michael@0: /*
michael@0:  * jsimd_i386.c
michael@0:  *
michael@0:  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0:  * Copyright 2009-2011 D. R. Commander
michael@0:  * 
michael@0:  * Based on the x86 SIMD extension for IJG JPEG library,
michael@0:  * Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0:  * For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0:  *
michael@0:  * This file contains the interface between the "normal" portions
michael@0:  * of the library and the SIMD implementations when running on a
michael@0:  * 32-bit x86 architecture.
michael@0:  */
michael@0: 
michael@0: #define JPEG_INTERNALS
michael@0: #include "../jinclude.h"
michael@0: #include "../jpeglib.h"
michael@0: #include "../jsimd.h"
michael@0: #include "../jdct.h"
michael@0: #include "../jsimddct.h"
michael@0: #include "jsimd.h"
michael@0: 
michael@0: /*
michael@0:  * In the PIC cases, we have no guarantee that constants will keep
michael@0:  * their alignment. This macro allows us to verify it at runtime.
michael@0:  */
michael@0: #define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0)
michael@0: 
michael@0: #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
michael@0: 
michael@0: static unsigned int simd_support = ~0;
michael@0: 
michael@0: /*
michael@0:  * Check what SIMD accelerations are supported.
michael@0:  *
michael@0:  * FIXME: This code is racy under a multi-threaded environment.
michael@0:  */
michael@0: LOCAL(void)
michael@0: init_simd (void)
michael@0: {
michael@0:   char *env = NULL;
michael@0: 
michael@0:   if (simd_support != ~0U)
michael@0:     return;
michael@0: 
michael@0:   simd_support = jpeg_simd_cpu_support();
michael@0: 
michael@0:   /* Force different settings through environment variables */
michael@0:   env = getenv("JSIMD_FORCEMMX");
michael@0:   if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0:     simd_support &= JSIMD_MMX;
michael@0:   env = getenv("JSIMD_FORCE3DNOW");
michael@0:   if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0:     simd_support &= JSIMD_3DNOW|JSIMD_MMX;
michael@0:   env = getenv("JSIMD_FORCESSE");
michael@0:   if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0:     simd_support &= JSIMD_SSE|JSIMD_MMX;
michael@0:   env = getenv("JSIMD_FORCESSE2");
michael@0:   if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0:     simd_support &= JSIMD_SSE2;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_rgb_ycc (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_rgb_gray (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_ycc_rgb (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
michael@0:                        JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
michael@0:                        JDIMENSION output_row, int num_rows)
michael@0: {
michael@0:   void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0:   void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0: 
michael@0:   switch(cinfo->in_color_space)
michael@0:   {
michael@0:     case JCS_EXT_RGB:
michael@0:       sse2fct=jsimd_extrgb_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extrgb_ycc_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_RGBX:
michael@0:     case JCS_EXT_RGBA:
michael@0:       sse2fct=jsimd_extrgbx_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extrgbx_ycc_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGR:
michael@0:       sse2fct=jsimd_extbgr_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extbgr_ycc_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGRX:
michael@0:     case JCS_EXT_BGRA:
michael@0:       sse2fct=jsimd_extbgrx_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extbgrx_ycc_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XBGR:
michael@0:     case JCS_EXT_ABGR:
michael@0:       sse2fct=jsimd_extxbgr_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extxbgr_ycc_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XRGB:
michael@0:     case JCS_EXT_ARGB:
michael@0:       sse2fct=jsimd_extxrgb_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_extxrgb_ycc_convert_mmx;
michael@0:       break;
michael@0:     default:
michael@0:       sse2fct=jsimd_rgb_ycc_convert_sse2;
michael@0:       mmxfct=jsimd_rgb_ycc_convert_mmx;
michael@0:       break;
michael@0:   }
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
michael@0:     sse2fct(cinfo->image_width, input_buf,
michael@0:         output_buf, output_row, num_rows);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     mmxfct(cinfo->image_width, input_buf,
michael@0:         output_buf, output_row, num_rows);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_rgb_gray_convert (j_compress_ptr cinfo,
michael@0:                         JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
michael@0:                         JDIMENSION output_row, int num_rows)
michael@0: {
michael@0:   void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0:   void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0: 
michael@0:   switch(cinfo->in_color_space)
michael@0:   {
michael@0:     case JCS_EXT_RGB:
michael@0:       sse2fct=jsimd_extrgb_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extrgb_gray_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_RGBX:
michael@0:     case JCS_EXT_RGBA:
michael@0:       sse2fct=jsimd_extrgbx_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extrgbx_gray_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGR:
