michael@0: /*
michael@0:  *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS.  All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: 
michael@0: #include "vpx_config.h"
michael@0: #include "vp8_rtcd.h"
michael@0: #include "vpx_ports/x86.h"
michael@0: #include "vpx_mem/vpx_mem.h"
michael@0: #include "vp8/encoder/block.h"
michael@0: #include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
michael@0: 
michael@0: #include <mmintrin.h> /* MMX */
michael@0: #include <xmmintrin.h> /* SSE */
michael@0: #include <emmintrin.h> /* SSE2 */
michael@0: 
michael@0: #define SELECT_EOB(i, z) \
michael@0:     do { \
michael@0:         short boost = *zbin_boost_ptr; \
michael@0:         int cmp = (x[z] < boost) | (y[z] == 0); \
michael@0:         zbin_boost_ptr++; \
michael@0:         if (cmp) \
michael@0:             goto select_eob_end_##i; \
michael@0:         qcoeff_ptr[z] = y[z]; \
michael@0:         eob = i; \
michael@0:         zbin_boost_ptr = b->zrun_zbin_boost; \
michael@0:         select_eob_end_##i:; \
michael@0:     } while (0)
michael@0: 
michael@0: void vp8_regular_quantize_b_sse2(BLOCK *b, BLOCKD *d)
michael@0: {
michael@0:     char eob = 0;
michael@0:     short *zbin_boost_ptr  = b->zrun_zbin_boost;
michael@0:     short *qcoeff_ptr      = d->qcoeff;
michael@0:     DECLARE_ALIGNED_ARRAY(16, short, x, 16);
michael@0:     DECLARE_ALIGNED_ARRAY(16, short, y, 16);
michael@0: 
michael@0:     __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1;
michael@0:     __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
michael@0:     __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
michael@0:     __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
michael@0:     __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
michael@0:     __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
michael@0:     __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
michael@0:     __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
michael@0:     __m128i round0 = _mm_load_si128((__m128i *)(b->round));
michael@0:     __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
michael@0:     __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
michael@0:     __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
michael@0:     __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
michael@0:     __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
michael@0: 
michael@0:     vpx_memset(qcoeff_ptr, 0, 32);
michael@0: 
michael@0:     /* Duplicate to all lanes. */
michael@0:     zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
michael@0:     zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
michael@0: 
michael@0:     /* Sign of z: z >> 15 */
michael@0:     sz0 = _mm_srai_epi16(z0, 15);
michael@0:     sz1 = _mm_srai_epi16(z1, 15);
michael@0: 
michael@0:     /* x = abs(z): (z ^ sz) - sz */
michael@0:     x0 = _mm_xor_si128(z0, sz0);
michael@0:     x1 = _mm_xor_si128(z1, sz1);
michael@0:     x0 = _mm_sub_epi16(x0, sz0);
michael@0:     x1 = _mm_sub_epi16(x1, sz1);
michael@0: 
michael@0:     /* zbin[] + zbin_extra */
michael@0:     zbin0 = _mm_add_epi16(zbin0, zbin_extra);
michael@0:     zbin1 = _mm_add_epi16(zbin1, zbin_extra);
michael@0: 
michael@0:     /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
michael@0:      * the equation because boost is the only value which can change:
michael@0:      * x - (zbin[] + extra) >= boost */
michael@0:     x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
michael@0:     x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
michael@0: 
michael@0:     _mm_store_si128((__m128i *)(x), x_minus_zbin0);
michael@0:     _mm_store_si128((__m128i *)(x + 8), x_minus_zbin1);
michael@0: 
michael@0:     /* All the remaining calculations are valid whether they are done now with
michael@0:      * simd or later inside the loop one at a time. */
michael@0:     x0 = _mm_add_epi16(x0, round0);
michael@0:     x1 = _mm_add_epi16(x1, round1);
michael@0: 
michael@0:     y0 = _mm_mulhi_epi16(x0, quant0);
michael@0:     y1 = _mm_mulhi_epi16(x1, quant1);
michael@0: 
michael@0:     y0 = _mm_add_epi16(y0, x0);
michael@0:     y1 = _mm_add_epi16(y1, x1);
michael@0: 
michael@0:     /* Instead of shifting each value independently we convert the scaling
michael@0:      * factor with 1 << (16 - shift) so we can use multiply/return high half. */
michael@0:     y0 = _mm_mulhi_epi16(y0, quant_shift0);
michael@0:     y1 = _mm_mulhi_epi16(y1, quant_shift1);
michael@0: 
michael@0:     /* Return the sign: (y ^ sz) - sz */
michael@0:     y0 = _mm_xor_si128(y0, sz0);
michael@0:     y1 = _mm_xor_si128(y1, sz1);
michael@0:     y0 = _mm_sub_epi16(y0, sz0);
michael@0:     y1 = _mm_sub_epi16(y1, sz1);
michael@0: 
michael@0:     _mm_store_si128((__m128i *)(y), y0);
michael@0:     _mm_store_si128((__m128i *)(y + 8), y1);
michael@0: 
michael@0:     zbin_boost_ptr = b->zrun_zbin_boost;
michael@0: 
michael@0:     /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
michael@0:     SELECT_EOB(1, 0);
michael@0:     SELECT_EOB(2, 1);
michael@0:     SELECT_EOB(3, 4);
michael@0:     SELECT_EOB(4, 8);
michael@0:     SELECT_EOB(5, 5);
michael@0:     SELECT_EOB(6, 2);
michael@0:     SELECT_EOB(7, 3);
michael@0:     SELECT_EOB(8, 6);
michael@0:     SELECT_EOB(9, 9);
michael@0:     SELECT_EOB(10, 12);
michael@0:     SELECT_EOB(11, 13);
michael@0:     SELECT_EOB(12, 10);
michael@0:     SELECT_EOB(13, 7);
michael@0:     SELECT_EOB(14, 11);
michael@0:     SELECT_EOB(15, 14);
michael@0:     SELECT_EOB(16, 15);
michael@0: 
michael@0:     y0 = _mm_load_si128((__m128i *)(d->qcoeff));
michael@0:     y1 = _mm_load_si128((__m128i *)(d->qcoeff + 8));
michael@0: 
michael@0:     /* dqcoeff = qcoeff * dequant */
michael@0:     y0 = _mm_mullo_epi16(y0, dequant0);
michael@0:     y1 = _mm_mullo_epi16(y1, dequant1);
michael@0: 
michael@0:     _mm_store_si128((__m128i *)(d->dqcoeff), y0);
michael@0:     _mm_store_si128((__m128i *)(d->dqcoeff + 8), y1);
michael@0: 
michael@0:     *d->eob = eob;
michael@0: }
michael@0: 
michael@0: void vp8_fast_quantize_b_sse2(BLOCK *b, BLOCKD *d)
michael@0: {
michael@0:   __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
michael@0:   __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
michael@0:   __m128i round0 = _mm_load_si128((__m128i *)(b->round));
michael@0:   __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
michael@0:   __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
michael@0:   __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
michael@0:   __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
michael@0:   __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
michael@0:   __m128i inv_zig_zag0 = _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag));
michael@0:   __m128i inv_zig_zag1 = _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag + 8));
michael@0: 
michael@0:   __m128i sz0, sz1, x0, x1, y0, y1, xdq0, xdq1, zeros, ones;
michael@0: 
michael@0:   /* sign of z: z >> 15 */
michael@0:   sz0 = _mm_srai_epi16(z0, 15);
michael@0:   sz1 = _mm_srai_epi16(z1, 15);
michael@0: 
michael@0:   /* x = abs(z): (z ^ sz) - sz */
michael@0:   x0 = _mm_xor_si128(z0, sz0);
michael@0:   x1 = _mm_xor_si128(z1, sz1);
michael@0:   x0 = _mm_sub_epi16(x0, sz0);
michael@0:   x1 = _mm_sub_epi16(x1, sz1);
michael@0: 
michael@0:   /* x += round */
michael@0:   x0 = _mm_add_epi16(x0, round0);
michael@0:   x1 = _mm_add_epi16(x1, round1);
michael@0: 
michael@0:   /* y = (x * quant) >> 16 */
michael@0:   y0 = _mm_mulhi_epi16(x0, quant_fast0);
michael@0:   y1 = _mm_mulhi_epi16(x1, quant_fast1);
michael@0: 
michael@0:   /* x = abs(y) = (y ^ sz) - sz */
michael@0:   y0 = _mm_xor_si128(y0, sz0);
michael@0:   y1 = _mm_xor_si128(y1, sz1);
michael@0:   x0 = _mm_sub_epi16(y0, sz0);
michael@0:   x1 = _mm_sub_epi16(y1, sz1);
michael@0: 
michael@0:   /* qcoeff = x */
michael@0:   _mm_store_si128((__m128i *)(d->qcoeff), x0);
michael@0:   _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
michael@0: 
michael@0:   /* x * dequant */
michael@0:   xdq0 = _mm_mullo_epi16(x0, dequant0);
michael@0:   xdq1 = _mm_mullo_epi16(x1, dequant1);
michael@0: 
michael@0:   /* dqcoeff = x * dequant */
michael@0:   _mm_store_si128((__m128i *)(d->dqcoeff), xdq0);
michael@0:   _mm_store_si128((__m128i *)(d->dqcoeff + 8), xdq1);
michael@0: 
michael@0:   /* build a mask for the zig zag */
michael@0:   zeros = _mm_setzero_si128();
michael@0: 
michael@0:   x0 = _mm_cmpeq_epi16(x0, zeros);
michael@0:   x1 = _mm_cmpeq_epi16(x1, zeros);
michael@0: 
michael@0:   ones = _mm_cmpeq_epi16(zeros, zeros);
michael@0: 
michael@0:   x0 = _mm_xor_si128(x0, ones);
michael@0:   x1 = _mm_xor_si128(x1, ones);
michael@0: 
michael@0:   x0 = _mm_and_si128(x0, inv_zig_zag0);
michael@0:   x1 = _mm_and_si128(x1, inv_zig_zag1);
michael@0: 
michael@0:   x0 = _mm_max_epi16(x0, x1);
michael@0: 
michael@0:   /* now down to 8 */
michael@0:   x1 = _mm_shuffle_epi32(x0, 0xE); // 0b00001110
michael@0: 
michael@0:   x0 = _mm_max_epi16(x0, x1);
michael@0: 
michael@0:   /* only 4 left */
michael@0:   x1 = _mm_shufflelo_epi16(x0, 0xE); // 0b00001110
michael@0: 
michael@0:   x0 = _mm_max_epi16(x0, x1);
michael@0: 
michael@0:   /* okay, just 2! */
michael@0:   x1 = _mm_shufflelo_epi16(x0, 0x1); // 0b00000001
michael@0: 
michael@0:   x0 = _mm_max_epi16(x0, x1);
michael@0: 
michael@0:   *d->eob = 0xFF & _mm_cvtsi128_si32(x0);
michael@0: }