Quelle  vp9_quantize_test.cc   Sprache: C

/*
 *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

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

#include "gtest/gtest.h"

#include "./vp9_rtcd.h"
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
#include "test/acm_random.h"
#include "test/bench.h"
#include "test/buffer.h"
#include "test/clear_system_state.h"
#include "test/register_state_check.h"
#include "test/util.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_scan.h"
#include "vp9/encoder/vp9_block.h"
#include "vpx/vpx_codec.h"
#include "vpx/vpx_integer.h"
#include "vpx_ports/vpx_timer.h"

using libvpx_test::ACMRandom;
using libvpx_test::Buffer;

namespace {
const int number_of_iterations = 100;

typedef void (*QuantizeFunc)(const tran_low_t *coeff, intptr_t count,
                             const macroblock_plane *mb_plane,
                             tran_low_t *qcoeff, tran_low_t *dqcoeff,
                             const int16_t *dequant, uint16_t *eob,
                             const struct ScanOrder *const scan_order);
typedef std::tuple<QuantizeFunc, QuantizeFunc, vpx_bit_depth_t,
                   int /*max_size*/, bool /*is_fp*/>
    QuantizeParam;

// Wrapper for 32x32 version which does not use count
typedef void (*Quantize32x32Func)(const tran_low_t *coeff,
                                  const macroblock_plane *const mb_plane,
                                  tran_low_t *qcoeff, tran_low_t *dqcoeff,
                                  const int16_t *dequant, uint16_t *eob,
                                  const struct ScanOrder *const scan_order);

template <Quantize32x32Func fn>
void Quant32x32Wrapper(const tran_low_t *coeff, intptr_t count,
                       const macroblock_plane *const mb_plane,
                       tran_low_t *qcoeff, tran_low_t *dqcoeff,
                       const int16_t *dequant, uint16_t *eob,
                       const struct ScanOrder *const scan_order) {
  (void)count;
  fn(coeff, mb_plane, qcoeff, dqcoeff, dequant, eob, scan_order);
}

// Wrapper for FP version which does not use zbin or quant_shift.
typedef void (*QuantizeFPFunc)(const tran_low_t *coeff, intptr_t count,
                               const macroblock_plane *const mb_plane,
                               tran_low_t *qcoeff, tran_low_t *dqcoeff,
                               const int16_t *dequant, uint16_t *eob,
                               const struct ScanOrder *const scan_order);

template <QuantizeFPFunc fn>
void QuantFPWrapper(const tran_low_t *coeff, intptr_t count,
                    const macroblock_plane *const mb_plane, tran_low_t *qcoeff,
                    tran_low_t *dqcoeff, const int16_t *dequant, uint16_t *eob,
                    const struct ScanOrder *const scan_order) {
  fn(coeff, count, mb_plane, qcoeff, dqcoeff, dequant, eob, scan_order);
}

void GenerateHelperArrays(ACMRandom *rnd, int16_t *zbin, int16_t *round,
                          int16_t *quant, int16_t *quant_shift,
                          int16_t *dequant, int16_t *round_fp,
                          int16_t *quant_fp) {
  // Max when q == 0. Otherwise, it is 48 for Y and 42 for U/V.
  constexpr int kMaxQRoundingFactorFp = 64;

  for (int j = 0; j < 2; j++) {
    // The range is 4 to 1828 in the VP9 tables.
    const int qlookup = rnd->RandRange(1825) + 4;
    round_fp[j] = (kMaxQRoundingFactorFp * qlookup) >> 7;
    quant_fp[j] = (1 << 16) / qlookup;

    // Values determined by deconstructing vp9_init_quantizer().
    // zbin may be up to 1143 for 8 and 10 bit Y values, or 1200 for 12 bit Y
    // values or U/V values of any bit depth. This is because y_delta is not
    // factored into the vp9_ac_quant() call.
    zbin[j] = rnd->RandRange(1200);