michael@0:       sse2fct=jsimd_extbgr_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extbgr_gray_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGRX:
michael@0:     case JCS_EXT_BGRA:
michael@0:       sse2fct=jsimd_extbgrx_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extbgrx_gray_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XBGR:
michael@0:     case JCS_EXT_ABGR:
michael@0:       sse2fct=jsimd_extxbgr_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extxbgr_gray_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XRGB:
michael@0:     case JCS_EXT_ARGB:
michael@0:       sse2fct=jsimd_extxrgb_gray_convert_sse2;
michael@0:       mmxfct=jsimd_extxrgb_gray_convert_mmx;
michael@0:       break;
michael@0:     default:
michael@0:       sse2fct=jsimd_rgb_gray_convert_sse2;
michael@0:       mmxfct=jsimd_rgb_gray_convert_mmx;
michael@0:       break;
michael@0:   }
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
michael@0:     sse2fct(cinfo->image_width, input_buf,
michael@0:         output_buf, output_row, num_rows);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     mmxfct(cinfo->image_width, input_buf,
michael@0:         output_buf, output_row, num_rows);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
michael@0:                        JSAMPIMAGE input_buf, JDIMENSION input_row,
michael@0:                        JSAMPARRAY output_buf, int num_rows)
michael@0: {
michael@0:   void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
michael@0:   void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
michael@0: 
michael@0:   switch(cinfo->out_color_space)
michael@0:   {
michael@0:     case JCS_EXT_RGB:
michael@0:       sse2fct=jsimd_ycc_extrgb_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extrgb_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_RGBX:
michael@0:     case JCS_EXT_RGBA:
michael@0:       sse2fct=jsimd_ycc_extrgbx_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extrgbx_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGR:
michael@0:       sse2fct=jsimd_ycc_extbgr_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extbgr_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGRX:
michael@0:     case JCS_EXT_BGRA:
michael@0:       sse2fct=jsimd_ycc_extbgrx_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extbgrx_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XBGR:
michael@0:     case JCS_EXT_ABGR:
michael@0:       sse2fct=jsimd_ycc_extxbgr_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extxbgr_convert_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XRGB:
michael@0:     case JCS_EXT_ARGB:
michael@0:       sse2fct=jsimd_ycc_extxrgb_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_extxrgb_convert_mmx;
michael@0:       break;
michael@0:     default:
michael@0:       sse2fct=jsimd_ycc_rgb_convert_sse2;
michael@0:       mmxfct=jsimd_ycc_rgb_convert_mmx;
michael@0:       break;
michael@0:   }
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
michael@0:     sse2fct(cinfo->output_width, input_buf,
michael@0:         input_row, output_buf, num_rows);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     mmxfct(cinfo->output_width, input_buf,
michael@0:         input_row, output_buf, num_rows);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v2_downsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v1_downsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                        JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0:         compptr->v_samp_factor, compptr->width_in_blocks,
michael@0:         input_data, output_data);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0:         compptr->v_samp_factor, compptr->width_in_blocks,
michael@0:         input_data, output_data);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                        JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0:         compptr->v_samp_factor, compptr->width_in_blocks,
michael@0:         input_data, output_data);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0:         compptr->v_samp_factor, compptr->width_in_blocks,
michael@0:         input_data, output_data);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v2_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v1_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v2_upsample (j_decompress_ptr cinfo,
michael@0:                      jpeg_component_info * compptr, 
michael@0:                      JSAMPARRAY input_data,
michael@0:                      JSAMPARRAY * output_data_ptr)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
michael@0:         cinfo->output_width, input_data, output_data_ptr);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor,
michael@0:         cinfo->output_width, input_data, output_data_ptr);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v1_upsample (j_decompress_ptr cinfo,
michael@0:                      jpeg_component_info * compptr, 
michael@0:                      JSAMPARRAY input_data,
michael@0:                      JSAMPARRAY * output_data_ptr)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
michael@0:         cinfo->output_width, input_data, output_data_ptr);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor,
michael@0:         cinfo->output_width, input_data, output_data_ptr);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v2_fancy_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v1_fancy_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
michael@0:                            jpeg_component_info * compptr, 
