FreeRDP/libfreerdp/codec/rfx_sse2.c

/**
 * FreeRDP: A Remote Desktop Protocol Implementation
 * RemoteFX Codec Library - SSE2 Optimizations
 *
 * Copyright 2011 Stephen Erisman
 * Copyright 2011 Norbert Federa <nfedera@thinstuff.com>
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include <xmmintrin.h>
#include <emmintrin.h>

#include "rfx_types.h"
#include "rfx_sse2.h"

#ifdef _MSC_VER
#define	__attribute__(...)
#endif

#define CACHE_LINE_BYTES	64

#define _mm_between_epi16(_val, _min, _max) \
	do { _val = _mm_min_epi16(_max, _mm_max_epi16(_val, _min)); } while (0)

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_prefetch_buffer(char * buffer, int num_bytes)
{
	__m128i * buf = (__m128i*) buffer;
	unsigned int i;
	for (i = 0; i < (num_bytes / sizeof(__m128i)); i+=(CACHE_LINE_BYTES / sizeof(__m128i)))
	{
		_mm_prefetch((char*)(&buf[i]), _MM_HINT_NTA);
	}
}

static void rfx_decode_ycbcr_to_rgb_sse2(INT16* y_r_buffer, INT16* cb_g_buffer, INT16* cr_b_buffer)
{	
	__m128i zero = _mm_setzero_si128();
	__m128i max = _mm_set1_epi16(255);

	__m128i* y_r_buf = (__m128i*) y_r_buffer;
	__m128i* cb_g_buf = (__m128i*) cb_g_buffer;
	__m128i* cr_b_buf = (__m128i*) cr_b_buffer;

	__m128i y;
	__m128i cr;
	__m128i cb;
	__m128i r;
	__m128i g;
	__m128i b;

	int i;

	__m128i r_cr = _mm_set1_epi16(22986);	//  1.403 << 14
	__m128i g_cb = _mm_set1_epi16(-5636);	// -0.344 << 14
	__m128i g_cr = _mm_set1_epi16(-11698);	// -0.714 << 14
	__m128i b_cb = _mm_set1_epi16(28999);	//  1.770 << 14
	__m128i c4096 = _mm_set1_epi16(4096);

	for (i = 0; i < (4096 * sizeof(INT16) / sizeof(__m128i)); i += (CACHE_LINE_BYTES / sizeof(__m128i)))
	{
		_mm_prefetch((char*)(&y_r_buf[i]), _MM_HINT_NTA);
		_mm_prefetch((char*)(&cb_g_buf[i]), _MM_HINT_NTA);
		_mm_prefetch((char*)(&cr_b_buf[i]), _MM_HINT_NTA);
	}
	for (i = 0; i < (4096 * sizeof(INT16) / sizeof(__m128i)); i++)
	{
		/*
		In order to use SSE2 signed 16-bit integer multiplication we need to convert
		the floating point factors to signed int without loosing information.
		The result of this multiplication is 32 bit and we have two SSE instructions
		that return either the hi or lo word.
		Thus we will multiply the factors by the highest possible 2^n, take the 
		upper 16 bits of the signed 32-bit result (_mm_mulhi_epi16) and correct	this
		result by multiplying it by 2^(16-n).
		For the given factors in the conversion matrix the best possible n is 14.

		Example for calculating r:
		r = (y>>5) + 128 + (cr*1.403)>>5                       // our base formula
		r = (y>>5) + 128 + (HIWORD(cr*(1.403<<14)<<2))>>5      // see above
		r = (y+4096)>>5 + (HIWORD(cr*22986)<<2)>>5             // simplification
		r = ((y+4096)>>2 + HIWORD(cr*22986)) >> 3
		*/

		/* y = (y_r_buf[i] + 4096) >> 2 */
		y = _mm_load_si128(&y_r_buf[i]);
		y = _mm_add_epi16(y, c4096);
		y = _mm_srai_epi16(y, 2);
		/* cb = cb_g_buf[i]; */
		cb = _mm_load_si128(&cb_g_buf[i]);
		/* cr = cr_b_buf[i]; */
		cr = _mm_load_si128(&cr_b_buf[i]);

