Add avx2 highbd_quantize_b

- First pass encoding time reduces ~10.9% on i7-6700
  at 100 frames, 1080p.
- avx2 works for coeff number >= 8 cases; coeff number < 8
  case will be implemented by sse2.
- Unit test is added type B/FP/DC.

Change-Id: Ibe5b7807c64e6dfc2d59c470ed50a6e8ca94ef7c
diff --git a/aom_dsp/aom_dsp.mk b/aom_dsp/aom_dsp.mk
index 6e2d563..2274369 100644
--- a/aom_dsp/aom_dsp.mk
+++ b/aom_dsp/aom_dsp.mk
@@ -292,6 +292,7 @@
 DSP_SRCS-$(HAVE_SSE2)   += x86/quantize_sse2.c
 ifeq ($(CONFIG_HIGHBITDEPTH),yes)
 DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/highbd_quantize_intrin_avx2.c
 endif
 ifeq ($(ARCH_X86_64),yes)
 DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3_x86_64.asm
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 36c960d..6b0a46e 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -545,7 +545,7 @@
 
     if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
       add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
-      specialize qw/aom_highbd_quantize_b sse2/;
+      specialize qw/aom_highbd_quantize_b sse2 avx2/;
 
       add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
       specialize qw/aom_highbd_quantize_b_32x32 sse2/;
diff --git a/aom_dsp/x86/highbd_quantize_intrin_avx2.c b/aom_dsp/x86/highbd_quantize_intrin_avx2.c
new file mode 100644
index 0000000..2bbf15e
--- /dev/null
+++ b/aom_dsp/x86/highbd_quantize_intrin_avx2.c
@@ -0,0 +1,171 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+static INLINE void init_one_qp(const __m128i *p, __m256i *qp) {
+  const __m128i sign = _mm_srai_epi16(*p, 15);
+  const __m128i dc = _mm_unpacklo_epi16(*p, sign);
+  const __m128i ac = _mm_unpackhi_epi16(*p, sign);
+  *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1);
+}
+
+static INLINE void update_qp(__m256i *qp) {
+  int i;
+  for (i = 0; i < 5; ++i) {
+    qp[i] = _mm256_permute2x128_si256(qp[i], qp[i], 0x11);
+  }
+}
+
+static INLINE void init_qp(const int16_t *zbin_ptr, const int16_t *round_ptr,
+                           const int16_t *quant_ptr, const int16_t *dequant_ptr,
+                           const int16_t *quant_shift_ptr, __m256i *qp) {
+  const __m128i zbin = _mm_loadu_si128((const __m128i *)zbin_ptr);
+  const __m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
+  const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
+  const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr);
+  const __m128i quant_shift = _mm_loadu_si128((const __m128i *)quant_shift_ptr);
+  init_one_qp(&zbin, &qp[0]);
+  init_one_qp(&round, &qp[1]);
+  init_one_qp(&quant, &qp[2]);
+  init_one_qp(&dequant, &qp[3]);
+  init_one_qp(&quant_shift, &qp[4]);
+}
+
+// Note:
+// *x is vector multiplied by *y which is 16 int32_t parallel multiplication
+// and right shift 16.  The output, 16 int32_t is save in *p.
+static INLINE void mm256_mul_shift_epi32(const __m256i *x, const __m256i *y,
+                                         __m256i *p) {
+  __m256i prod_lo = _mm256_mul_epi32(*x, *y);
+  __m256i prod_hi = _mm256_srli_epi64(*x, 32);
+  const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
+  prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
+
+  prod_lo = _mm256_srli_epi64(prod_lo, 16);
+  const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1);
+  prod_lo = _mm256_and_si256(prod_lo, mask);
+  prod_hi = _mm256_srli_epi64(prod_hi, 16);
+
+  prod_hi = _mm256_slli_epi64(prod_hi, 32);
+  *p = _mm256_or_si256(prod_lo, prod_hi);
+}
+
+static INLINE void quantize(const __m256i *qp, __m256i *c,
+                            const int16_t *iscan_ptr, tran_low_t *qcoeff,
+                            tran_low_t *dqcoeff, __m256i *eob) {
+  const __m256i abs = _mm256_abs_epi32(*c);
+  const __m256i flag1 = _mm256_cmpgt_epi32(abs, qp[0]);
+  __m256i flag2 = _mm256_cmpeq_epi32(abs, qp[0]);
+  flag2 = _mm256_or_si256(flag1, flag2);
+  const int32_t nzflag = _mm256_movemask_epi8(flag2);
+
+  if (LIKELY(nzflag)) {
+    __m256i q = _mm256_add_epi32(abs, qp[1]);
+    __m256i tmp;
+    mm256_mul_shift_epi32(&q, &qp[2], &tmp);
+    q = _mm256_add_epi32(tmp, q);
+
+    mm256_mul_shift_epi32(&q, &qp[4], &q);
+    __m256i dq = _mm256_mullo_epi32(q, qp[3]);
+
+    q = _mm256_sign_epi32(q, *c);
+    dq = _mm256_sign_epi32(dq, *c);
+    q = _mm256_and_si256(q, flag2);
+    dq = _mm256_and_si256(dq, flag2);
+
+    _mm256_storeu_si256((__m256i *)qcoeff, q);
+    _mm256_storeu_si256((__m256i *)dqcoeff, dq);
+
+    const __m128i isc = _mm_loadu_si128((const __m128i *)iscan_ptr);
+    const __m128i zr = _mm_setzero_si128();
+    const __m128i lo = _mm_unpacklo_epi16(isc, zr);
+    const __m128i hi = _mm_unpackhi_epi16(isc, zr);
+    const __m256i iscan =
+        _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
+
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i zc = _mm256_cmpeq_epi32(dq, zero);
+    const __m256i nz = _mm256_cmpeq_epi32(zc, zero);
+    __m256i cur_eob = _mm256_sub_epi32(iscan, nz);
+    cur_eob = _mm256_and_si256(cur_eob, nz);
+    *eob = _mm256_max_epi32(cur_eob, *eob);
+  } else {
+    const __m256i zero = _mm256_setzero_si256();
+    _mm256_storeu_si256((__m256i *)qcoeff, zero);
+    _mm256_storeu_si256((__m256i *)dqcoeff, zero);
+  }
+}
+
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                int skip_block, const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  const unsigned int step = 8;
+
+  if (LIKELY(!skip_block)) {
+    __m256i qp[5], coeff;
+    init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp);
+    coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
+
+    __m256i eob = _mm256_setzero_si256();
+    quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
+
+    coeff_ptr += step;
+    qcoeff_ptr += step;
+    dqcoeff_ptr += step;
+    iscan += step;
+    n_coeffs -= step;
+
+    update_qp(qp);
+
+    while (n_coeffs > 0) {
+      coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
+      quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
+
+      coeff_ptr += step;
+      qcoeff_ptr += step;
+      dqcoeff_ptr += step;
+      iscan += step;
+      n_coeffs -= step;
+    }
+    {
+      __m256i eob_s;
+      eob_s = _mm256_shuffle_epi32(eob, 0xe);
+      eob = _mm256_max_epi16(eob, eob_s);
+      eob_s = _mm256_shufflelo_epi16(eob, 0xe);
+      eob = _mm256_max_epi16(eob, eob_s);
+      eob_s = _mm256_shufflelo_epi16(eob, 1);
+      eob = _mm256_max_epi16(eob, eob_s);
+      const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob),
+                                              _mm256_extractf128_si256(eob, 1));
+      *eob_ptr = _mm_extract_epi16(final_eob, 0);
+    }
+  } else {
+    do {
+      const __m256i zero = _mm256_setzero_si256();
+      _mm256_storeu_si256((__m256i *)qcoeff_ptr, zero);
+      _mm256_storeu_si256((__m256i *)dqcoeff_ptr, zero);
+      qcoeff_ptr += step;
+      dqcoeff_ptr += step;
+      n_coeffs -= step;
+    } while (n_coeffs > 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/av1/encoder/av1_quantize.c b/av1/encoder/av1_quantize.c
index 84d4bbf..0d6ea01 100644
--- a/av1/encoder/av1_quantize.c
+++ b/av1/encoder/av1_quantize.c
@@ -899,14 +899,29 @@
 
   switch (qparam->log_scale) {
     case 0:
-      aom_highbd_quantize_b(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round,
-                            p->quant, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
-                            pd->dequant, eob_ptr, sc->scan, sc->iscan
+      if (LIKELY(n_coeffs >= 8)) {
+        aom_highbd_quantize_b(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                              p->round, p->quant, p->quant_shift, qcoeff_ptr,
+                              dqcoeff_ptr, pd->dequant, eob_ptr, sc->scan,
+                              sc->iscan
 #if CONFIG_AOM_QM
-                            ,
-                            qm_ptr, iqm_ptr
+                              ,
+                              qm_ptr, iqm_ptr
 #endif
-                            );
+                              );
+      } else {
+        // TODO(luoyi): Need SIMD (e.g. sse2) for smaller block size
+        // quantization
+        aom_highbd_quantize_b_c(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                                p->round, p->quant, p->quant_shift, qcoeff_ptr,
+                                dqcoeff_ptr, pd->dequant, eob_ptr, sc->scan,
+                                sc->iscan
+#if CONFIG_AOM_QM
+                                ,
+                                qm_ptr, iqm_ptr
+#endif
+                                );
+      }
       break;
     case 1:
       aom_highbd_quantize_b_32x32(coeff_ptr, n_coeffs, skip_block, p->zbin,
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc
index c522f9e..2d17163 100644
--- a/test/quantize_func_test.cc
+++ b/test/quantize_func_test.cc
@@ -12,6 +12,7 @@
 #include "third_party/googletest/src/googletest/include/gtest/gtest.h"
 
 #include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
 #include "./av1_rtcd.h"
 #include "aom/aom_codec.h"
 #include "aom_ports/aom_timer.h"
@@ -59,7 +60,10 @@
   HBD_QUAN_FUNC;
 }
 
-typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, TX_SIZE, aom_bit_depth_t>
+typedef enum { TYPE_B, TYPE_DC, TYPE_FP } QuantType;
+
+typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, TX_SIZE, QuantType,
+                        aom_bit_depth_t>
     QuantizeParam;
 
 typedef struct {
@@ -73,7 +77,7 @@
  protected:
   QuantizeTest()
       : quant_ref_(GET_PARAM(0)), quant_(GET_PARAM(1)), tx_size_(GET_PARAM(2)),
-        bd_(GET_PARAM(3)) {}
+        type_(GET_PARAM(3)), bd_(GET_PARAM(4)) {}
 
   virtual ~QuantizeTest() {}
 
@@ -114,8 +118,17 @@
 
     // Testing uses luminance quantization table
     const int16_t *zbin = qtab_->quant.y_zbin[q];
-    const int16_t *round_fp = qtab_->quant.y_round_fp[q];
-    const int16_t *quant_fp = qtab_->quant.y_quant_fp[q];
+
+    const int16_t *round = 0;
+    const int16_t *quant = 0;
+    if (type_ == TYPE_B) {
+      round = qtab_->quant.y_round[q];
+      quant = qtab_->quant.y_quant[q];
+    } else if (type_ == TYPE_FP) {
+      round = qtab_->quant.y_round_fp[q];
+      quant = qtab_->quant.y_quant_fp[q];
+    }
+
     const int16_t *quant_shift = qtab_->quant.y_quant_shift[q];
     const int16_t *dequant = qtab_->dequant.y_dequant[q];
 
@@ -124,13 +137,13 @@
 
       memset(qcoeff_ref, 0, 5 * n_coeffs * sizeof(*qcoeff_ref));
 
-      quant_ref_(coeff_ptr, n_coeffs, skip_block, zbin, round_fp, quant_fp,
+      quant_ref_(coeff_ptr, n_coeffs, skip_block, zbin, round, quant,
                  quant_shift, qcoeff_ref, dqcoeff_ref, dequant, &eob[0],
                  sc->scan, sc->iscan);
 
       ASM_REGISTER_STATE_CHECK(quant_(
-          coeff_ptr, n_coeffs, skip_block, zbin, round_fp, quant_fp,
-          quant_shift, qcoeff, dqcoeff, dequant, &eob[1], sc->scan, sc->iscan));
+          coeff_ptr, n_coeffs, skip_block, zbin, round, quant, quant_shift,
+          qcoeff, dqcoeff, dequant, &eob[1], sc->scan, sc->iscan));
 
       for (int j = 0; j < n_coeffs; ++j) {
         ASSERT_EQ(qcoeff_ref[j], qcoeff[j])
@@ -215,6 +228,7 @@
   QuantizeFunc quant_ref_;
   QuantizeFunc quant_;
   TX_SIZE tx_size_;
+  QuantType type_;
   aom_bit_depth_t bd_;
 };
 
@@ -283,39 +297,46 @@
 
 #if HAVE_AVX2
 const QuantizeParam kQParamArrayAvx2[] = {
-  make_tuple(&av1_quantize_fp_c, &av1_quantize_fp_avx2, TX_16X16, AOM_BITS_8),
-  make_tuple(&av1_quantize_fp_32x32_c, &av1_quantize_fp_32x32_avx2, TX_32X32,
+  make_tuple(&av1_quantize_fp_c, &av1_quantize_fp_avx2, TX_16X16, TYPE_FP,
              AOM_BITS_8),
+  make_tuple(&av1_quantize_fp_32x32_c, &av1_quantize_fp_32x32_avx2, TX_32X32,
+             TYPE_FP, AOM_BITS_8),
 #if CONFIG_HIGHBITDEPTH
   make_tuple(&highbd_quan16x16_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan16x16_wrapper<av1_highbd_quantize_fp_avx2>, TX_16X16,
-             AOM_BITS_8),
+             TYPE_FP, AOM_BITS_8),
   make_tuple(&highbd_quan16x16_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan16x16_wrapper<av1_highbd_quantize_fp_avx2>, TX_16X16,
-             AOM_BITS_10),
+             TYPE_FP, AOM_BITS_10),
   make_tuple(&highbd_quan16x16_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan16x16_wrapper<av1_highbd_quantize_fp_avx2>, TX_16X16,
-             AOM_BITS_12),
+             TYPE_FP, AOM_BITS_12),
   make_tuple(&highbd_quan32x32_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan32x32_wrapper<av1_highbd_quantize_fp_avx2>, TX_32X32,
-             AOM_BITS_8),
+             TYPE_FP, AOM_BITS_8),
   make_tuple(&highbd_quan32x32_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan32x32_wrapper<av1_highbd_quantize_fp_avx2>, TX_32X32,
-             AOM_BITS_10),
+             TYPE_FP, AOM_BITS_10),
   make_tuple(&highbd_quan32x32_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan32x32_wrapper<av1_highbd_quantize_fp_avx2>, TX_32X32,
-             AOM_BITS_12),
+             TYPE_FP, AOM_BITS_12),
 #if CONFIG_TX64X64
   make_tuple(&highbd_quan64x64_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan64x64_wrapper<av1_highbd_quantize_fp_avx2>, TX_64X64,
-             AOM_BITS_8),
+             TYPE_FP, AOM_BITS_8),
   make_tuple(&highbd_quan64x64_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan64x64_wrapper<av1_highbd_quantize_fp_avx2>, TX_64X64,
-             AOM_BITS_10),
+             TYPE_FP, AOM_BITS_10),
   make_tuple(&highbd_quan64x64_wrapper<av1_highbd_quantize_fp_c>,
              &highbd_quan64x64_wrapper<av1_highbd_quantize_fp_avx2>, TX_64X64,
-             AOM_BITS_12),
+             TYPE_FP, AOM_BITS_12),
 #endif  // CONFIG_TX64X64
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_avx2, TX_16X16,
+             TYPE_B, AOM_BITS_8),
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_avx2, TX_16X16,
+             TYPE_B, AOM_BITS_10),
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_avx2, TX_16X16,
+             TYPE_B, AOM_BITS_12),
 #endif  // CONFIG_HIGHBITDEPTH
 };
 
@@ -324,8 +345,24 @@
 #endif  // HAVE_AVX2
 
 #if HAVE_SSE2
-const QuantizeParam kQParamArraySSE2[] = { make_tuple(
-    &av1_quantize_fp_c, &av1_quantize_fp_sse2, TX_16X16, AOM_BITS_8) };
+const QuantizeParam kQParamArraySSE2[] = {
+  make_tuple(&av1_quantize_fp_c, &av1_quantize_fp_sse2, TX_16X16, TYPE_FP,
+             AOM_BITS_8),
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_sse2, TX_16X16,
+             TYPE_B, AOM_BITS_8),
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_sse2, TX_16X16,
+             TYPE_B, AOM_BITS_10),
+  make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_sse2, TX_16X16,
+             TYPE_B, AOM_BITS_12),
+  make_tuple(&aom_highbd_quantize_b_32x32_c, &aom_highbd_quantize_b_32x32_sse2,
+             TX_32X32, TYPE_B, AOM_BITS_8),
+  make_tuple(&aom_highbd_quantize_b_32x32_c, &aom_highbd_quantize_b_32x32_sse2,
+             TX_32X32, TYPE_B, AOM_BITS_10),
+  make_tuple(&aom_highbd_quantize_b_32x32_c, &aom_highbd_quantize_b_32x32_sse2,
+             TX_32X32, TYPE_B, AOM_BITS_12),
+#endif
+};
 
 INSTANTIATE_TEST_CASE_P(SSE2, QuantizeTest,
                         ::testing::ValuesIn(kQParamArraySSE2));
@@ -333,7 +370,8 @@
 
 #if !CONFIG_HIGHBITDEPTH && HAVE_SSSE3 && ARCH_X86_64
 const QuantizeParam kQ16x16ParamArraySSSE3[] = {
-  make_tuple(&av1_quantize_fp_c, &av1_quantize_fp_ssse3, TX_16X16, AOM_BITS_8),
+  make_tuple(&av1_quantize_fp_c, &av1_quantize_fp_ssse3, TX_16X16, TYPE_FP,
+             AOM_BITS_8),
 };
 INSTANTIATE_TEST_CASE_P(SSSE3, QuantizeTest,
                         ::testing::ValuesIn(kQ16x16ParamArraySSSE3));
@@ -341,7 +379,7 @@
 // TODO(any):
 //  The following test does not pass yet
 const QuantizeParam kQ32x32ParamArraySSSE3[] = { make_tuple(
-    av1_quantize_fp_32x32_c, av1_quantize_fp_32x32_ssse3, TX_32X32,
+    av1_quantize_fp_32x32_c, av1_quantize_fp_32x32_ssse3, TX_32X32, TYPE_FP,
     AOM_BITS_8) };
 INSTANTIATE_TEST_CASE_P(DISABLED_SSSE3, QuantizeTest,
                         ::testing::ValuesIn(kQ32x32ParamArraySSSE3));