This program is still to be converted. Because the Windows source uses ImageBuffer.cpp and ImageBuffer.h and gdilib, this will take some time.
avxpackedintegerthreshold.asm
; Name: avxpackedintegerthreshold.asm ; ; Build: nasm -f elf32 -o avxpackedintegerthreshold.o avxpackedintegerthreshold.asm ; ; Source: Modern x86 Assembly Language Programming p. 425 ; ; Remark: Until ImageBuffer.cpp is converted, this example is of little use. extern NUM_PIXELS_MAX global AvxPiThreshold global AvxPiCalcMean ; Image threshold data structure (see AvxPackedIntegerThreshold.h) struc ITD .PbSrc resd 1 .PbMask resd 1 .NumPixels resd 1 .Threshold resb 1 .Pad resb 3 .NumMaskedPixels resd 1 .SumMaskedPixels resd 1 .MeanMaskedPixels resq 1 endstruc ; Marco AvxPiCalcMeanUpdateSums ; ; Description: The following macro updates sum_masked_pixels in ymm4. ; It also resets any necessary intermediate values in ; order to prevent an overflow condition. ; ; Register contents: ; ymm3:ymm2 = packed word sum_masked_pixels ; ymm4 = packed dword sum_masked_pixels ; ymm7 = packed zero ; ; Temp registers: ; ymm0, ymm1, ymm5, ymm6 %macro AvxPiCalcMeanUpdateSums 0 ; Promote packed word sum_masked_pixels to dword vpunpcklwd ymm0,ymm2,ymm7 vpunpcklwd ymm1,ymm3,ymm7 vpunpckhwd ymm5,ymm2,ymm7 vpunpckhwd ymm6,ymm3,ymm7 ; Update packed dword sums in sum_masked_pixels vpaddd ymm0,ymm0,ymm1 vpaddd ymm5,ymm5,ymm6 vpaddd ymm4,ymm4,ymm0 vpaddd ymm4,ymm4,ymm5 ; Reset intermediate values xor edx,edx ;reset update counter vpxor ymm2,ymm2,ymm2 ;reset sum_masked_pixels lo vpxor ymm3,ymm3,ymm3 ;reset sum_masked_pixels hi %endmacro ; Custom segment for constant values section .data PixelScale times 32 db 0 ;uint8 to int8 scale value R8_MinusOne dq -1.0 ;invalid mean value section .text ; extern "C" bool AvxPiThreshold(ITD* itd); ; ; Description: The following function performs image thresholding ; of an 8 bits-per-pixel grayscale image. ; ; Returns: 0 = invalid size or unaligned image buffer ; 1 = success ; ; Requires: AVX2 AvxPiThreshold: push ebp mov ebp,esp push esi push edi ; Load and verify the argument values in ITD structure mov edx,[ebp+8] ;edx = 'itd' xor eax,eax ;set error return code mov ecx,[edx+ITD.NumPixels] ;ecx = NumPixels test ecx,ecx jz .done ;jump if num_pixels == 0 cmp ecx,[NUM_PIXELS_MAX] ja .done ;jump if num_pixels too big test ecx,1fh jnz .done ;jump if num_pixels % 32 != 0 shr ecx,5 ;ecx = number of packed pixels mov esi,[edx+ITD.PbSrc] ;esi = PbSrc test esi,1fh jnz .done ;jump if misaligned mov edi,[edx+ITD.PbMask] ;edi = PbMask test edi,1fh jnz .done ;jump if misaligned ; Initialize packed threshold vpbroadcastb ymm0,[edx+ITD.Threshold] ;ymm0 = packed threshold vmovdqa ymm7, [PixelScale] ;ymm7 = uint8 to int8 SF vpsubb ymm2,ymm0,ymm7 ;ymm1 = scaled threshold ; Create the mask image .@1: vmovdqa ymm0, [esi] ;load next packed pixel vpsubb ymm1,ymm0,ymm7 ;ymm1 = scaled image pixels vpcmpgtb ymm3,ymm1,ymm2 ;compare against threshold vmovdqa [edi],ymm3 ;save packed threshold mask add esi,32 add edi,32 dec ecx jnz .