ref: d757d7e998d2e247a6fea424e8cb67e5f2b17951
parent: f63eb66ecd9e0128a0775658b25bd259460ad3b8
parent: 6cc76ec73f02c4c6e78cd77f73b09ea385915f34
author: James Zern <[email protected]>
date: Wed Nov 23 22:31:25 EST 2016
Merge changes Icc4ead05,Ib019964b,I3b5fd3b3,Ieedadee2 * changes: Update vpx_idct4x4_16_add_neon() to pass SingleExtremeCoeff test Refine 8-bit 4x4 idct NEON intrinsics Add idct speed test. Update partial_idct_test.cc to support high bitdepth
--- a/test/partial_idct_test.cc
+++ b/test/partial_idct_test.cc
@@ -25,15 +25,28 @@
#include "vp9/common/vp9_blockd.h"
#include "vp9/common/vp9_scan.h"
#include "vpx/vpx_integer.h"
+#include "vpx_ports/vpx_timer.h"
using libvpx_test::ACMRandom;
namespace {+
typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef uint16_t Pixel;
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride,
+ int bd);
+#else // !CONFIG_VP9_HIGHBITDEPTH
+typedef uint8_t Pixel;
typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
-typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, InvTxfmFunc, TX_SIZE, int>
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, InvTxfmFunc, TX_SIZE, int,
+ int>
PartialInvTxfmParam;
const int kMaxNumCoeffs = 1024;
+const int kCountTestBlock = 1000;
// https://bugs.chromium.org/p/webm/issues/detail?id=1332
// The functions specified do not pass with INT16_MIN/MAX. They fail at the
@@ -52,16 +65,11 @@
int16_t MinSupportedCoeff(InvTxfmFunc a) {(void)a;
-#if !CONFIG_EMULATE_HARDWARE
-#if HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_VP9_HIGHBITDEPTH
+#if HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_VP9_HIGHBITDEPTH && \
+ !CONFIG_EMULATE_HARDWARE
if (a == vpx_idct8x8_64_add_ssse3 || a == vpx_idct8x8_12_add_ssse3) {return -23625 + 1;
}
-#elif HAVE_NEON
- if (a == vpx_idct4x4_16_add_neon) {- return std::numeric_limits<int16_t>::min() + 1;
- }
-#endif
#endif // !CONFIG_EMULATE_HARDWARE
return std::numeric_limits<int16_t>::min();
}
@@ -70,11 +78,14 @@
public:
virtual ~PartialIDctTest() {} virtual void SetUp() {+ rnd_.Reset(ACMRandom::DeterministicSeed());
ftxfm_ = GET_PARAM(0);
full_itxfm_ = GET_PARAM(1);
partial_itxfm_ = GET_PARAM(2);
tx_size_ = GET_PARAM(3);
last_nonzero_ = GET_PARAM(4);
+ bit_depth_ = GET_PARAM(5);
+ mask_ = (1 << bit_depth_) - 1;
switch (tx_size_) {case TX_4X4: size_ = 4; break;
@@ -83,14 +94,26 @@
case TX_32X32: size_ = 32; break;
default: FAIL() << "Wrong Size!"; break;
}
- block_size_ = size_ * size_;
+ // Randomize stride_ to a value less than or equal to 1024
+ stride_ = rnd_(1024) + 1;
+ if (stride_ < size_) {+ stride_ = size_;
+ }
+ // Align stride_ to 16 if it's bigger than 16.
