger)
// Space out A and B from RRRRRGGGGGGBBBBB to 00000RRRRR00000GGGGGG00000BBBBB
// 31 = 11111 binary
// 63 = 111111 binary
unsigned int A32 = (unsigned int)A;
unsigned int A_spaced = A32 & 31; // B
A_spaced |= (A32 & (63 << 5)) << 5; // G
A_spaced |= (A32 & (31 << 11)) << 11; // R
unsigned int B32 = (unsigned int)B;
unsigned int B_spaced = B32 & 31; // B
B_spaced |= (B32 & (63 << 5)) << 5; // G
B_spaced |= (B32 & (31 << 11)) << 11; // R
// multiply and add the alpha to give a result RRRRRrrrrrGGGGGGgggggBBBBBbbbbb,
// where RGB are the most significant bits we want to keep
unsigned int C_spaced = (A_spaced * Alpha) + (B_spaced * (32 - Alpha));
// remap back to RRRRRGGGGGBBBBB
res = (unsigned short)(((C_spaced >> 5) & 31) + ((C_spaced >> 10) & (63 << 5)) + ((C_spaced >> 16) & (31 << 11)));
return res;
}
You need to profile this to see if it is faster, it assumes that multiplications you save are slower than the extra bit-manipulations you replace them with.
can you please suggest better way wrt less temp variable
There is no advantage to remove temporary variables from the implementation. When you compile with optimizations turned on (e.g. -O2 or /O2) those temp variables will get optimized away.
Two adjustments I would make to your code:
Use uint16_t instead of unsigned short. For most platforms, it won't matter since sizeof(uint16_t)==sizeof(unsigned short), but it helps to be definitive.
No point in converting alpha from an 8-bit value to a 5-bit value. You'll get better accuracy with blending if you let alpha have the full range
Some of your bit-shifting looks weird. It might work. But I use a simpler approach.
Here's an adjustment to your implementation:
#include <stdint.h>
#define MAKE_RGB565(r, g, b) ((r << 11) | (g << 5) | (b))
uint16_t blend_rgb565(uint16_t a, uint16_t b, uint8_t Alpha)
{
const uint8_t invAlpha = 255 - Alpha;
uint16_t A_r = a >> 11;
uint16_t A_g = (a >> 5) & 0x3f;
uint16_t A_b = a & 0x1f;
uint16_t B_r = b >> 11;
uint16_t B_g = (b >> 5) & 0x3f;
uint16_t B_b = b & 0x1f;
uint32_t C_r = (A_r * invAlpha + B_r * Alpha) / 255;
uint32_t C_g = (A_g * invAlpha + B_g * Alpha) / 255;
uint32_t C_b = (A_b * invAlpha + B_b * Alpha) / 255;
return MAKE_RGB565(C_r, C_g, C_b);
}
But the bigger issue is that this function works on exactly one one pair of pixel colors. If you are invoking this function across an entire image or pair of images, the overhead of using the function call is going to be a major performance issue - even with compiler optimizations and inlining. So if you are calling this function row x col times, you should probably manually inline the code into your loop that is enumerating over every pixel on an image (or pair of images).
In the same vein as #samgak's answer, you can implement more efficiently on a 64 bits architecture by "post-masking", as follows:
rrrrrggggggbbbbb
Replicate to a long long (by shifting or mapping the long long to four shorts)
---------------- rrrrrggggggbbbbb rrrrrggggggbbbbb rrrrrggggggbbbbb
Mask out the useless bits
---------------- rrrrr----------- -----gggggg----- -----------bbbbb
Multiply by α
-----------rrrrr rrrrr----------- ggggggggggg----- ------bbbbbbbbbb
Mask out the low order bits
-----------rrrrr ---------------- gggggg---------- ------bbbbb-----
Pack
rrrrrgggggbbbbb
Another saving is possible by rewriting
(1 - α) X + α Y
as
X + α (Y - X)
(or X - α (X - Y) to avoid negatives). This spares a multiply (at the expense of a comparison).
Update:
The "saving" above cannot work because the negatives should be handled component-wise.
When you and some data with a mask you get some result which is of the same size as the data/mask.
