jdsample.c

jpeg编解码器

METHODDEF(void)
h2v1_fancy_upsample_orig (j_decompress_ptr cinfo, jpeg_component_info * compptr,
			  JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr);

METHODDEF(void)
h2v1_fancy_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
{
  if (MMXAvailable)
    h2v1_fancy_upsample_mmx(cinfo, compptr, input_data, output_data_ptr);
  else
    h2v1_fancy_upsample_orig(cinfo, compptr, input_data, output_data_ptr);
}

METHODDEF(void)
h2v1_fancy_upsample_mmx (j_decompress_ptr cinfo, jpeg_component_info * compptr,
		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
{
  JSAMPARRAY output_data = *output_data_ptr;
#if defined(HAVE_MMX_INTEL_MNEMONICS) /* see in h2v2 for comments */
   register JSAMPROW inptr, outptr;
#else
   JSAMPROW inptr, outptr;
#endif
  int inrow, hsize = compptr->downsampled_width;

  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
    inptr = input_data[inrow];
    outptr = output_data[inrow];

#if defined(HAVE_MMX_INTEL_MNEMONICS) 
	_asm {
			mov ecx, hsize			;// horizontal line size
			mov esi, inptr			;// input buffer pointer
			mov edi, outptr			;// output buffer pointer

			pxor mm6, mm6			;// zero register
			movq mm7, [esi]			;// input register

			;// Special 1st column case - process low 8 bytes of mm7
			movq mm0, mm7			;// move 1st quadword into mm7
			movq mm1, mm7			;// make a copy
			movq mm2, mm7			;// make a copy

			punpcklbw mm0, mm6		;// unpack lower values; inptr[0][1][2][3]
			movq mm3, mm0			;// make a copy
			pmullw mm0, mul3w		;// multiply by 3
			psllq mm1, 8			;// shift 1 byte for previous values; inptr[-1][0][1][2]
			movq mm5, mm7			;// copy original data
			pand mm5, mask2			;// mask out all but lower byte for "previous" state
			paddb mm1, mm5			;// add in byte to quadword
			psrlq mm2, 8			;// shift right for "next" state; inptr[1][2][3][4]
			punpcklbw mm1, mm6		;// unpack 
			punpcklbw mm2, mm6		;// unpack

			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias 
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte
			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave
			movq [edi], mm2			;// write out results

			;// process high 8 bytes of mm7
			movq mm0, mm7			;// copy input data
			movq mm1, mm7			;// copy input data
			movq mm2, mm7			;// copy input data
			movq mm3, mm7			;// copy input data

			punpckhbw mm0, mm6		;// unpack hi data
			pmullw mm0, mul3w		;// multiply by 3
			psllq mm1, 8			;// shift 1 byte for previous values; inptr[-1][0][1][2]
			psrlq mm2, 8			;// shift right for "next" state; inptr[1][2][3][4]
			movq mm7, [esi+8]		;// get next quadword from input buffer
			movq mm5, mm7			;// make copy
			psllq mm5, 56			;// shift left to isolate LSB
			paddb mm2, mm5			;// add in byte for "next" state

			punpckhbw mm1, mm6		;// unpack
			punpckhbw mm2, mm6		;// unpack
			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias 
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte
			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave
			movq [edi+8], mm2		;// write out results

			add edi, 16			;// increment output buffer pointer
			add esi, 8			;// increment input buffer pointer
			sub ecx, 8			;// increment column counter
			cmp ecx, 8			;// cmp with 8
			jle last_col			;// if less that goto last column

			;// Main Loop - process low 8 bytes of mm7
		col_loop:
			movq mm0, mm7			;// copy input data
			movq mm1, mm7			;// copy input data
			movq mm2, mm7			;// copy input data
			punpcklbw mm0, mm6		;// unpack lo data
			pmullw mm0, mul3w		;// multiply by 3; i[0][1][2][3]
			psllq mm1, 8			;// shift left to get previous byte
			movq mm5, mm3			;// retrieve copy of "previous" state
			psrlq mm5, 56			;// shift to get LSB
			paddb mm1, mm5			;// add in byte
			psrlq mm2, 8			;// shift rt for "next" state
			punpcklbw mm1, mm6		;// unpack
			punpcklbw mm2, mm6		;// unpack
			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias 
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias 
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte
			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave

			movq [edi], mm2			;// write out results

			;// process high 8 bytes of mm7
			movq mm0, mm7			;// copy input data
			movq mm1, mm7			;// copy input data
			movq mm2, mm7			;// copy input data
			movq mm3, mm7			;// copy input data

			punpckhbw mm0, mm6		;// unpack hi data
			pmullw mm0, mul3w		;// multiply by 3; i[0][1][2][3]
			psllq mm1, 8			;// shift left to get previous byte		    
			psrlq mm2, 8			;// shift rt for "next" state
			movq mm7, [esi+8]		;// get next quadword from input buffer
			movq mm5, mm7			;// make copy
			psllq mm5, 56			;// shift left for LSB
			paddb mm2, mm5			;// add in byte

			punpckhbw mm1, mm6		;// unpack
			punpckhbw mm2, mm6		;// unpack
			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias 
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte

			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave

			movq [edi+8], mm2		;// write out results

			add edi, 16			;// increment output buffer pointer
			add esi, 8			;// increment input buffer pointer
			sub ecx, 8			;// increment column counter
			cmp ecx, 8			;// cmp with 8
			jg col_loop			;// if &gt; 8 goto main loop

		last_col:
			;// Special last column case - process low 8 bytes of mm7
			movq mm0, mm7			;// copy input data
			movq mm1, mm7			;// copy input data
			movq mm2, mm7			;// copy input data

			punpcklbw mm0, mm6		;// unpack lo data

			pmullw mm0, mul3w		;// multiply by 3; i[0][1][2][3]

			psllq mm1, 8			;// shift left to get previous byte
			
			movq mm5, mm3			;// retrieve copy of "previous" state
			psrlq mm5, 56			;// shift left for MSB
			paddb mm1, mm5			;// add in byte

			psrlq mm2, 8			;// shift rt for "next" state
			punpcklbw mm1, mm6		;// unpack
			punpcklbw mm2, mm6		;// unpack
			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte
			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave
			movq [edi], mm2			;// write out results

			;// Special last column case - process hi 8 bytes of mm7
			movq mm0, mm7			;// copy input data
			movq mm1, mm7			;// copy input data
			movq mm2, mm7			;// copy input data
			punpckhbw mm0, mm6		;// unpack hi data

			pmullw mm0, mul3w		;// multiply by 3; i[0][1][2][3]
			psllq mm1, 8			;// shift left to get previous byte
			psrlq mm2, 8			;// shift rt for "next" state
			pand mm7, mask1			;// mask out all but MSB
			paddb mm2, mm7			;// add in byte
			punpckhbw mm1, mm6		;// unpack
			punpckhbw mm2, mm6		;// unpack
			paddw mm1, mm0			;// add in result from multiply to "previous" data
			paddw mm1, bias1w		;// add in bias
			paddw mm2, mm0			;// add in result from multiply to "next" data
			paddw mm2, bias2w		;// add in bias
			psrlw mm1, 2			;// convert from word to byte
			psrlw mm2, 2			;// convert from word to byte
			psllq mm2, 8			;// prepare for interleave
			paddb mm2, mm1			;// do interleave
			movq [edi+8], mm2		;// write out results
			emms
		}
#endif
#if defined(HAVE_MMX_ATT_MNEMONICS)   
  __asm__ (
    "movl %0, %%ecx          \n\t"  // horizontal line size
    "movl %1, %%esi          \n\t"  // input buffer pointer
    "movl %2, %%edi         \n\t"  // output buffer pointer

    "pxor %%mm6,%%mm6            \n\t"  // zero register
    "movq (%%esi),%%mm7          \n\t"  // input register

    // Special 1st column case - process low 8 bytes of mm7
    "movq %%mm7,%%mm0            \n\t"  // move 1st quadword into mm7
    "movq %%mm7,%%mm1            \n\t"  // make a copy
    "movq %%mm7,%%mm2            \n\t"  // make a copy

    "punpcklbw %%mm6,%%mm0       \n\t" // unpack lower values; inptr[0][1][2][3]
    "movq %%mm0,%%mm3            \n\t"  // make a copy
    "pmullw mul3w,%%mm0         \n\t" // multiply by 3
    "psllq $8,%%mm1              \n\t"  // shift 1 byte for previous values; inptr[-1][0][1][2]
    "movq %%mm7,%%mm5            \n\t"  // copy original data
    "pand mask2,%%mm5           \n\t"  // mask out all but lower byte for "previous" state
    "paddb %%mm5,%%mm1           \n\t"  // add in byte to quadword
    "psrlq $8,%%mm2              \n\t"  // shift right for "next" state; inptr[1][2][3][4]
    "punpcklbw %%mm6,%%mm1       \n\t" // unpack 
    "punpcklbw %%mm6,%%mm2       \n\t" // unpack

    "paddw %%mm0,%%mm1           \n\t"  // add in result from multiply to "previous" data
    "paddw bias1w,%%mm1         \n\t" // add in bias 
    "paddw %%mm0,%%mm2           \n\t"  // add in result from multiply to "next" data
    "paddw bias2w,%%mm2         \n\t" // add in bias
    "psrlw $2,%%mm1              \n\t"  // convert from word to byte
    "psrlw $2,%%mm2              \n\t"  // convert from word to byte
    "psllq $8,%%mm2              \n\t"  // prepare for interleave
    "paddb %%mm1,%%mm2           \n\t"  // do interleave
    "movq %%mm2,(%%edi)          \n\t"  // write out results