+ if (stride_ > 16) {+ stride_ &= ~15;
+ }
+
+ input_block_size_ = size_ * size_;
+ output_block_size_ = size_ * stride_;
+
input_block_ = reinterpret_cast<tran_low_t *>(
- vpx_memalign(16, sizeof(*input_block_) * block_size_));
- output_block_ = reinterpret_cast<uint8_t *>(
- vpx_memalign(16, sizeof(*output_block_) * block_size_));
- output_block_ref_ = reinterpret_cast<uint8_t *>(
- vpx_memalign(16, sizeof(*output_block_ref_) * block_size_));
+ vpx_memalign(16, sizeof(*input_block_) * input_block_size_));
+ output_block_ = reinterpret_cast<Pixel *>(
+ vpx_memalign(16, sizeof(*output_block_) * output_block_size_));
+ output_block_ref_ = reinterpret_cast<Pixel *>(
+ vpx_memalign(16, sizeof(*output_block_ref_) * output_block_size_));
}
virtual void TearDown() {@@ -103,44 +126,76 @@
libvpx_test::ClearSystemState();
}
+ void InitMem() {+ memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
+ for (int j = 0; j < output_block_size_; ++j) {+ output_block_[j] = output_block_ref_[j] = rnd_.Rand16() & mask_;
+ }
+ }
+
+ void InitInput() {+ const int max_coeff = 32766 / 4;
+ int max_energy_leftover = max_coeff * max_coeff;
+ for (int j = 0; j < last_nonzero_; ++j) {+ int16_t coeff = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
+ (rnd_.Rand16() - 32768) / 65536);
+ max_energy_leftover -= coeff * coeff;
+ if (max_energy_leftover < 0) {+ max_energy_leftover = 0;
+ coeff = 0;
+ }
+ input_block_[vp9_default_scan_orders[tx_size_].scan[j]] = coeff;
+ }
+ }
+
+ void Exec(InvTxfmFunc func, void *out) {+#if CONFIG_VP9_HIGHBITDEPTH
+ func(input_block_, CONVERT_TO_BYTEPTR(out), stride_, bit_depth_);
+#else
+ func(input_block_, reinterpret_cast<uint8_t *>(out), stride_);
+#endif
+ }
+
protected:
int last_nonzero_;
TX_SIZE tx_size_;
tran_low_t *input_block_;
- uint8_t *output_block_;
- uint8_t *output_block_ref_;
+ Pixel *output_block_;
+ Pixel *output_block_ref_;
int size_;
- int block_size_;
+ int stride_;
+ int input_block_size_;
+ int output_block_size_;
+ int bit_depth_;
+ int mask_;
FwdTxfmFunc ftxfm_;
InvTxfmFunc full_itxfm_;
InvTxfmFunc partial_itxfm_;
+ ACMRandom rnd_;
};
TEST_P(PartialIDctTest, RunQuantCheck) {- ACMRandom rnd(ACMRandom::DeterministicSeed());
-
- const int count_test_block = 1000;
-
DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
- for (int i = 0; i < count_test_block; ++i) {- // clear out destination buffer
- memset(input_block_, 0, sizeof(*input_block_) * block_size_);
- memset(output_block_, 0, sizeof(*output_block_) * block_size_);
- memset(output_block_ref_, 0, sizeof(*output_block_ref_) * block_size_);
+ for (int i = 0; i < kCountTestBlock; ++i) {+ InitMem();
ACMRandom rnd(ACMRandom::DeterministicSeed());
- for (int i = 0; i < count_test_block; ++i) {- // Initialize a test block with input range [-255, 255].
- if (i == 0) {- for (int j = 0; j < block_size_; ++j) input_extreme_block[j] = 255;
- } else if (i == 1) {- for (int j = 0; j < block_size_; ++j) input_extreme_block[j] = -255;
+ for (int j = 0; j < kCountTestBlock; ++j) {+ // Initialize a test block with input range [-mask_, mask_].
