I'm trying to find an optimized method for RGB8 (actually grayscale) to RGB32 image conversion.
Source is an 8 bits grey image, Destination should be an 32 bits grey image (BGRA) with 4th channel (alpha) to be ignored. Source address is not guaranteed to be 16 byte aligned, Count is a multiple of 16, Destination address is 16 byte aligned.
- INPUT: 8 bits single channel grey image
- OUTPUT: 32 bits BGRA (alpha channel ignored)
- COUNT: Image size is a multiple of 16
- CPU: x86-32 (SSE2/SSE3 allowed)
Here is my optimized assembly code. Is there an even faster way of conversion?
void ConvertGreyToRgb32Assembler(__m128i* Source, __m128i* Destination, unsigned int Count) {
static unsigned int __declspec(align(64)) Masks[] = {
0x80000000, 0x80010101, 0x80020202, 0x80030303,
0x80040404, 0x80050505, 0x80060606, 0x80070707,
0x80080808, 0x80090909, 0x800a0a0a, 0x800b0b0b,
0x800c0c0c, 0x800d0d0d, 0x800e0e0e, 0x800f0f0f
};
__asm {
mov esi, Source
mov edi, Destination
mov edx, Count
xor ecx, ecx
movdqa xmm4, xmmword ptr [Masks + 0]
movdqa xmm5, xmmword ptr [Masks + 16]
movdqa xmm6, xmmword ptr [Masks + 32]
movdqa xmm7, xmmword ptr [Masks + 48]
l1:
movdqu xmm0, xmmword ptr [esi + ecx]
movdqa xmm1, xmm0
movdqa xmm2, xmm0
movdqa xmm3, xmm0
pshufb xmm0, xmm4
pshufb xmm1, xmm5
pshufb xmm2, xmm6
pshufb xmm3, xmm7
movntdq [edi + 0], xmm0
movntdq [edi + 16], xmm1
movntdq [edi + 32], xmm2
movntdq [edi + 48], xmm3
add edi, 64
add ecx, 16
cmp ecx, edx
jb l1
}
}
There is another approach using several PUNPCKLBW and PUNPCKHBW but that seems to be slightly slower.
Update: This is the basic non optimized algorithm:
BGRA* Destination = ...
unsigned char* Source ...
for (unsigned int i = 0; i < Size; i++) {
Destination[i].Blue = Source[i];
Destination[i].Green = Source[i];
Destination[i].Red = Source[i];
}
PS: I also tried using C code with MS VS2008 SSE compiler intrinsics. It turned out that the compiler generated a lot of unnecessary memory moves which causes the code to be 10-20% slower than pure assembly.
Update 2: This is the same code by using intrinsics only.
void ConvertGreyToRgb32Assembler(__m128i* Source, __m128i* Destination, unsigned int Count) {
static const unsigned int __declspec(align(64)) Masks[] = {
0x80000000, 0x80010101, 0x80020202, 0x80030303,
0x80040404, 0x80050505, 0x80060606, 0x80070707,
0x80080808, 0x80090909, 0x800a0a0a, 0x800b0b0b,
0x800c0c0c, 0x800d0d0d, 0x800e0e0e, 0x800f0f0f
};
register __m128i m0 = _mm_load_si128((__m128i*) (Masks + 0));
register __m128i m1 = _mm_load_si128((__m128i*) (Masks + 4));
register __m128i m2 = _mm_load_si128((__m128i*) (Masks + 8));
register __m128i m3 = _mm_load_si128((__m128i*) (Masks + 12));
for (unsigned int i = 0; i < Count / 16; i++) {
__m128i r0 = _mm_load_si128(Source + i);
_mm_stream_si128(Destination + (i * 4) + 0, _mm_shuffle_epi8(r0, m0));
_mm_stream_si128(Destination + (i * 4) + 1, _mm_shuffle_epi8(r0, m1));
_mm_stream_si128(Destination + (i * 4) + 2, _mm_shuffle_epi8(r0, m2));
_mm_stream_si128(Destination + (i * 4) + 3, _mm_shuffle_epi8(r0, m3));
}
}
Update 3: This is the compiler generated code (beautified) (Visual Studio 2012, all optimization on):
push ebp
mov ebp, esp
mov edx, dword ptr [ebp+8]
movdqa xmm1, xmmword ptr ds:[Masks + 0]
movdqa xmm2, xmmword ptr ds:[Masks + 16]
movdqa xmm3, xmmword ptr ds:[Masks + 32]
movdqa xmm4, xmmword ptr ds:[Masks + 48]
push esi
test ecx, ecx
je l2
lea esi, [ecx-1]
shr esi, 4
inc esi
l1:
mov ecx, edx
movdqu xmm0, xmmword ptr [ecx]
mov ecx, eax
movdqa xmm5, xmm0
pshufb xmm5, xmm1
movdqa xmmword ptr [ecx], xmm5
movdqa xmm5, xmm0
pshufb xmm5, xmm2
movdqa xmmword ptr [eax+10h], xmm5
movdqa xmm5, xmm0
pshufb xmm5, xmm3
movdqa xmmword ptr [eax+20h], xmm5
lea ecx, [eax+30h]
add edx, 10h
add eax, 40h
dec esi
pshufb xmm0, xmm4
movdqa xmmword ptr [ecx], xmm0
jne l1
l2:
pop esi
pop ebp
ret
It seems that interleaving movdqa with pshufb is some what faster.