		/* (y + HIWORD(cr*22986)) >> 3 */
		r = _mm_add_epi16(y, _mm_mulhi_epi16(cr, r_cr));
		r = _mm_srai_epi16(r, 3);
		/* y_r_buf[i] = MINMAX(r, 0, 255); */
		_mm_between_epi16(r, zero, max);
		_mm_store_si128(&y_r_buf[i], r);

		/* (y + HIWORD(cb*-5636) + HIWORD(cr*-11698)) >> 3 */
		g = _mm_add_epi16(y, _mm_mulhi_epi16(cb, g_cb));
		g = _mm_add_epi16(g, _mm_mulhi_epi16(cr, g_cr));
		g = _mm_srai_epi16(g, 3);
		/* cb_g_buf[i] = MINMAX(g, 0, 255); */
		_mm_between_epi16(g, zero, max);
		_mm_store_si128(&cb_g_buf[i], g);

		/* (y + HIWORD(cb*28999)) >> 3 */
		b = _mm_add_epi16(y, _mm_mulhi_epi16(cb, b_cb));
		b = _mm_srai_epi16(b, 3);
		/* cr_b_buf[i] = MINMAX(b, 0, 255); */
		_mm_between_epi16(b, zero, max);
		_mm_store_si128(&cr_b_buf[i], b);
	}
}

/* The encodec YCbCr coeffectients are represented as 11.5 fixed-point numbers. See rfx_encode.c */
static void rfx_encode_rgb_to_ycbcr_sse2(INT16* y_r_buffer, INT16* cb_g_buffer, INT16* cr_b_buffer)
{
	__m128i min = _mm_set1_epi16(-128 << 5);
	__m128i max = _mm_set1_epi16(127 << 5);

	__m128i* y_r_buf = (__m128i*) y_r_buffer;
	__m128i* cb_g_buf = (__m128i*) cb_g_buffer;
	__m128i* cr_b_buf = (__m128i*) cr_b_buffer;

	__m128i y;
	__m128i cr;
	__m128i cb;
	__m128i r;
	__m128i g;
	__m128i b;

	__m128i y_r  = _mm_set1_epi16(9798);   //  0.299000 << 15
	__m128i y_g  = _mm_set1_epi16(19235);  //  0.587000 << 15
	__m128i y_b  = _mm_set1_epi16(3735);   //  0.114000 << 15
	__m128i cb_r = _mm_set1_epi16(-5535);  // -0.168935 << 15
	__m128i cb_g = _mm_set1_epi16(-10868); // -0.331665 << 15
	__m128i cb_b = _mm_set1_epi16(16403);  //  0.500590 << 15
	__m128i cr_r = _mm_set1_epi16(16377);  //  0.499813 << 15
	__m128i cr_g = _mm_set1_epi16(-13714); // -0.418531 << 15
	__m128i cr_b = _mm_set1_epi16(-2663);  // -0.081282 << 15

	int i;

	for (i = 0; i < (4096 * sizeof(INT16) / sizeof(__m128i)); i += (CACHE_LINE_BYTES / sizeof(__m128i)))
	{
		_mm_prefetch((char*)(&y_r_buf[i]), _MM_HINT_NTA);
		_mm_prefetch((char*)(&cb_g_buf[i]), _MM_HINT_NTA);
		_mm_prefetch((char*)(&cr_b_buf[i]), _MM_HINT_NTA);
	}
	for (i = 0; i < (4096 * sizeof(INT16) / sizeof(__m128i)); i++)
	{
		/*
		In order to use SSE2 signed 16-bit integer multiplication we need to convert
		the floating point factors to signed int without loosing information.
		The result of this multiplication is 32 bit and using SSE2 we get either the
		product's hi or lo word.
		Thus we will multiply the factors by the highest possible 2^n and take the
		upper 16 bits of the signed 32-bit result (_mm_mulhi_epi16).
		Since the final result needs to be scaled by << 5 and also in in order to keep
		the precision within the upper 16 bits we will also have to scale the RGB
		values used in the multiplication by << 5+(16-n).
		*/

		/* r = y_r_buf[i]; */
		r = _mm_load_si128(&y_r_buf[i]);