What I want to do, is to take the masked bits in the result (where there was 1 in the mask) and shift them to the right so they are next to each other and I can perform a CTZ (Count Trailing Zeroes) on them.
I didn't know how to name such a procedure so Google has failed me. The operation should preferably not be a loop solution, this has to be as fast operation as possible.
And here is an incredible image made in MS Paint.
This operation is known as compress right. It is implemented as part of BMI2 as the PEXT instruction, in Intel processors as of Haswell.
Unfortunately, without hardware support is it a quite annoying operation. Of course there is an obvious solution, just moving the bits one by one in a loop, here is the one given by Hackers Delight:
unsigned compress(unsigned x, unsigned m) {
unsigned r, s, b; // Result, shift, mask bit.
r = 0;
s = 0;
do {
b = m & 1;
r = r | ((x & b) << s);
s = s + b;
x = x >> 1;
m = m >> 1;
} while (m != 0);
return r;
}
But there is an other way, also given by Hackers Delight, which does less looping (number of iteration logarithmic in the number of bits) but more per iteration:
unsigned compress(unsigned x, unsigned m) {
unsigned mk, mp, mv, t;
int i;
x = x & m; // Clear irrelevant bits.
mk = ~m << 1; // We will count 0's to right.
for (i = 0; i < 5; i++) {
mp = mk ^ (mk << 1); // Parallel prefix.
mp = mp ^ (mp << 2);
mp = mp ^ (mp << 4);
mp = mp ^ (mp << 8);
mp = mp ^ (mp << 16);
mv = mp & m; // Bits to move.
m = m ^ mv | (mv >> (1 << i)); // Compress m.
t = x & mv;
x = x ^ t | (t >> (1 << i)); // Compress x.
mk = mk & ~mp;
}
return x;
}
Notice that a lot of the values there depend only on m. Since you only have 512 different masks, you could precompute those and simplify the code to something like this (not tested)
unsigned compress(unsigned x, int maskindex) {
unsigned t;
int i;
x = x & masks[maskindex][0];
for (i = 0; i < 5; i++) {
t = x & masks[maskindex][i + 1];
x = x ^ t | (t >> (1 << i));
}
return x;
}
Of course all of these can be turned into "not a loop" by unrolling, the second and third ways are probably more suitable for that. That's a bit of cheat however.
You can use the pack-by-multiplication technique similar to the one described here. This way you don't need any loop and can mix the bits in any order.
For example with the mask 0b10101001 == 0xA9 like above and 8-bit data abcdefgh (with a-h is the 8 bits) you can use the below expression to get 0000aceh
uint8_t compress_maskA9(uint8_t x)
{
const uint8_t mask1 = 0xA9 & 0xF0;
const uint8_t mask2 = 0xA9 & 0x0F;
return (((x & mask1)*0x03000000 >> 28) & 0x0C) | ((x & mask2)*0x50000000 >> 30);
}
In this specific case there are some overlaps of the 4 bits while adding (which incur unexpected carry) during the multiplication step, so I've split them into 2 parts, the first one extracts bit a and c, then e and h will be extracted in the latter part. There are other ways to split the bits as well, like a & h then c & e. You can see the results compared to Harold's function live on ideone
An alternate way with only one multiplication
const uint32_t X = (x << 8) | x;
return (X & 0x8821)*0x12050000 >> 28;
I got this by duplicating the bits so that they're spaced out farther, leaving enough space to avoid the carry. This is often better than splitting into 2 multiplications
If you want the result's bits reversed (i.e. heca0000) you can easily change the magic numbers accordingly
// result: he00 | 00ca;
return (((x & 0x09)*0x88000000 >> 28) & 0x0C) | (((x & 0xA0)*0x04800000) >> 30);
or you can also extract the 3 bits e, c and a at the same time, leaving h separately (as I mentioned above, there are often multiple solutions) and you need only one multiplication
return ((x & 0xA8)*0x12400000 >> 29) | (x & 0x01) << 3; // result: 0eca | h000
But there might be a better alternative like the above second snippet
const uint32_t X = (x << 8) | x;
return (X & 0x2881)*0x80290000 >> 28
Correctness check: http://ideone.com/PYUkty
For a larger number of masks you can precompute the magic numbers correspond to those masks and store them in an array so that you can look them up immediately for use. I calculated those mask by hand but you can do that automatically
Explanation
We have abcdefgh & mask1 = a0c00000. Multiply it with magic1
........................a0c00000
× 00000011000000000000000000000000 (magic1 = 0x03000000)
────────────────────────────────
a0c00000........................