    // round may be up to 685 for Y values or 914 for U/V.
    round[j] = rnd->RandRange(914);
    // quant ranges from 1 to -32703
    quant[j] = static_cast<int>(rnd->RandRange(32704)) - 32703;
    // quant_shift goes up to 1 << 16.
    quant_shift[j] = rnd->RandRange(16384);
    // dequant maxes out at 1828 for all cases.
    dequant[j] = rnd->RandRange(1828);
  }
  for (int j = 2; j < 8; j++) {
    zbin[j] = zbin[1];
    round_fp[j] = round_fp[1];
    quant_fp[j] = quant_fp[1];
    round[j] = round[1];
    quant[j] = quant[1];
    quant_shift[j] = quant_shift[1];
    dequant[j] = dequant[1];
  }
}

class VP9QuantizeBase : public AbstractBench {
 public:
  VP9QuantizeBase(vpx_bit_depth_t bit_depth, int max_size, bool is_fp)
      : bit_depth_(bit_depth), max_size_(max_size), is_fp_(is_fp),
        coeff_(Buffer<tran_low_t>(max_size_, max_size_, 0, 16)),
        qcoeff_(Buffer<tran_low_t>(max_size_, max_size_, 0, 32)),
        dqcoeff_(Buffer<tran_low_t>(max_size_, max_size_, 0, 32)) {
    // TODO(jianj): SSSE3 and AVX2 tests fail on extreme values.
#if HAVE_NEON
    max_value_ = (1 << (7 + bit_depth_)) - 1;
#else
    max_value_ = (1 << bit_depth_) - 1;
#endif

    mb_plane_ = reinterpret_cast<macroblock_plane *>(
        vpx_memalign(16, sizeof(macroblock_plane)));

    zbin_ptr_ = mb_plane_->zbin =
        reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*zbin_ptr_)));
    round_fp_ptr_ = mb_plane_->round_fp = reinterpret_cast<int16_t *>(
        vpx_memalign(16, 8 * sizeof(*round_fp_ptr_)));
    quant_fp_ptr_ = mb_plane_->quant_fp = reinterpret_cast<int16_t *>(
        vpx_memalign(16, 8 * sizeof(*quant_fp_ptr_)));
    round_ptr_ = mb_plane_->round =
        reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*round_ptr_)));
    quant_ptr_ = mb_plane_->quant =
        reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*quant_ptr_)));
    quant_shift_ptr_ = mb_plane_->quant_shift = reinterpret_cast<int16_t *>(
        vpx_memalign(16, 8 * sizeof(*quant_shift_ptr_)));
    dequant_ptr_ = reinterpret_cast<int16_t *>(
        vpx_memalign(16, 8 * sizeof(*dequant_ptr_)));

    r_ptr_ = (is_fp_) ? round_fp_ptr_ : round_ptr_;
    q_ptr_ = (is_fp_) ? quant_fp_ptr_ : quant_ptr_;
  }

  ~VP9QuantizeBase() override {
    vpx_free(mb_plane_);
    vpx_free(zbin_ptr_);
    vpx_free(round_fp_ptr_);
    vpx_free(quant_fp_ptr_);
    vpx_free(round_ptr_);
    vpx_free(quant_ptr_);
    vpx_free(quant_shift_ptr_);
    vpx_free(dequant_ptr_);
    mb_plane_ = nullptr;
    zbin_ptr_ = nullptr;
    round_fp_ptr_ = nullptr;
    quant_fp_ptr_ = nullptr;
    round_ptr_ = nullptr;
    quant_ptr_ = nullptr;
    quant_shift_ptr_ = nullptr;
    dequant_ptr_ = nullptr;
    libvpx_test::ClearSystemState();
  }

 protected:
  macroblock_plane *mb_plane_;
  int16_t *zbin_ptr_;
  int16_t *quant_fp_ptr_;
  int16_t *round_fp_ptr_;
  int16_t *round_ptr_;
  int16_t *quant_ptr_;
  int16_t *quant_shift_ptr_;
  int16_t *dequant_ptr_;
  const vpx_bit_depth_t bit_depth_;
  int max_value_;
  const int max_size_;
  const bool is_fp_;
  Buffer<tran_low_t> coeff_;
  Buffer<tran_low_t> qcoeff_;
  Buffer<tran_low_t> dqcoeff_;
  int16_t *r_ptr_;
  int16_t *q_ptr_;
  int count_;
  const ScanOrder *scan_;
  uint16_t eob_;
};

class VP9QuantizeTest : public VP9QuantizeBase,
                        public ::testing::TestWithParam<QuantizeParam> {
 public:
  VP9QuantizeTest()
      : VP9QuantizeBase(GET_PARAM(2), GET_PARAM(3), GET_PARAM(4)),
        quantize_op_(GET_PARAM(0)), ref_quantize_op_(GET_PARAM(1)) {}

 protected:
  void Run() override;
  void Speed(bool is_median);
  const QuantizeFunc quantize_op_;
  const QuantizeFunc ref_quantize_op_;
};

void VP9QuantizeTest::Run() {
  quantize_op_(coeff_.TopLeftPixel(), count_, mb_plane_, qcoeff_.TopLeftPixel(),
               dqcoeff_.TopLeftPixel(), dequant_ptr_, &eob_, scan_);
}

void VP9QuantizeTest::Speed(bool is_median) {
  ACMRandom rnd(ACMRandom::DeterministicSeed());
  ASSERT_TRUE(coeff_.Init());
  ASSERT_TRUE(qcoeff_.Init());
  ASSERT_TRUE(dqcoeff_.Init());
  TX_SIZE starting_sz, ending_sz;

  if (max_size_ == 16) {
    starting_sz = TX_4X4;
    ending_sz = TX_16X16;
  } else {
    starting_sz = TX_32X32;
    ending_sz = TX_32X32;
  }

  for (TX_SIZE sz = starting_sz; sz <= ending_sz; ++sz) {
    // zbin > coeff, zbin < coeff.
    for (int i = 0; i < 2; ++i) {
      // TX_TYPE defines the scan order. That is not relevant to the speed test.
      // Pick the first one.
      const TX_TYPE tx_type = DCT_DCT;
      count_ = (4 << sz) * (4 << sz);
      scan_ = &vp9_scan_orders[sz][tx_type];

      GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_,
                           quant_shift_ptr_, dequant_ptr_, round_fp_ptr_,
                           quant_fp_ptr_);

      if (i == 0) {
        // When |coeff values| are less than zbin the results are 0.
        int threshold = 100;
        if (max_size_ == 32) {
          // For 32x32, the threshold is halved. Double it to keep the values
          // from clearing it.
          threshold = 200;
        }
        for (int j = 0; j < 8; ++j) zbin_ptr_[j] = threshold;
        coeff_.Set(&rnd, -99, 99);
      } else if (i == 1) {
        for (int j = 0; j < 8; ++j) zbin_ptr_[j] = 50;
        coeff_.Set(&rnd, -500, 500);
      }

      const char *type =
          (i == 0) ? "Bypass calculations " : "Full calculations ";
      char block_size[16];
      snprintf(block_size, sizeof(block_size), "%dx%d", 4 << sz, 4 << sz);
      char title[100];
      snprintf(title, sizeof(title), "%25s %8s ", type, block_size);

      if (is_median) {
        RunNTimes(10000000 / count_);
        PrintMedian(title);
      } else {
        Buffer<tran_low_t> ref_qcoeff =
            Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
        ASSERT_TRUE(ref_qcoeff.Init());
        Buffer<tran_low_t> ref_dqcoeff =
            Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
        ASSERT_TRUE(ref_dqcoeff.Init());
        uint16_t ref_eob = 0;

        const int kNumTests = 5000000;
        vpx_usec_timer timer, simd_timer;

        vpx_usec_timer_start(&timer);
        for (int n = 0; n < kNumTests; ++n) {
          ref_quantize_op_(coeff_.TopLeftPixel(), count_, mb_plane_,
                           ref_qcoeff.TopLeftPixel(),
                           ref_dqcoeff.TopLeftPixel(), dequant_ptr_, &ref_eob,
                           scan_);
        }
        vpx_usec_timer_mark(&timer);

        vpx_usec_timer_start(&simd_timer);
        for (int n = 0; n < kNumTests; ++n) {
          quantize_op_(coeff_.TopLeftPixel(), count_, mb_plane_,
                       qcoeff_.TopLeftPixel(), dqcoeff_.TopLeftPixel(),
                       dequant_ptr_, &eob_, scan_);
        }
        vpx_usec_timer_mark(&simd_timer);

        const int elapsed_time =
            static_cast<int>(vpx_usec_timer_elapsed(&timer));
        const int simd_elapsed_time =
            static_cast<int>(vpx_usec_timer_elapsed(&simd_timer));
        printf("%s c_time = %d \t simd_time = %d \t Gain = %f \n", title,
               elapsed_time, simd_elapsed_time,
               ((float)elapsed_time / simd_elapsed_time));
      }
    }
  }
}

// This quantizer compares the AC coefficients to the quantization step size to
// determine if further multiplication operations are needed.
// Based on vp9_quantize_fp_sse2().
inline void quant_fp_nz(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
                        const struct macroblock_plane *const mb_plane,
                        tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
                        const int16_t *dequant_ptr, uint16_t *eob_ptr,
                        const struct ScanOrder *const scan_order,
                        int is_32x32) {
  int i, eob = -1;
  const int thr = dequant_ptr[1] >> (1 + is_32x32);
  const int16_t *round_ptr = mb_plane->round_fp;
  const int16_t *quant_ptr = mb_plane->quant_fp;
  const int16_t *scan = scan_order->scan;

  // Quantization pass: All coefficients with index >= zero_flag are
  // skippable. Note: zero_flag can be zero.
  for (i = 0; i < n_coeffs; i += 16) {
    int y;
    int nzflag_cnt = 0;
    int abs_coeff[16];
    int coeff_sign[16];

    // count nzflag for each row (16 tran_low_t)
    for (y = 0; y < 16; ++y) {
      const int rc = i + y;
      const int coeff = coeff_ptr[rc];
      coeff_sign[y] = (coeff >> 31);
      abs_coeff[y] = (coeff ^ coeff_sign[y]) - coeff_sign[y];
      // The first 16 are skipped in the sse2 code.  Do the same here to match.
      if (i >= 16 && (abs_coeff[y] <= thr)) {
        nzflag_cnt++;
      }
    }

    for (y = 0; y < 16; ++y) {
      const int rc = i + y;
      // If all of the AC coeffs in a row has magnitude less than the
      // quantization step_size/2, quantize to zero.
      if (nzflag_cnt < 16) {
        int tmp;
        int _round;

        if (is_32x32) {
          _round = ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
        } else {
          _round = round_ptr[rc != 0];
        }
        tmp = clamp(abs_coeff[y] + _round, INT16_MIN, INT16_MAX);
        tmp = (tmp * quant_ptr[rc != 0]) >> (16 - is_32x32);
        qcoeff_ptr[rc] = (tmp ^ coeff_sign[y]) - coeff_sign[y];
        dqcoeff_ptr[rc] =
            static_cast<tran_low_t>(qcoeff_ptr[rc] * dequant_ptr[rc != 0]);

        if (is_32x32) {
          dqcoeff_ptr[rc] = static_cast<tran_low_t>(qcoeff_ptr[rc] *
                                                    dequant_ptr[rc != 0] / 2);
        } else {
          dqcoeff_ptr[rc] =
              static_cast<tran_low_t>(qcoeff_ptr[rc] * dequant_ptr[rc != 0]);
        }
      } else {
        qcoeff_ptr[rc] = 0;
        dqcoeff_ptr[rc] = 0;
      }
    }
  }

  // Scan for eob.
  for (i = 0; i < n_coeffs; i++) {
    // Use the scan order to find the correct eob.
    const int rc = scan[i];
    if (qcoeff_ptr[rc]) {
      eob = i;
    }
  }
  *eob_ptr = eob + 1;
}

void quantize_fp_nz_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
                      const struct macroblock_plane *mb_plane,
                      tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
                      const int16_t *dequant_ptr, uint16_t *eob_ptr,
                      const struct ScanOrder *const scan_order) {
  quant_fp_nz(coeff_ptr, n_coeffs, mb_plane, qcoeff_ptr, dqcoeff_ptr,
              dequant_ptr, eob_ptr, scan_order, 0);
}

void quantize_fp_32x32_nz_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
                            const struct macroblock_plane *mb_plane,
                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
                            const int16_t *dequant_ptr, uint16_t *eob_ptr,
                            const struct ScanOrder *const scan_order) {
  quant_fp_nz(coeff_ptr, n_coeffs, mb_plane, qcoeff_ptr, dqcoeff_ptr,
              dequant_ptr, eob_ptr, scan_order, 1);
}

TEST_P(VP9QuantizeTest, OperationCheck) {
  ACMRandom rnd(ACMRandom::DeterministicSeed());
  ASSERT_TRUE(coeff_.Init());
  ASSERT_TRUE(qcoeff_.Init());
  ASSERT_TRUE(dqcoeff_.Init());
  Buffer<tran_low_t> ref_qcoeff =
      Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
  ASSERT_TRUE(ref_qcoeff.Init());
  Buffer<tran_low_t> ref_dqcoeff =
      Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
  ASSERT_TRUE(ref_dqcoeff.Init());
  uint16_t ref_eob = 0;
  eob_ = 0;

  for (int i = 0; i < number_of_iterations; ++i) {
    TX_SIZE sz;
    if (max_size_ == 16) {
      sz = static_cast<TX_SIZE>(i % 3);  // TX_4X4, TX_8X8 TX_16X16
    } else {
      sz = TX_32X32;
    }
    const TX_TYPE tx_type = static_cast<TX_TYPE>((i >> 2) % 3);
    scan_ = &vp9_scan_orders[sz][tx_type];
    count_ = (4 << sz) * (4 << sz);
    coeff_.Set(&rnd, -max_value_, max_value_);
    GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_,
                         quant_shift_ptr_, dequant_ptr_, round_fp_ptr_,
                         quant_fp_ptr_);
    ref_quantize_op_(coeff_.TopLeftPixel(), count_, mb_plane_,
                     ref_qcoeff.TopLeftPixel(), ref_dqcoeff.TopLeftPixel(),
                     dequant_ptr_, &ref_eob, scan_);

    ASM_REGISTER_STATE_CHECK(quantize_op_(
        coeff_.TopLeftPixel(), count_, mb_plane_, qcoeff_.TopLeftPixel(),
        dqcoeff_.TopLeftPixel(), dequant_ptr_, &eob_, scan_));

    EXPECT_TRUE(qcoeff_.CheckValues(ref_qcoeff));
    EXPECT_TRUE(dqcoeff_.CheckValues(ref_dqcoeff));

    EXPECT_EQ(eob_, ref_eob);

    if (HasFailure()) {
      printf("Failure on iteration %d.\n", i);
      qcoeff_.PrintDifference(ref_qcoeff);
      dqcoeff_.PrintDifference(ref_dqcoeff);
      return;
    }
  }
}

TEST_P(VP9QuantizeTest, EOBCheck) {
  ACMRandom rnd(ACMRandom::DeterministicSeed());
  ASSERT_TRUE(coeff_.Init());
  ASSERT_TRUE(qcoeff_.Init());
  ASSERT_TRUE(dqcoeff_.Init());
  Buffer<tran_low_t> ref_qcoeff =
      Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
  ASSERT_TRUE(ref_qcoeff.Init());
  Buffer<tran_low_t> ref_dqcoeff =
      Buffer<tran_low_t>(max_size_, max_size_, 0, 32);
  ASSERT_TRUE(ref_dqcoeff.Init());
  uint16_t ref_eob = 0;
  eob_ = 0;
  const uint32_t max_index = max_size_ * max_size_ - 1;

  for (int i = 0; i < number_of_iterations; ++i) {
    TX_SIZE sz;
    if (max_size_ == 16) {
      sz = static_cast<TX_SIZE>(i % 3);  // TX_4X4, TX_8X8 TX_16X16
    } else {
      sz = TX_32X32;
    }
    const TX_TYPE tx_type = static_cast<TX_TYPE>((i >> 2) % 3);
    scan_ = &vp9_scan_orders[sz][tx_type];
    count_ = (4 << sz) * (4 << sz);
    // Two random entries
    coeff_.Set(0);
    coeff_.TopLeftPixel()[rnd.RandRange(count_) & max_index] =
        static_cast<int>(rnd.RandRange(max_value_ * 2)) - max_value_;
    coeff_.TopLeftPixel()[rnd.RandRange(count_) & max_index] =
        static_cast<int>(rnd.RandRange(max_value_ * 2)) - max_value_;
    GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_,
                         quant_shift_ptr_, dequant_ptr_, round_fp_ptr_,
                         quant_fp_ptr_);
    ref_quantize_op_(coeff_.TopLeftPixel(), count_, mb_plane_,
                     ref_qcoeff.TopLeftPixel(), ref_dqcoeff.TopLeftPixel(),
                     dequant_ptr_, &ref_eob, scan_);

    ASM_REGISTER_STATE_CHECK(quantize_op_(
        coeff_.TopLeftPixel(), count_, mb_plane_, qcoeff_.TopLeftPixel(),
        dqcoeff_.TopLeftPixel(), dequant_ptr_, &eob_, scan_));

    EXPECT_TRUE(qcoeff_.CheckValues(ref_qcoeff));
    EXPECT_TRUE(dqcoeff_.CheckValues(ref_dqcoeff));

    EXPECT_EQ(eob_, ref_eob);

    if (HasFailure()) {
      printf("Failure on iteration %d.\n", i);
      qcoeff_.PrintDifference(ref_qcoeff);
      dqcoeff_.PrintDifference(ref_dqcoeff);
      return;
    }
  }
}

TEST_P(VP9QuantizeTest, DISABLED_Speed) { Speed(false); }

TEST_P(VP9QuantizeTest, DISABLED_SpeedMedian) { Speed(true); }

using std::make_tuple;

#if HAVE_SSE2
#if CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
    SSE2, VP9QuantizeTest,
    ::testing::Values(
        make_tuple(vpx_quantize_b_sse2, vpx_quantize_b_c, VPX_BITS_8, 16,
                   false),
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_sse2>,
                   &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 16, true),
        make_tuple(vpx_highbd_quantize_b_sse2, vpx_highbd_quantize_b_c,
                   VPX_BITS_8, 16, false),
        make_tuple(vpx_highbd_quantize_b_sse2, vpx_highbd_quantize_b_c,
                   VPX_BITS_10, 16, false),
        make_tuple(vpx_highbd_quantize_b_sse2, vpx_highbd_quantize_b_c,
                   VPX_BITS_12, 16, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_sse2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_8, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_sse2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_10, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_sse2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_12, 32, false)));

#else
INSTANTIATE_TEST_SUITE_P(
    SSE2, VP9QuantizeTest,
    ::testing::Values(make_tuple(vpx_quantize_b_sse2, vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_sse2>,
                                 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8,
                                 16, true)));
#endif  // CONFIG_VP9_HIGHBITDEPTH
#endif  // HAVE_SSE2

#if HAVE_SSSE3
INSTANTIATE_TEST_SUITE_P(
    SSSE3, VP9QuantizeTest,
    ::testing::Values(make_tuple(vpx_quantize_b_ssse3, vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_ssse3>,
                                 &Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                                 VPX_BITS_8, 32, false),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_ssse3>,
                                 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8,
                                 16, true),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_ssse3>,
                                 &QuantFPWrapper<quantize_fp_32x32_nz_c>,
                                 VPX_BITS_8, 32, true)));
#endif  // HAVE_SSSE3

#if HAVE_AVX
INSTANTIATE_TEST_SUITE_P(
    AVX, VP9QuantizeTest,
    ::testing::Values(make_tuple(vpx_quantize_b_avx, vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_avx>,
                                 &Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                                 VPX_BITS_8, 32, false)));
#endif  // HAVE_AVX

#if VPX_ARCH_X86_64 && HAVE_AVX2
#if CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
    AVX2, VP9QuantizeTest,
    ::testing::Values(
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_avx2>,
                   &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 16, true),
        make_tuple(&QuantFPWrapper<vp9_highbd_quantize_fp_avx2>,
                   &QuantFPWrapper<vp9_highbd_quantize_fp_c>, VPX_BITS_12, 16,
                   true),
        make_tuple(&QuantFPWrapper<vp9_highbd_quantize_fp_32x32_avx2>,
                   &QuantFPWrapper<vp9_highbd_quantize_fp_32x32_c>, VPX_BITS_12,
                   32, true),
        make_tuple(vpx_quantize_b_avx2, vpx_quantize_b_c, VPX_BITS_8, 16,
                   false),
        make_tuple(vpx_highbd_quantize_b_avx2, vpx_highbd_quantize_b_c,
                   VPX_BITS_8, 16, false),
        make_tuple(vpx_highbd_quantize_b_avx2, vpx_highbd_quantize_b_c,
                   VPX_BITS_10, 16, false),
        make_tuple(vpx_highbd_quantize_b_avx2, vpx_highbd_quantize_b_c,
                   VPX_BITS_12, 16, false),
        make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_avx2>,
                   &Quant32x32Wrapper<vpx_quantize_b_32x32_c>, VPX_BITS_8, 32,
                   false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_avx2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_8, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_avx2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_10, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_avx2>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_12, 32, false)));
#else
INSTANTIATE_TEST_SUITE_P(
    AVX2, VP9QuantizeTest,
    ::testing::Values(make_tuple(&QuantFPWrapper<vp9_quantize_fp_avx2>,
                                 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8,
                                 16, true),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_avx2>,
                                 &QuantFPWrapper<quantize_fp_32x32_nz_c>,
                                 VPX_BITS_8, 32, true),
                      make_tuple(vpx_quantize_b_avx2, vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_avx2>,
                                 &Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                                 VPX_BITS_8, 32, false)));
#endif  // CONFIG_VP9_HIGHBITDEPTH
#endif  // HAVE_AVX2

#if HAVE_NEON
#if CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
    NEON, VP9QuantizeTest,
    ::testing::Values(
        make_tuple(&vpx_quantize_b_neon, &vpx_quantize_b_c, VPX_BITS_8, 16,
                   false),
        make_tuple(vpx_highbd_quantize_b_neon, vpx_highbd_quantize_b_c,
                   VPX_BITS_8, 16, false),
        make_tuple(vpx_highbd_quantize_b_neon, vpx_highbd_quantize_b_c,
                   VPX_BITS_10, 16, false),
        make_tuple(vpx_highbd_quantize_b_neon, vpx_highbd_quantize_b_c,
                   VPX_BITS_12, 16, false),
        make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_neon>,
                   &Quant32x32Wrapper<vpx_quantize_b_32x32_c>, VPX_BITS_8, 32,
                   false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_neon>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_8, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_neon>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_10, 32, false),
        make_tuple(&Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_neon>,
                   &Quant32x32Wrapper<vpx_highbd_quantize_b_32x32_c>,
                   VPX_BITS_12, 32, false),
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_neon>,
                   &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8, 16, true),
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_neon>,
                   &QuantFPWrapper<vp9_quantize_fp_32x32_c>, VPX_BITS_8, 32,
                   true)));
#else
INSTANTIATE_TEST_SUITE_P(
    NEON, VP9QuantizeTest,
    ::testing::Values(make_tuple(&vpx_quantize_b_neon, &vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_neon>,
                                 &Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                                 VPX_BITS_8, 32, false),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_neon>,
                                 &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8,
                                 16, true),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_neon>,
                                 &QuantFPWrapper<vp9_quantize_fp_32x32_c>,
                                 VPX_BITS_8, 32, true)));
#endif  // CONFIG_VP9_HIGHBITDEPTH
#endif  // HAVE_NEON

#if HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
    VSX, VP9QuantizeTest,
    ::testing::Values(make_tuple(&vpx_quantize_b_vsx, &vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&vpx_quantize_b_32x32_vsx,
                                 &vpx_quantize_b_32x32_c, VPX_BITS_8, 32,
                                 false),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_vsx>,
                                 &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8,
                                 16, true),
                      make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_vsx>,
                                 &QuantFPWrapper<vp9_quantize_fp_32x32_c>,
                                 VPX_BITS_8, 32, true)));
#endif  // HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH

#if HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
    LSX, VP9QuantizeTest,
    ::testing::Values(make_tuple(&vpx_quantize_b_lsx, &vpx_quantize_b_c,
                                 VPX_BITS_8, 16, false),
                      make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_lsx>,
                                 &Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                                 VPX_BITS_8, 32, false)));
#endif  // HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH

// Only useful to compare "Speed" test results.
INSTANTIATE_TEST_SUITE_P(
    DISABLED_C, VP9QuantizeTest,
    ::testing::Values(
        make_tuple(&vpx_quantize_b_c, &vpx_quantize_b_c, VPX_BITS_8, 16, false),
        make_tuple(&Quant32x32Wrapper<vpx_quantize_b_32x32_c>,
                   &Quant32x32Wrapper<vpx_quantize_b_32x32_c>, VPX_BITS_8, 32,
                   false),
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_c>,
                   &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8, 16, true),
        make_tuple(&QuantFPWrapper<quantize_fp_nz_c>,
                   &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 16, true),
        make_tuple(&QuantFPWrapper<quantize_fp_32x32_nz_c>,
                   &QuantFPWrapper<quantize_fp_32x32_nz_c>, VPX_BITS_8, 32,
                   true),
        make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_c>,
                   &QuantFPWrapper<vp9_quantize_fp_32x32_c>, VPX_BITS_8, 32,
                   true)));
}  // namespace