		/* g = cb_g_buf[i]; */
		g = _mm_load_si128(&cb_g_buf[i]);

		/* b = cr_b_buf[i]; */
		b = _mm_load_si128(&cr_b_buf[i]);

		/* r<<6; g<<6; b<<6 */
		r = _mm_slli_epi16(r, 6);
		g = _mm_slli_epi16(g, 6);
		b = _mm_slli_epi16(b, 6);

		/* y = HIWORD(r*y_r) + HIWORD(g*y_g) + HIWORD(b*y_b) + min */
		y = _mm_mulhi_epi16(r, y_r);
		y = _mm_add_epi16(y, _mm_mulhi_epi16(g, y_g));
		y = _mm_add_epi16(y, _mm_mulhi_epi16(b, y_b));
		y = _mm_add_epi16(y, min);
		/* y_r_buf[i] = MINMAX(y, 0, (255 << 5)) - (128 << 5); */
		_mm_between_epi16(y, min, max);
		_mm_store_si128(&y_r_buf[i], y);

		/* cb = HIWORD(r*cb_r) + HIWORD(g*cb_g) + HIWORD(b*cb_b) */
		cb = _mm_mulhi_epi16(r, cb_r);
		cb = _mm_add_epi16(cb, _mm_mulhi_epi16(g, cb_g));
		cb = _mm_add_epi16(cb, _mm_mulhi_epi16(b, cb_b));
		/* cb_g_buf[i] = MINMAX(cb, (-128 << 5), (127 << 5)); */
		_mm_between_epi16(cb, min, max);
		_mm_store_si128(&cb_g_buf[i], cb);

		/* cr = HIWORD(r*cr_r) + HIWORD(g*cr_g) + HIWORD(b*cr_b) */
		cr = _mm_mulhi_epi16(r, cr_r);
		cr = _mm_add_epi16(cr, _mm_mulhi_epi16(g, cr_g));
		cr = _mm_add_epi16(cr, _mm_mulhi_epi16(b, cr_b));
		/* cr_b_buf[i] = MINMAX(cr, (-128 << 5), (127 << 5)); */
		_mm_between_epi16(cr, min, max);
		_mm_store_si128(&cr_b_buf[i], cr);
	}
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_quantization_decode_block_sse2(INT16* buffer, const int buffer_size, const UINT32 factor)
{
	__m128i a;
	__m128i * ptr = (__m128i*) buffer;
	__m128i * buf_end = (__m128i*) (buffer + buffer_size);

	if (factor == 0)
		return;

	do
	{
		a = _mm_load_si128(ptr);
		a = _mm_slli_epi16(a, factor);
		_mm_store_si128(ptr, a);

		ptr++;
	} while(ptr < buf_end);
}

static void rfx_quantization_decode_sse2(INT16* buffer, const UINT32* quantization_values)
{
	_mm_prefetch_buffer((char*) buffer, 4096 * sizeof(INT16));

	rfx_quantization_decode_block_sse2(buffer, 4096, 5);

	rfx_quantization_decode_block_sse2(buffer, 1024, quantization_values[8] - 6); /* HL1 */
	rfx_quantization_decode_block_sse2(buffer + 1024, 1024, quantization_values[7] - 6); /* LH1 */
	rfx_quantization_decode_block_sse2(buffer + 2048, 1024, quantization_values[9] - 6); /* HH1 */
	rfx_quantization_decode_block_sse2(buffer + 3072, 256, quantization_values[5] - 6); /* HL2 */
	rfx_quantization_decode_block_sse2(buffer + 3328, 256, quantization_values[4] - 6); /* LH2 */
	rfx_quantization_decode_block_sse2(buffer + 3584, 256, quantization_values[6] - 6); /* HH2 */
	rfx_quantization_decode_block_sse2(buffer + 3840, 64, quantization_values[2] - 6); /* HL3 */
	rfx_quantization_decode_block_sse2(buffer + 3904, 64, quantization_values[1] - 6); /* LH3 */
	rfx_quantization_decode_block_sse2(buffer + 3968, 64, quantization_values[3] - 6); /* HH3 */
	rfx_quantization_decode_block_sse2(buffer + 4032, 64, quantization_values[0] - 6); /* LL3 */
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_quantization_encode_block_sse2(INT16* buffer, const int buffer_size, const UINT32 factor)
{
	__m128i a;
	__m128i* ptr = (__m128i*) buffer;
	__m128i* buf_end = (__m128i*) (buffer + buffer_size);
	__m128i half;