+ if (j == 0) {+ for (int k = 0; k < input_block_size_; ++k) {+ input_extreme_block[k] = mask_;
+ }
+ } else if (j == 1) {+ for (int k = 0; k < input_block_size_; ++k) {+ input_extreme_block[k] = -mask_;
+ }
} else {- for (int j = 0; j < block_size_; ++j) {- input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+ for (int k = 0; k < input_block_size_; ++k) {+ input_extreme_block[k] = rnd.Rand8() % 2 ? mask_ : -mask_;
}
}
@@ -148,259 +203,356 @@
// quantization with maximum allowed step sizes
input_block_[0] = (output_ref_block[0] / 1336) * 1336;
- for (int j = 1; j < last_nonzero_; ++j) {- input_block_[vp9_default_scan_orders[tx_size_].scan[j]] =
- (output_ref_block[j] / 1828) * 1828;
+ for (int k = 1; k < last_nonzero_; ++k) {+ input_block_[vp9_default_scan_orders[tx_size_].scan[k]] =
+ (output_ref_block[k] / 1828) * 1828;
}
}
- ASM_REGISTER_STATE_CHECK(
- full_itxfm_(input_block_, output_block_ref_, size_));
- ASM_REGISTER_STATE_CHECK(
- partial_itxfm_(input_block_, output_block_, size_));
-
+ ASM_REGISTER_STATE_CHECK(Exec(full_itxfm_, output_block_ref_));
+ ASM_REGISTER_STATE_CHECK(Exec(partial_itxfm_, output_block_));
ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
- sizeof(*output_block_) * block_size_))
+ sizeof(*output_block_) * output_block_size_))
<< "Error: partial inverse transform produces different results";
}
}
TEST_P(PartialIDctTest, ResultsMatch) {- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = 1000;
- const int max_coeff = 32766 / 4;
- for (int i = 0; i < count_test_block; ++i) {- // clear out destination buffer
- memset(input_block_, 0, sizeof(*input_block_) * block_size_);
- memset(output_block_, 0, sizeof(*output_block_) * block_size_);
- memset(output_block_ref_, 0, sizeof(*output_block_ref_) * block_size_);
- int max_energy_leftover = max_coeff * max_coeff;
- for (int j = 0; j < last_nonzero_; ++j) {- int16_t coeff = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
- (rnd.Rand16() - 32768) / 65536);
- max_energy_leftover -= coeff * coeff;
- if (max_energy_leftover < 0) {- max_energy_leftover = 0;
- coeff = 0;
- }
- input_block_[vp9_default_scan_orders[tx_size_].scan[j]] = coeff;
- }
+ for (int i = 0; i < kCountTestBlock; ++i) {+ InitMem();
+ InitInput();
- ASM_REGISTER_STATE_CHECK(
- full_itxfm_(input_block_, output_block_ref_, size_));
- ASM_REGISTER_STATE_CHECK(
- partial_itxfm_(input_block_, output_block_, size_));
+ ASM_REGISTER_STATE_CHECK(Exec(full_itxfm_, output_block_ref_));
+ ASM_REGISTER_STATE_CHECK(Exec(partial_itxfm_, output_block_));
ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
- sizeof(*output_block_) * block_size_))
+ sizeof(*output_block_) * output_block_size_))
<< "Error: partial inverse transform produces different results";
}
}
TEST_P(PartialIDctTest, AddOutputBlock) {- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = 10;
- for (int i = 0; i < count_test_block; ++i) {- memset(input_block_, 0, sizeof(*input_block_) * block_size_);
+ for (int i = 0; i < kCountTestBlock; ++i) {+ InitMem();
for (int j = 0; j < last_nonzero_; ++j) {input_block_[vp9_default_scan_orders[tx_size_].scan[j]] = 10;
}
- for (int j = 0; j < block_size_; ++j) {- output_block_[j] = output_block_ref_[j] = rnd.Rand8();
- }
+ ASM_REGISTER_STATE_CHECK(Exec(full_itxfm_, output_block_ref_));
+ ASM_REGISTER_STATE_CHECK(Exec(partial_itxfm_, output_block_));
- ASM_REGISTER_STATE_CHECK(
- full_itxfm_(input_block_, output_block_ref_, size_));
- ASM_REGISTER_STATE_CHECK(
- partial_itxfm_(input_block_, output_block_, size_));
-
ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
- sizeof(*output_block_) * block_size_))
+ sizeof(*output_block_) * output_block_size_))
<< "Error: Transform results are not correctly added to output.";
}
}
TEST_P(PartialIDctTest, SingleExtremeCoeff) {- ACMRandom rnd(ACMRandom::DeterministicSeed());
const int16_t max_coeff = MaxSupportedCoeff(partial_itxfm_);
const int16_t min_coeff = MinSupportedCoeff(partial_itxfm_);
for (int i = 0; i < last_nonzero_; ++i) {- memset(input_block_, 0, sizeof(*input_block_) * block_size_);
+ memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
// Run once for min and once for max.