michael@0:                            JSAMPARRAY input_data,
michael@0:                            JSAMPARRAY * output_data_ptr)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0:     jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
michael@0:         compptr->downsampled_width, input_data, output_data_ptr);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor,
michael@0:         compptr->downsampled_width, input_data, output_data_ptr);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
michael@0:                            jpeg_component_info * compptr, 
michael@0:                            JSAMPARRAY input_data,
michael@0:                            JSAMPARRAY * output_data_ptr)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0:     jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
michael@0:         compptr->downsampled_width, input_data, output_data_ptr);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor,
michael@0:         compptr->downsampled_width, input_data, output_data_ptr);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v2_merged_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_h2v1_merged_upsample (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
michael@0:                             JSAMPIMAGE input_buf,
michael@0:                             JDIMENSION in_row_group_ctr,
michael@0:                             JSAMPARRAY output_buf)
michael@0: {
michael@0:   void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0:   void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0: 
michael@0:   switch(cinfo->out_color_space)
michael@0:   {
michael@0:     case JCS_EXT_RGB:
michael@0:       sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_RGBX:
michael@0:     case JCS_EXT_RGBA:
michael@0:       sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGR:
michael@0:       sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extbgr_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGRX:
michael@0:     case JCS_EXT_BGRA:
michael@0:       sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XBGR:
michael@0:     case JCS_EXT_ABGR:
michael@0:       sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XRGB:
michael@0:     case JCS_EXT_ARGB:
michael@0:       sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx;
michael@0:       break;
michael@0:     default:
michael@0:       sse2fct=jsimd_h2v2_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v2_merged_upsample_mmx;
michael@0:       break;
michael@0:   }
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0:     sse2fct(cinfo->output_width, input_buf,
michael@0:         in_row_group_ctr, output_buf);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     mmxfct(cinfo->output_width, input_buf,
michael@0:         in_row_group_ctr, output_buf);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
michael@0:                             JSAMPIMAGE input_buf,
michael@0:                             JDIMENSION in_row_group_ctr,
michael@0:                             JSAMPARRAY output_buf)
michael@0: {
michael@0:   void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0:   void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0: 
michael@0:   switch(cinfo->out_color_space)
michael@0:   {
michael@0:     case JCS_EXT_RGB:
michael@0:       sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_RGBX:
michael@0:     case JCS_EXT_RGBA:
michael@0:       sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGR:
michael@0:       sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extbgr_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_BGRX:
michael@0:     case JCS_EXT_BGRA:
michael@0:       sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XBGR:
michael@0:     case JCS_EXT_ABGR:
michael@0:       sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx;
michael@0:       break;
michael@0:     case JCS_EXT_XRGB:
michael@0:     case JCS_EXT_ARGB:
michael@0:       sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx;
michael@0:       break;
michael@0:     default:
michael@0:       sse2fct=jsimd_h2v1_merged_upsample_sse2;
michael@0:       mmxfct=jsimd_h2v1_merged_upsample_mmx;
michael@0:       break;
michael@0:   }
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) &&
michael@0:       IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0:     sse2fct(cinfo->output_width, input_buf,
michael@0:         in_row_group_ctr, output_buf);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     mmxfct(cinfo->output_width, input_buf,
michael@0:         in_row_group_ctr, output_buf);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_convsamp (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(DCTELEM) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_convsamp_float (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(FAST_FLOAT) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_SSE)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_3DNOW)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0:                 DCTELEM * workspace)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_convsamp_sse2(sample_data, start_col, workspace);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_convsamp_mmx(sample_data, start_col, workspace);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0:                       FAST_FLOAT * workspace)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
michael@0:   else if (simd_support & JSIMD_SSE)
michael@0:     jsimd_convsamp_float_sse(sample_data, start_col, workspace);