    // process high 8 bytes of mm7
    "movq %%mm7,%%mm0            \n\t"  // copy input data
    "movq %%mm7,%%mm1            \n\t"  // copy input data
    "movq %%mm7,%%mm2            \n\t"  // copy input data
    "movq %%mm7,%%mm3            \n\t"  // copy input data

    "punpckhbw %%mm6,%%mm0       \n\t" // unpack hi data
    "pmullw mul3w,%%mm0         \n\t" // multiply by 3
    "psllq $8,%%mm1              \n\t"  // shift 1 byte for previous values; inptr[-1][0][1][2]
    "psrlq $8,%%mm2              \n\t"  // shift right for "next" state; inptr[1][2][3][4]
    "movq 8(%%esi),%%mm7         \n\t" // get next quadword from input buffer
    "movq %%mm7,%%mm5            \n\t"  // make copy
    "psllq $56,%%mm5             \n\t"  // shift left to isolate LSB
    "paddb %%mm5,%%mm2           \n\t"  // add in byte for "next" state

    "punpckhbw %%mm6,%%mm1       \n\t" // unpack
    "punpckhbw %%mm6,%%mm2       \n\t" // unpack
    "paddw %%mm0,%%mm1           \n\t"  // add in result from multiply to "previous" data
    "paddw bias1w,%%mm1         \n\t" // add in bias 
    "paddw %%mm0,%%mm2           \n\t"  // add in result from multiply to "next" data
    "paddw bias2w,%%mm2         \n\t" // add in bias
    "psrlw $2,%%mm1              \n\t"  // convert from word to byte
    "psrlw $2,%%mm2              \n\t"  // convert from word to byte
    "psllq $8,%%mm2              \n\t"  // prepare for interleave
    "paddb %%mm1,%%mm2           \n\t"  // do interleave
    "movq %%mm2,8(%%edi)         \n\t" // write out results

    "addl $16,%%edi              \n\t" // increment output buffer pointer
    "addl $8,%%esi               \n\t" // increment input buffer pointer
    "subl $8,%%ecx               \n\t" // increment column counter
    "cmpl $8,%%ecx               \n\t" // cmp with 8
    "jle last_col                \n\t"  // if less that goto last column

    // Main Loop - process low 8 bytes of mm7
  "col_loop_a:                   \n\t"
    "movq %%mm7,%%mm0            \n\t"  // copy input data
    "movq %%mm7,%%mm1            \n\t"  // copy input data
    "movq %%mm7,%%mm2            \n\t"  // copy input data
    "punpcklbw %%mm6,%%mm0       \n\t" // unpack lo data
    "pmullw mul3w,%%mm0         \n\t" // multiply by 3; i[0][1][2][3]
    "psllq $8,%%mm1              \n\t"  // shift left to get previous byte
    "movq %%mm3,%%mm5            \n\t"  // retrieve copy of "previous" state
    "psrlq $56,%%mm5             \n\t"  // shift to get LSB
    "paddb %%mm5,%%mm1           \n\t"  // add in byte
    "psrlq $8,%%mm2              \n\t"  // shift rt for "next" state
    "punpcklbw %%mm6,%%mm1       \n\t" // unpack
    "punpcklbw %%mm6,%%mm2       \n\t" // unpack
    "paddw %%mm0,%%mm1           \n\t"  // add in result from multiply to "previous" data
    "paddw bias1w,%%mm1         \n\t" // add in bias 
    "paddw %%mm0,%%mm2           \n\t"  // add in result from multiply to "next" data
    "paddw bias2w,%%mm2         \n\t" // add in bias 
    "psrlw $2,%%mm1              \n\t"  // convert from word to byte
    "psrlw $2,%%mm2              \n\t"  // convert from word to byte
    "psllq $8,%%mm2              \n\t"  // prepare for interleave
    "paddb %%mm1,%%mm2           \n\t"  // do interleave

    "movq %%mm2,(%%edi)          \n\t"  // write out results

    // process high 8 bytes of mm7
    "movq %%mm7,%%mm0            \n\t"  // copy input data
    "movq %%mm7,%%mm1            \n\t"  // copy input data
    "movq %%mm7,%%mm2            \n\t"  // copy input data
    "movq %%mm7,%%mm3            \n\t"  // copy input data

    "punpckhbw %%mm6,%%mm0       \n\t" // unpack hi data
    "pmullw mul3w,%%mm0         \n\t" // multiply by 3; i[0][1][2][3]
    "psllq $8,%%mm1              \n\t"  // shift left to get previous byte                 
    "psrlq $8,%%mm2              \n\t"  // shift rt for "next" state
    "movq 8(%%esi),%%mm7         \n\t" // get next quadword from input buffer
    "movq %%mm7,%%mm5            \n\t"  // make copy
    "psllq $56,%%mm5             \n\t"  // shift left for LSB