	if (factor == 0)
		return;

	half = _mm_set1_epi16(1 << (factor - 1));
	do
	{
		a = _mm_load_si128(ptr);
		a = _mm_add_epi16(a, half);
		a = _mm_srai_epi16(a, factor);
		_mm_store_si128(ptr, a);

		ptr++;
	} while(ptr < buf_end);
}

static void rfx_quantization_encode_sse2(INT16* buffer, const UINT32* quantization_values)
{
	_mm_prefetch_buffer((char*) buffer, 4096 * sizeof(INT16));

	rfx_quantization_encode_block_sse2(buffer, 1024, quantization_values[8] - 6); /* HL1 */
	rfx_quantization_encode_block_sse2(buffer + 1024, 1024, quantization_values[7] - 6); /* LH1 */
	rfx_quantization_encode_block_sse2(buffer + 2048, 1024, quantization_values[9] - 6); /* HH1 */
	rfx_quantization_encode_block_sse2(buffer + 3072, 256, quantization_values[5] - 6); /* HL2 */
	rfx_quantization_encode_block_sse2(buffer + 3328, 256, quantization_values[4] - 6); /* LH2 */
	rfx_quantization_encode_block_sse2(buffer + 3584, 256, quantization_values[6] - 6); /* HH2 */
	rfx_quantization_encode_block_sse2(buffer + 3840, 64, quantization_values[2] - 6); /* HL3 */
	rfx_quantization_encode_block_sse2(buffer + 3904, 64, quantization_values[1] - 6); /* LH3 */
	rfx_quantization_encode_block_sse2(buffer + 3968, 64, quantization_values[3] - 6); /* HH3 */
	rfx_quantization_encode_block_sse2(buffer + 4032, 64, quantization_values[0] - 6); /* LL3 */

	rfx_quantization_encode_block_sse2(buffer, 4096, 5);
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_horiz_sse2(INT16* l, INT16* h, INT16* dst, int subband_width)
{
	int y, n;
	INT16* l_ptr = l;
	INT16* h_ptr = h;
	INT16* dst_ptr = dst;
	int first;
	int last;
	__m128i l_n;
	__m128i h_n;
	__m128i h_n_m;
	__m128i tmp_n;
	__m128i dst_n;
	__m128i dst_n_p;
	__m128i dst1;
	__m128i dst2;

	for (y = 0; y < subband_width; y++)
	{
		/* Even coefficients */
		for (n = 0; n < subband_width; n+=8)
		{
			/* dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); */
			
			l_n = _mm_load_si128((__m128i*) l_ptr);

			h_n = _mm_load_si128((__m128i*) h_ptr);
			h_n_m = _mm_loadu_si128((__m128i*) (h_ptr - 1));
			if (n == 0)
			{
				first = _mm_extract_epi16(h_n_m, 1);
				h_n_m = _mm_insert_epi16(h_n_m, first, 0);
			}
			
			tmp_n = _mm_add_epi16(h_n, h_n_m);
			tmp_n = _mm_add_epi16(tmp_n, _mm_set1_epi16(1));
			tmp_n = _mm_srai_epi16(tmp_n, 1);
			
			dst_n = _mm_sub_epi16(l_n, tmp_n);
			
			_mm_store_si128((__m128i*) l_ptr, dst_n);
			
			l_ptr+=8;
			h_ptr+=8;
		}
		l_ptr -= subband_width;
		h_ptr -= subband_width;
		
		/* Odd coefficients */
		for (n = 0; n < subband_width; n+=8)
		{
			/* dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); */
			
			h_n = _mm_load_si128((__m128i*) h_ptr);
			
			h_n = _mm_slli_epi16(h_n, 1);
			
			dst_n = _mm_load_si128((__m128i*) (l_ptr));
			dst_n_p = _mm_loadu_si128((__m128i*) (l_ptr + 1));
			if (n == subband_width - 8)
			{
				last = _mm_extract_epi16(dst_n_p, 6);
				dst_n_p = _mm_insert_epi16(dst_n_p, last, 7);
			}
			
			tmp_n = _mm_add_epi16(dst_n_p, dst_n);
			tmp_n = _mm_srai_epi16(tmp_n, 1);
			
			tmp_n = _mm_add_epi16(tmp_n, h_n);
			
			dst1 = _mm_unpacklo_epi16(dst_n, tmp_n);
			dst2 = _mm_unpackhi_epi16(dst_n, tmp_n);
			
			_mm_store_si128((__m128i*) dst_ptr, dst1);
			_mm_store_si128((__m128i*) (dst_ptr + 8), dst2);
			
			l_ptr+=8;
			h_ptr+=8;
			dst_ptr+=16;
		}
	}
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_vert_sse2(INT16* l, INT16* h, INT16* dst, int subband_width)
{
	int x, n;
	INT16* l_ptr = l;
	INT16* h_ptr = h;
	INT16* dst_ptr = dst;
	__m128i l_n;
	__m128i h_n;
	__m128i tmp_n;
	__m128i h_n_m;
	__m128i dst_n;
	__m128i dst_n_m;
	__m128i dst_n_p;
	
	int total_width = subband_width + subband_width;

	/* Even coefficients */
	for (n = 0; n < subband_width; n++)
	{
		for (x = 0; x < total_width; x+=8)
		{
			/* dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); */
			
			l_n = _mm_load_si128((__m128i*) l_ptr);
			h_n = _mm_load_si128((__m128i*) h_ptr);
			
			tmp_n = _mm_add_epi16(h_n, _mm_set1_epi16(1));;
			if (n == 0)
				tmp_n = _mm_add_epi16(tmp_n, h_n);
			else
			{
				h_n_m = _mm_loadu_si128((__m128i*) (h_ptr - total_width));
				tmp_n = _mm_add_epi16(tmp_n, h_n_m);
			}
			tmp_n = _mm_srai_epi16(tmp_n, 1);
			
			dst_n = _mm_sub_epi16(l_n, tmp_n);
			_mm_store_si128((__m128i*) dst_ptr, dst_n);
			
			l_ptr+=8;
			h_ptr+=8;
			dst_ptr+=8;
		}
		dst_ptr+=total_width;
	}
	
	h_ptr = h;
	dst_ptr = dst + total_width;
	
	/* Odd coefficients */
	for (n = 0; n < subband_width; n++)
	{
		for (x = 0; x < total_width; x+=8)
		{
			/* dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); */
			
			h_n = _mm_load_si128((__m128i*) h_ptr);
			dst_n_m = _mm_load_si128((__m128i*) (dst_ptr - total_width));
			h_n = _mm_slli_epi16(h_n, 1);
			
			tmp_n = dst_n_m;
			if (n == subband_width - 1)
				tmp_n = _mm_add_epi16(tmp_n, dst_n_m);
			else
			{
				dst_n_p = _mm_loadu_si128((__m128i*) (dst_ptr + total_width));
				tmp_n = _mm_add_epi16(tmp_n, dst_n_p);
			}
			tmp_n = _mm_srai_epi16(tmp_n, 1);
			
			dst_n = _mm_add_epi16(tmp_n, h_n);
			_mm_store_si128((__m128i*) dst_ptr, dst_n);

			h_ptr+=8;
			dst_ptr+=8;
		}
		dst_ptr+=total_width;
	}
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_sse2(INT16* buffer, INT16* idwt, int subband_width)
{
	INT16 *hl, *lh, *hh, *ll;
	INT16 *l_dst, *h_dst;

	_mm_prefetch_buffer((char*) idwt, subband_width * 4 * sizeof(INT16));