for (int j = 0; j < 2; ++j) {const int coeff = j ? min_coeff : max_coeff;
- memset(output_block_, 0, sizeof(*output_block_) * block_size_);
- memset(output_block_ref_, 0, sizeof(*output_block_ref_) * block_size_);
+ memset(output_block_, 0, sizeof(*output_block_) * output_block_size_);
+ memset(output_block_ref_, 0,
+ sizeof(*output_block_ref_) * output_block_size_);
input_block_[vp9_default_scan_orders[tx_size_].scan[i]] = coeff;
- ASM_REGISTER_STATE_CHECK(
- full_itxfm_(input_block_, output_block_ref_, size_));
- ASM_REGISTER_STATE_CHECK(
- partial_itxfm_(input_block_, output_block_, size_));
+ ASM_REGISTER_STATE_CHECK(Exec(full_itxfm_, output_block_ref_));
+ ASM_REGISTER_STATE_CHECK(Exec(partial_itxfm_, output_block_));
ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
- sizeof(*output_block_) * block_size_))
+ sizeof(*output_block_) * output_block_size_))
<< "Error: Fails with single coeff of " << coeff << " at " << i
<< ".";
}
}
}
+
+TEST_P(PartialIDctTest, DISABLED_Speed) {+ // Keep runtime stable with transform size.
+ const int kCountSpeedTestBlock = 500000000 / input_block_size_;
+ InitMem();
+ InitInput();
+
+ for (int i = 0; i < kCountSpeedTestBlock; ++i) {+ ASM_REGISTER_STATE_CHECK(Exec(full_itxfm_, output_block_ref_));
+ }
+ vpx_usec_timer timer;
+ vpx_usec_timer_start(&timer);
+ for (int i = 0; i < kCountSpeedTestBlock; ++i) {+ Exec(partial_itxfm_, output_block_);
+ }
+ libvpx_test::ClearSystemState();
+ vpx_usec_timer_mark(&timer);
+ const int elapsed_time =
+ static_cast<int>(vpx_usec_timer_elapsed(&timer) / 1000);
+ printf("idct%dx%d_%d (bitdepth %d) time: %5d ms ", size_, size_,+ last_nonzero_, bit_depth_, elapsed_time);
+
+ ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+ sizeof(*output_block_) * output_block_size_))
+ << "Error: partial inverse transform produces different results";
+}
+
using std::tr1::make_tuple;
+#if CONFIG_VP9_HIGHBITDEPTH
+
INSTANTIATE_TEST_CASE_P(
C, PartialIDctTest,
+ ::testing::Values(
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1024_add_c, TX_32X32, 1024, 8),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1024_add_c, TX_32X32, 1024, 10),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1024_add_c, TX_32X32, 1024, 12),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_34_add_c, TX_32X32, 34, 8),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_34_add_c, TX_32X32, 34, 10),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_34_add_c, TX_32X32, 34, 12),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_c, TX_32X32, 1, 8),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_c, TX_32X32, 1, 10),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_c, TX_32X32, 1, 12),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_c, TX_16X16, 256, 8),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_c, TX_16X16, 256, 10),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_c, TX_16X16, 256, 12),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_c, TX_16X16, 10, 8),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_c, TX_16X16, 10, 10),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_c, TX_16X16, 10, 12),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_1_add_c, TX_16X16, 1, 8),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_1_add_c, TX_16X16, 1, 10),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_1_add_c, TX_16X16, 1, 12),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_c, TX_8X8, 64, 8),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_c, TX_8X8, 64, 10),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_c, TX_8X8, 64, 12),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_c, TX_8X8, 12, 8),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_c, TX_8X8, 12, 10),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_c, TX_8X8, 12, 12),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_1_add_c, TX_8X8, 1, 8),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_1_add_c, TX_8X8, 1, 10),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_1_add_c, TX_8X8, 1, 12),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_c, TX_4X4, 16, 8),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_c, TX_4X4, 16, 10),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_c, TX_4X4, 16, 12),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_1_add_c, TX_4X4, 1, 8),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_1_add_c, TX_4X4, 1, 10),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_1_add_c, TX_4X4, 1, 12)));