Update 4: This seems to be the optimal hand optimized code:
__asm {
mov esi, Source
mov edi, Destination
mov ecx, Count
movdqu xmm0, xmmword ptr [esi]
movdqa xmm4, xmmword ptr [Masks + 0]
movdqa xmm5, xmmword ptr [Masks + 16]
movdqa xmm6, xmmword ptr [Masks + 32]
movdqa xmm7, xmmword ptr [Masks + 48]
l1:
dec ecx
lea edi, [ edi + 64 ]
lea esi, [ esi + 16 ]
movdqa xmm1, xmm0
movdqa xmm2, xmm0
movdqa xmm3, xmm0
pshufb xmm0, xmm4
movdqa [edi - 64], xmm0
pshufb xmm1, xmm5
movdqa [edi - 48], xmm1
pshufb xmm2, xmm6
movdqa [edi - 32], xmm2
pshufb xmm3, xmm7
movdqa [edi - 16], xmm3
movdqu xmm0, xmmword ptr [esi]
ja l1
}
Update 5: This conversion algorithm uses punpck instruction. However this conversion routine is a bit slower than using masks and pushfb.
for (unsigned int i = 0; i < Count; i += 16) {
register __m128i r0 = _mm_load_si128(Source++);
register __m128i r1 = _mm_unpackhi_epi8(r0, r0);
register __m128i r2 = _mm_unpacklo_epi8(r0, r0);
register __m128i r3 = _mm_unpackhi_epi8(r1, r1);
register __m128i r4 = _mm_unpacklo_epi8(r1, r1);
register __m128i r5 = _mm_unpackhi_epi8(r2, r2);
register __m128i r6 = _mm_unpacklo_epi8(r2, r2);
_mm_store_si128(Destination++, r6);
_mm_store_si128(Destination++, r5);
_mm_store_si128(Destination++, r4);
_mm_store_si128(Destination++, r3);
}
Update 6: For the sake of completeness this is the inverse method to convert from 32 bits back to 8 bits grey image.
static void ConvertRgb32ToGrey(const __m128i* Source, __m128i* Destination, unsigned int Count) {
static const unsigned char __declspec(align(64)) Masks[] = {
0x00, 0x04, 0x08, 0x0c, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80, 0x80, 0x00, 0x04, 0x08, 0x0c, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x00, 0x04, 0x08, 0x0c, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x00, 0x04, 0x08, 0x0c,
};
register __m128i m0 = _mm_load_si128((__m128i*) (Masks + 0));
register __m128i m1 = _mm_load_si128((__m128i*) (Masks + 16));
register __m128i m2 = _mm_load_si128((__m128i*) (Masks + 32));
register __m128i m3 = _mm_load_si128((__m128i*) (Masks + 48));
for (unsigned int i = 0; i < Count / 64; i++) {
__m128i a = _mm_load_si128(Source + (i * 4) + 0);
__m128i b = _mm_load_si128(Source + (i * 4) + 1);
__m128i c = _mm_load_si128(Source + (i * 4) + 2);
__m128i d = _mm_load_si128(Source + (i * 4) + 3);
a = _mm_shuffle_epi8(a, m0);
b = _mm_shuffle_epi8(b, m1);
c = _mm_shuffle_epi8(c, m2);
d = _mm_shuffle_epi8(d, m3);
__m128i e = _mm_or_si128(a, b);
__m128i f = _mm_or_si128(c, d);
__m128i g = _mm_or_si128(e, f);
_mm_stream_si128(Destination + i, g);
}
}