michael@0:   else if (simd_support & JSIMD_3DNOW)
michael@0:     jsimd_convsamp_float_3dnow(sample_data, start_col, workspace);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_fdct_islow (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(DCTELEM) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_fdct_ifast (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(DCTELEM) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_fdct_float (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(FAST_FLOAT) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_3DNOW)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_fdct_islow (DCTELEM * data)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0:     jsimd_fdct_islow_sse2(data);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_fdct_islow_mmx(data);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_fdct_ifast (DCTELEM * data)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0:     jsimd_fdct_ifast_sse2(data);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_fdct_ifast_mmx(data);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_fdct_float (FAST_FLOAT * data)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
michael@0:     jsimd_fdct_float_sse(data);
michael@0:   else if (simd_support & JSIMD_3DNOW)
michael@0:     jsimd_fdct_float_3dnow(data);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_quantize (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (sizeof(DCTELEM) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_quantize_float (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (sizeof(FAST_FLOAT) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_SSE)
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_3DNOW)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
michael@0:                 DCTELEM * workspace)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_quantize_sse2(coef_block, divisors, workspace);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_quantize_mmx(coef_block, divisors, workspace);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
michael@0:                       FAST_FLOAT * workspace)
michael@0: {
michael@0:   if (simd_support & JSIMD_SSE2)
michael@0:     jsimd_quantize_float_sse2(coef_block, divisors, workspace);
michael@0:   else if (simd_support & JSIMD_SSE)
michael@0:     jsimd_quantize_float_sse(coef_block, divisors, workspace);
michael@0:   else if (simd_support & JSIMD_3DNOW)
michael@0:     jsimd_quantize_float_3dnow(coef_block, divisors, workspace);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_idct_2x2 (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_idct_4x4 (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0:                 JDIMENSION output_col)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0:     jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0:                 JDIMENSION output_col)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0:     jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_idct_islow (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_idct_ifast (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   /* The code is optimised for these values only */
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(IFAST_MULT_TYPE) != 2)
michael@0:     return 0;
michael@0:   if (IFAST_SCALE_BITS != 2)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_MMX)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(int)
michael@0: jsimd_can_idct_float (void)
michael@0: {
michael@0:   init_simd();
michael@0: 
michael@0:   if (DCTSIZE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JCOEF) != 2)
michael@0:     return 0;
michael@0:   if (BITS_IN_JSAMPLE != 8)
michael@0:     return 0;
michael@0:   if (sizeof(JDIMENSION) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(FAST_FLOAT) != 4)
michael@0:     return 0;
michael@0:   if (sizeof(FLOAT_MULT_TYPE) != 4)
michael@0:     return 0;
michael@0: 
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
michael@0:     return 1;
michael@0:   if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
michael@0:     return 1;
michael@0:   if (simd_support & JSIMD_3DNOW)
michael@0:     return 1;
michael@0: 
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0:                 JDIMENSION output_col)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
michael@0:     jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0:                 JDIMENSION output_col)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
michael@0:     jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0:   else if (simd_support & JSIMD_MMX)
michael@0:     jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0: }
michael@0: 
michael@0: GLOBAL(void)
michael@0: jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0:                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0:                 JDIMENSION output_col)
michael@0: {
michael@0:   if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
michael@0:     jsimd_idct_float_sse2(compptr->dct_table, coef_block,
michael@0:         output_buf, output_col);
michael@0:   else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
michael@0:     jsimd_idct_float_sse(compptr->dct_table, coef_block,
michael@0:         output_buf, output_col);
michael@0:   else if (simd_support & JSIMD_3DNOW)
michael@0:     jsimd_idct_float_3dnow(compptr->dct_table, coef_block,
michael@0:         output_buf, output_col);
michael@0: }
michael@0: