Faster way for extracting and combining bits from UINT16 to UINT8

Question

I'm searching for a faster way for my required special extract and combine operation as described below:

+-------+-------+-------+-------+-------+-------+-------+-------+
| BIT 7 | BIT 6 | BIT 5 | BIT 4 | BIT 3 | BIT 2 | BIT 1 | BIT 0 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   D1  |  D0   |  C1   |  C0   |  B1   |  B0   |  A1   |   A0  |
+-------+-------+-------+-------+-------+-------+-------+-------+

A = A0 OR A1
B = B0 OR B1
C = C0 OR C1
D = D0 OR D1

+-------+-------+-------+-------+-------+-------+-------+-------+
| BIT 7 | BIT 6 | BIT 5 | BIT 4 | BIT 3 | BIT 2 | BIT 1 | BIT 0 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |   D   |   C   |   B   |   A   |
+-------+-------+-------+-------+-------+-------+-------+-------+

For sake of simplicity above is only an 8-bit example, the same applies for 16 bit values. It should be implemented as fast as possible on dsPIC33F microcontroller.

The easy way in C is:

PairFlags |= (ChannelFlags & 0x0003) ? 0x0001 : 0;
PairFlags |= (ChannelFlags & 0x000C) ? 0x0002 : 0;
PairFlags |= (ChannelFlags & 0x0030) ? 0x0004 : 0;
PairFlags |= (ChannelFlags & 0x00C0) ? 0x0008 : 0;
PairFlags |= (ChannelFlags & 0x0300) ? 0x0010 : 0;
PairFlags |= (ChannelFlags & 0x0C00) ? 0x0020 : 0;
PairFlags |= (ChannelFlags & 0x3000) ? 0x0040 : 0;
PairFlags |= (ChannelFlags & 0xC000) ? 0x0080 : 0;

This will produce approx. 40 instructions (with O3) which corresponds to 1µs in my case.

The amount of instruction cycles should be reduced if possible. Is there a faster way either in C or inline assembly?

Answer 1

The following should work for reducing a 16-bit value to 8 bits (with each bit of output formed by ORing a pair of bits of input):

// Set even bits to bits in pair ORed together, and odd bits to 0...
PairFlags = (ChannelFlags | (ChannelFlags >> 1)) & 0x5555; // '0h0g0f0e0d0c0b0a'
// Compress the '00' or '01' bit pairs down to single '0' or '1' bits...
PairFlags = (PairFlags ^ (PairFlags >> 1)) & 0x3333; // '00hg00fe00dc00ba'
PairFlags = (PairFlags ^ (PairFlags >> 2)) & 0x0F0F; // '0000hgfe0000dcba'
PairFlags = (PairFlags ^ (PairFlags >> 4)) & 0x00FF; // '00000000hgfedcba'

Note: The ^ can be replaced by | in the above for the same result.

Answer 2

Assuming I got everything right (not tested), this seems to generate good, branch-free code at least on gcc and clang for x86 (-O3):

uint8_t convert (uint8_t ChannelFlags)
{
  return ( ((ChannelFlags & A1A0)!=0) << A_POS ) |
         ( ((ChannelFlags & B1B0)!=0) << B_POS ) |
         ( ((ChannelFlags & C1C0)!=0) << C_POS ) |
         ( ((ChannelFlags & D1D0)!=0) << D_POS ) ;  
}

This masks out each individual bitset, then check against zero to end up with 1 or 0 in a temporary int. This value is shifted in position in the result, before everything is finally bitwise OR:ed together. Full code:

#include <stdint.h>

#define A1A0  (3u << 0)
#define B1B0  (3u << 2)
#define C1C0  (3u << 4)
#define D1D0  (3u << 6)

#define A_POS 0
#define B_POS 1
#define C_POS 2
#define D_POS 3

uint8_t convert (uint8_t ChannelFlags)
{
  return ( ((ChannelFlags & A1A0)!=0) << A_POS ) |
         ( ((ChannelFlags & B1B0)!=0) << B_POS ) |
         ( ((ChannelFlags & C1C0)!=0) << C_POS ) |
         ( ((ChannelFlags & D1D0)!=0) << D_POS ) ;  
}

clang disassembly x86 gives 18 instructions branch free:

convert:                                # @convert
        test    dil, 3
        setne   al
        test    dil, 12
        setne   cl
        add     cl, cl
        or      cl, al
        test    dil, 48
        setne   al
        shl     al, 2
        or      al, cl
        mov     ecx, edi
        shr     cl, 7
        shr     dil, 6
        and     dil, 1
        or      dil, cl
        shl     dil, 3
        or      al, dil
        ret