	/* Inverse DWT in horizontal direction, results in 2 sub-bands in L, H order in tmp buffer idwt. */
	/* The 4 sub-bands are stored in HL(0), LH(1), HH(2), LL(3) order. */
	/* The lower part L uses LL(3) and HL(0). */
	/* The higher part H uses LH(1) and HH(2). */

	ll = buffer + subband_width * subband_width * 3;
	hl = buffer;
	l_dst = idwt;

	rfx_dwt_2d_decode_block_horiz_sse2(ll, hl, l_dst, subband_width);

	lh = buffer + subband_width * subband_width;
	hh = buffer + subband_width * subband_width * 2;
	h_dst = idwt + subband_width * subband_width * 2;
	
	rfx_dwt_2d_decode_block_horiz_sse2(lh, hh, h_dst, subband_width);

	/* Inverse DWT in vertical direction, results are stored in original buffer. */
	rfx_dwt_2d_decode_block_vert_sse2(l_dst, h_dst, buffer, subband_width);
}

static void rfx_dwt_2d_decode_sse2(INT16* buffer, INT16* dwt_buffer)
{
	_mm_prefetch_buffer((char*) buffer, 4096 * sizeof(INT16));
	
	rfx_dwt_2d_decode_block_sse2(buffer + 3840, dwt_buffer, 8);
	rfx_dwt_2d_decode_block_sse2(buffer + 3072, dwt_buffer, 16);
	rfx_dwt_2d_decode_block_sse2(buffer, dwt_buffer, 32);
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_encode_block_vert_sse2(INT16* src, INT16* l, INT16* h, int subband_width)
{
	int total_width;
	int x;
	int n;
	__m128i src_2n;
	__m128i src_2n_1;
	__m128i src_2n_2;
	__m128i h_n;
	__m128i h_n_m;
	__m128i l_n;

	total_width = subband_width << 1;

	for (n = 0; n < subband_width; n++)
	{
		for (x = 0; x < total_width; x += 8)
		{
			src_2n = _mm_load_si128((__m128i*) src);
			src_2n_1 = _mm_load_si128((__m128i*) (src + total_width));
			if (n < subband_width - 1)
				src_2n_2 = _mm_load_si128((__m128i*) (src + 2 * total_width));
			else
				src_2n_2 = src_2n;

			/* h[n] = (src[2n + 1] - ((src[2n] + src[2n + 2]) >> 1)) >> 1 */

			h_n = _mm_add_epi16(src_2n, src_2n_2);
			h_n = _mm_srai_epi16(h_n, 1);
			h_n = _mm_sub_epi16(src_2n_1, h_n);
			h_n = _mm_srai_epi16(h_n, 1);

			_mm_store_si128((__m128i*) h, h_n);

			if (n == 0)
				h_n_m = h_n;
			else
				h_n_m = _mm_load_si128((__m128i*) (h - total_width));

			/* l[n] = src[2n] + ((h[n - 1] + h[n]) >> 1) */

			l_n = _mm_add_epi16(h_n_m, h_n);
			l_n = _mm_srai_epi16(l_n, 1);
			l_n = _mm_add_epi16(l_n, src_2n);

			_mm_store_si128((__m128i*) l, l_n);

			src += 8;
			l += 8;
			h += 8;
		}
		src += total_width;
	}
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_encode_block_horiz_sse2(INT16* src, INT16* l, INT16* h, int subband_width)
{
	int y;
	int n;
	int first;
	__m128i src_2n;
	__m128i src_2n_1;
	__m128i src_2n_2;
	__m128i h_n;
	__m128i h_n_m;
	__m128i l_n;

	for (y = 0; y < subband_width; y++)
	{
		for (n = 0; n < subband_width; n += 8)
		{
			/* The following 3 Set operations consumes more than half of the total DWT processing time! */
			src_2n = _mm_set_epi16(src[14], src[12], src[10], src[8], src[6], src[4], src[2], src[0]);
			src_2n_1 = _mm_set_epi16(src[15], src[13], src[11], src[9], src[7], src[5], src[3], src[1]);
			src_2n_2 = _mm_set_epi16(n == subband_width - 8 ? src[14] : src[16],
				src[14], src[12], src[10], src[8], src[6], src[4], src[2]);