+ a0c00000......................... (the leading "a" bit is outside int's range
──────────────────────────────── so it'll be truncated)
r1 = acc.............................
=> (r1 >> 28) & 0x0C = 0000ac00
Similarly we multiply abcdefgh & mask2 = 0000e00h with magic2
........................0000e00h
× 01010000000000000000000000000000 (magic2 = 0x50000000)
────────────────────────────────
e00h............................
+ 0h..............................
────────────────────────────────
r2 = eh..............................
=> (r2 >> 30) = 000000eh
Combine them together we have the expected result
((r1 >> 28) & 0x0C) | (r2 >> 30) = 0000aceh
And here's the demo for the second snippet
abcdefghabcdefgh
& 1000100000100001 (0x8821)
────────────────────────────────
a000e00000c0000h
× 00010010000001010000000000000000 (0x12050000)
────────────────────────────────
000h
00e00000c0000h
+ 0c0000h
a000e00000c0000h
────────────────────────────────
= acehe0h0c0c00h0h
& 11110000000000000000000000000000
────────────────────────────────
= aceh
For the reversed order case:
abcdefghabcdefgh
& 0010100010000001 (0x2881)
────────────────────────────────
00c0e000a000000h
x 10000000001010010000000000000000 (0x80290000)
────────────────────────────────
000a000000h
00c0e000a000000h
+ 0e000a000000h
h
────────────────────────────────
hecaea00a0h0h00h
& 11110000000000000000000000000000
────────────────────────────────
= heca
Related:
How to create a byte out of 8 bool values (and vice versa)?
Redistribute least significant bits from a 4-byte array to a nibble
I'm working in C++ with an array of unsigned char representing pixels in an image. Each pixel has 3 channel (R,G,B). The image is represented linearly, sort of like
RGBRGBRGBRGB.....
How do I split each of the R,G and B, into separate arrays efficiently?
I tried:
for(int pos = 0; pos < srcWidth * srcHeight; pos++) {
int rgbPos = pos * 3;
splitChannels[0][pos] = rgbSrcData[rgbPos];
splitChannels[1][pos] = rgbSrcData[rgbPos + 1];
splitChannels[2][pos] = rgbSrcData[rgbPos + 2];
}
But this is surprisingly slow.
Thanks!
My attempt: load and store the bytes four by four. Byte scrambling will be tedious but possibly throughput will improve.
// Load 4 interleaved pixels
unsigned int RGB0= ((int*)rgbSrcData)[i];
unsigned int RGB1= ((int*)rgbSrcData)[i + 1];
unsigned int RGB2= ((int*)rgbSrcData)[i + 2];
// Rearrange and store 4 unpacked pixels
((int*)splitChannels[0])[j]=
(RGB0 & 0xFF) | (RGB0 >> 24) | (RGB1 & 0xFF0000) | ((RGB2 & 0xFF00) << 16);
((int*)splitChannels[1])[j]=
((RGB0 & 0xFF00) >> 8) | (RGB1 & 0xFF) | (RGB1 >> 24) | (RGB2 & 0xFF0000) >> 16;
((int*)splitChannels[2])[j]=
((RGB0 & 0xFF0000) >> 16) | (RGB1 & 0xFF00) | ((RGB2 & 0xFF) >> 16) | (RGB2 & 0xFF000000);
(CAUTION: not unchecked !) A shift-only version is also possible.
An SSE solution would be more complex (the stride 3 does not get along with powers of 2).
A great technique to use to make it run faster is loop unwinding.
You can read about it here: http://en.wikipedia.org/wiki/Loop_unwinding
I am wanting to write to a texture from my directX 11 compute shader. However I have no idea how to display this onto the screen nor am I sure what sort of buffer I should be using to do this.
welcome on stackoverflow :)
The type of resource to choose is RWTexture2D<float4> since you can print this directly on screen via a swapchain.