@1 ;repeat until done mov eax,1 ;set return code .done: pop edi pop esi pop ebp ret ; extern "C" bool AvxPiCalcMean(ITD* itd); ; ; Description: The following function calculates the mean value all ; above-threshold image pixels using the mask created by ; function AvxPiThreshold_. ; ; Returns: 0 = invalid image size or unaligned image buffer ; 1 = success ; ; Requires: AVX2, POPCNT AvxPiCalcMean: push ebp mov ebp,esp push ebx push esi push edi ; Load and verify the argument values in ITD structure mov eax,[ebp+8] ;eax = 'itd' mov ecx,[eax+ITD.NumPixels] ;ecx = NumPixels test ecx,ecx jz .error ;jump if num_pixels == 0 cmp ecx,[NUM_PIXELS_MAX] ja .error ;jump if num_pixels too big test ecx,1fh jnz .error ;jump if num_pixels % 32 != 0 shr ecx,5 ;ecx = number of packed pixels mov edi,[eax+ITD.PbMask] ;edi = PbMask test edi,1fh jnz .error ;jump if PbMask not aligned mov esi,[eax+ITD.PbSrc] ;esi = PbSrc test esi,1fh jnz .error ;jump if PbSrc not aligned ; Initialize values for mean calculation xor edx,edx ;edx = update counter vpxor ymm7,ymm7,ymm7 ;ymm7 = packed zero vmovdqa ymm2,ymm7 ;ymm2 = sum_masked_pixels (16 words) vmovdqa ymm3,ymm7 ;ymm3 = sum_masked_pixels (16 words) vmovdqa ymm4,ymm7 ;ymm4 = sum_masked_pixels (8 dwords) xor ebx,ebx ;ebx = num_masked_pixels (1 dword) ; Register usage for processing loop ; esi = PbSrc, edi = PbMask, eax = scratch register ; ebx = num_pixels_masked, ecx = NumPixels / 32, edx = update counter ; ; ymm0 = packed pixel, ymm1 = packed mask ; ymm3:ymm2 = sum_masked_pixels (32 words) ; ymm4 = sum_masked_pixels (8 dwords) ; ymm5 = scratch register ; ymm6 = scratch register ; ymm7 = packed zero .@1: vmovdqa ymm0, [esi] ;load next packed pixel vmovdqa ymm1, [edi] ;load next packed mask ; Update mum_masked_pixels vpmovmskb eax,ymm1 popcnt eax,eax add ebx,eax ; Update sum_masked_pixels (word values) vpand ymm6,ymm0,ymm1 ;set non-masked pixels to zero vpunpcklbw ymm0,ymm6,ymm7 vpunpckhbw ymm1,ymm6,ymm7 ;ymm1:ymm0 = masked pixels (words) vpaddw ymm2,ymm2,ymm0 vpaddw ymm3,ymm3,ymm1 ;ymm3:ymm2 = sum_masked_pixels ; Check and see if it's necessary to update the dword sum_masked_pixels ; in xmm4 and num_masked_pixels in ebx inc edx cmp edx,255 jb .noUpdate AvxPiCalcMeanUpdateSums .noUpdate: add esi,32 add edi,32 dec ecx jnz .@1 ;repeat loop until done ; Main processing loop is finished. If necessary, perform final update ; of sum_masked_pixels in xmm4 & num_masked_pixels in ebx. test edx,edx jz .@2 AvxPiCalcMeanUpdateSums ; Compute and save final sum_masked_pixels & num_masked_pixels .@2: vextracti128 xmm0,ymm4,1 vpaddd xmm1,xmm0,xmm4 vphaddd xmm2,xmm1,xmm7 vphaddd xmm3,xmm2,xmm7 vmovd edx,xmm3 ;edx = final sum_mask_pixels mov eax,[ebp+8] ;eax = 'itd' mov [eax+ITD.SumMaskedPixels],edx ;save final sum_masked_pixels mov [eax+ITD.NumMaskedPixels],ebx ;save final num_masked_pixels ; Compute mean of masked pixels test ebx,ebx ;is num_mask_pixels zero? jz .noMean ;if yes, skip calc of mean vcvtsi2sd xmm0,xmm0,edx ;xmm0 = sum_masked_pixels vcvtsi2sd xmm1,xmm1,ebx ;xmm1 = num_masked_pixels vdivsd xmm0,xmm0,xmm1 ;xmm0 = mean_masked_pixels jmp .@3 .noMean: vmovsd xmm0,[R8_MinusOne] ;use -1.0 for no mean .@3: vmovsd [eax+ITD.MeanMaskedPixels],xmm0 ;save mean mov eax,1 ;set return code vzeroupper .done: pop edi pop esi pop ebx pop ebp ret .error: xor eax,eax ;set error return code jmp .done
build
nasm -f elf32 -o avxpackedintegerthreshold.o avxpackedintegerthreshold.asm