			/* h[n] = (src[2n + 1] - ((src[2n] + src[2n + 2]) >> 1)) >> 1 */

			h_n = _mm_add_epi16(src_2n, src_2n_2);
			h_n = _mm_srai_epi16(h_n, 1);
			h_n = _mm_sub_epi16(src_2n_1, h_n);
			h_n = _mm_srai_epi16(h_n, 1);

			_mm_store_si128((__m128i*) h, h_n);

			h_n_m = _mm_loadu_si128((__m128i*) (h - 1));
			if (n == 0)
			{
				first = _mm_extract_epi16(h_n_m, 1);
				h_n_m = _mm_insert_epi16(h_n_m, first, 0);
			}

			/* l[n] = src[2n] + ((h[n - 1] + h[n]) >> 1) */

			l_n = _mm_add_epi16(h_n_m, h_n);
			l_n = _mm_srai_epi16(l_n, 1);
			l_n = _mm_add_epi16(l_n, src_2n);

			_mm_store_si128((__m128i*) l, l_n);

			src += 16;
			l += 8;
			h += 8;
		}
	}
}

static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_encode_block_sse2(INT16* buffer, INT16* dwt, int subband_width)
{
	INT16 *hl, *lh, *hh, *ll;
	INT16 *l_src, *h_src;

	_mm_prefetch_buffer((char*) dwt, subband_width * 4 * sizeof(INT16));

	/* DWT in vertical direction, results in 2 sub-bands in L, H order in tmp buffer dwt. */

	l_src = dwt;
	h_src = dwt + subband_width * subband_width * 2;

	rfx_dwt_2d_encode_block_vert_sse2(buffer, l_src, h_src, subband_width);

	/* DWT in horizontal direction, results in 4 sub-bands in HL(0), LH(1), HH(2), LL(3) order, stored in original buffer. */
	/* The lower part L generates LL(3) and HL(0). */
	/* The higher part H generates LH(1) and HH(2). */

	ll = buffer + subband_width * subband_width * 3;
	hl = buffer;

	lh = buffer + subband_width * subband_width;
	hh = buffer + subband_width * subband_width * 2;

	rfx_dwt_2d_encode_block_horiz_sse2(l_src, ll, hl, subband_width);
	rfx_dwt_2d_encode_block_horiz_sse2(h_src, lh, hh, subband_width);
}

static void rfx_dwt_2d_encode_sse2(INT16* buffer, INT16* dwt_buffer)
{
	_mm_prefetch_buffer((char*) buffer, 4096 * sizeof(INT16));
	
	rfx_dwt_2d_encode_block_sse2(buffer, dwt_buffer, 32);
	rfx_dwt_2d_encode_block_sse2(buffer + 3072, dwt_buffer, 16);
	rfx_dwt_2d_encode_block_sse2(buffer + 3840, dwt_buffer, 8);
}

void rfx_init_sse2(RFX_CONTEXT* context)
{
	DEBUG_RFX("Using SSE2 optimizations");

	IF_PROFILER(context->priv->prof_rfx_decode_ycbcr_to_rgb->name = "rfx_decode_ycbcr_to_rgb_sse2");
	IF_PROFILER(context->priv->prof_rfx_encode_rgb_to_ycbcr->name = "rfx_encode_rgb_to_ycbcr_sse2");
	IF_PROFILER(context->priv->prof_rfx_quantization_decode->name = "rfx_quantization_decode_sse2");
	IF_PROFILER(context->priv->prof_rfx_quantization_encode->name = "rfx_quantization_encode_sse2");
	IF_PROFILER(context->priv->prof_rfx_dwt_2d_decode->name = "rfx_dwt_2d_decode_sse2");
	IF_PROFILER(context->priv->prof_rfx_dwt_2d_encode->name = "rfx_dwt_2d_encode_sse2");

	context->decode_ycbcr_to_rgb = rfx_decode_ycbcr_to_rgb_sse2;
	context->encode_rgb_to_ycbcr = rfx_encode_rgb_to_ycbcr_sse2;
	context->quantization_decode = rfx_quantization_decode_sse2;
	context->quantization_encode = rfx_quantization_encode_sse2;
	context->dwt_2d_decode = rfx_dwt_2d_decode_sse2;
	context->dwt_2d_encode = rfx_dwt_2d_encode_sse2;
}