You can look at the DirectX SDK OIT sample:
They have declared a RWTexture2D<float4> frameBuffer that they access in the function SortAndRenderCS of OIT_CS.hlsl.
// convert the color to floats
float4 color[3];
color[0].r = (r0 >> 0 & 0xFF) / 255.0f;
color[0].g = (r0 >> 8 & 0xFF) / 255.0f;
color[0].b = (r0 >> 16 & 0xFF) / 255.0f;
color[0].a = (r0 >> 24 & 0xFF) / 255.0f;
color[1].r = (r1 >> 0 & 0xFF) / 255.0f;
color[1].g = (r1 >> 8 & 0xFF) / 255.0f;
color[1].b = (r1 >> 16 & 0xFF) / 255.0f;
color[1].a = (r1 >> 24 & 0xFF) / 255.0f;
color[2].r = (r2 >> 0 & 0xFF) / 255.0f;
color[2].g = (r2 >> 8 & 0xFF) / 255.0f;
color[2].b = (r2 >> 16 & 0xFF) / 255.0f;
color[2].a = (r2 >> 24 & 0xFF) / 255.0f;
float4 result = lerp(lerp(lerp(0, color[2], color[2].a), color[1], color[1].a), color[0], color[0].a);
result.a = 1.0f;
frameBuffer[nDTid.xy] = result;
As you can see they have r0, r1 and r2 uint values that are actually RGBA colors (a byte for each channel), they extract each channel using shifts and masks and normalized it.
You don't need to do that if you have already float4 values of course.
Then they do those lerps (for interpolation). Again you shouldn't need to do that.
What interest you is that they access frameBuffer using array notation and an uint2 for coordinates.
I use a population count (hamming weight) function intensively in a windows c application and have to optimize it as much as possible in order to boost performance. More than half the cases where I use the function I only need to know the value to a maximum of 15. The software will run on a wide range of processors, both old and new. I already make use of the POPCNT instruction when Intel's SSE4.2 or AMD's SSE4a is present, but would like to optimize the software implementation (used as a fall back if no SSE4 is present) as much as possible.
Currently I have the following software implementation of the function for 64bit (platform) mode:
int population_count64(unsigned __int64 w) {
w -= (w >> 1) & 0x5555555555555555ULL;
w = (w & 0x3333333333333333ULL) + ((w >> 2) & 0x3333333333333333ULL);
w = (w + (w >> 4)) & 0x0f0f0f0f0f0f0f0fULL;
return int((w * 0x0101010101010101ULL) >> 56);
}
So to summarize:
(1) I would like to know if it is possible to optimize this for the case when I only want to know the value to a maximum of 15.
(2) Is there a faster software implementation (for both Intel and AMD CPU's) than the function above (for unsigned 64bit integers)?
It is indeed possible to optimise your function for the "maximum 15" case. The following shaves off a few operations:
inline int population_count64_max15(unsigned __int64 w)
{
w -= (w >> 1) & 0x5555555555555555ULL;
w = (w & 0x3333333333333333ULL) + ((w >> 2) & 0x3333333333333333ULL);
return int((w * 0x1111111111111111ULL) >> 60);
}
Inlining the function (using the inline keyword as above) should also increase performance.
If you're on a 32-bit machine, split w into two 32-bit words, calculate the popcount separately for each half, then add up. This will get rid of some unneeded operations that are required to synthesize 64-bit operations from 32-bit ones (shifts, mults...). This also allows for increased parallelism if you interleave the calculations.
If you're compiling 64-bit code, you may try this:
int popcnt64(uint64_t w)
{
uint64_t w1 = (w & 0x2222222222222222) + ((w+w) & 0x2222222222222222);
uint64_t w2 = (w >> 1 & 0x2222222222222222) + (w >> 2 & 0x2222222222222222);
w1 = w1 + (w1 >> 4) & 0x0f0f0f0f0f0f0f0f;
w2 = w2 + (w2 >> 4) & 0x0f0f0f0f0f0f0f0f;
return (w1 + w2) * 0x0101010101010101 >> 57;
}
This contains more operations, but gives more opportunities of parallel execution to the CPU. On newer CPUs it should be slightly faster, on others it will be slightly slower.