+
+#if HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+ SSE2, PartialIDctTest,
+ ::testing::Values(
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_sse2, TX_32X32, 1, 8),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_sse2, TX_32X32, 1, 10),
+ make_tuple(&vpx_highbd_fdct32x32_c, &vpx_highbd_idct32x32_1024_add_c,
+ &vpx_highbd_idct32x32_1_add_sse2, TX_32X32, 1, 12),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_sse2, TX_16X16, 256, 8),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_sse2, TX_16X16, 256, 10),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_256_add_sse2, TX_16X16, 256, 12),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_sse2, TX_16X16, 10, 8),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_sse2, TX_16X16, 10, 10),
+ make_tuple(&vpx_highbd_fdct16x16_c, &vpx_highbd_idct16x16_256_add_c,
+ &vpx_highbd_idct16x16_10_add_sse2, TX_16X16, 10, 12),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_sse2, TX_8X8, 64, 8),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_sse2, TX_8X8, 64, 10),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_64_add_sse2, TX_8X8, 64, 12),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_sse2, TX_8X8, 12, 8),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_sse2, TX_8X8, 12, 10),
+ make_tuple(&vpx_highbd_fdct8x8_c, &vpx_highbd_idct8x8_64_add_c,
+ &vpx_highbd_idct8x8_12_add_sse2, TX_8X8, 12, 12),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_sse2, TX_4X4, 1, 8),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_sse2, TX_4X4, 1, 10),
+ make_tuple(&vpx_highbd_fdct4x4_c, &vpx_highbd_idct4x4_16_add_c,
+ &vpx_highbd_idct4x4_16_add_sse2, TX_4X4, 1, 12)));
+#endif // HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
+
+#else // !CONFIG_VP9_HIGHBITDEPTH
+
+INSTANTIATE_TEST_CASE_P(
+ C, PartialIDctTest,
::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_c, TX_32X32, 1024),
+ &vpx_idct32x32_1024_add_c, TX_32X32, 1024, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_135_add_c, TX_32X32, 135),
+ &vpx_idct32x32_135_add_c, TX_32X32, 135, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_c, TX_32X32, 34),
+ &vpx_idct32x32_34_add_c, TX_32X32, 34, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1_add_c, TX_32X32, 1),
+ &vpx_idct32x32_1_add_c, TX_32X32, 1, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_256_add_c, TX_16X16, 256),
+ &vpx_idct16x16_256_add_c, TX_16X16, 256, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_10_add_c, TX_16X16, 10),
+ &vpx_idct16x16_10_add_c, TX_16X16, 10, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_1_add_c, TX_16X16, 1),
+ &vpx_idct16x16_1_add_c, TX_16X16, 1, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_c, TX_8X8, 64),
+ &vpx_idct8x8_64_add_c, TX_8X8, 64, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_c, TX_8X8, 12),
+ &vpx_idct8x8_12_add_c, TX_8X8, 12, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_1_add_c, TX_8X8, 1),
+ &vpx_idct8x8_1_add_c, TX_8X8, 1, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_16_add_c, TX_4X4, 16),
+ &vpx_idct4x4_16_add_c, TX_4X4, 16, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_1_add_c, TX_4X4, 1)));
+ &vpx_idct4x4_1_add_c, TX_4X4, 1, 8)));
#if HAVE_NEON && !CONFIG_EMULATE_HARDWARE
-#if CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_CASE_P(
NEON, PartialIDctTest,
::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1_add_neon, TX_32X32, 1),
+ &vpx_idct32x32_1024_add_neon, TX_32X32, 1024,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_neon, TX_32X32, 34),
- make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_1_add_neon, TX_16X16, 1),
- make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_neon, TX_8X8, 64),
- make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_neon, TX_8X8, 12),
- make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_1_add_neon, TX_8X8, 1),
- make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_16_add_neon, TX_4X4, 16),
- make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_1_add_neon, TX_4X4, 1)));
-#else // !CONFIG_VP9_HIGHBITDEPTH
-INSTANTIATE_TEST_CASE_P(
- NEON, PartialIDctTest,
- ::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_neon, TX_32X32, 1024),
+ &vpx_idct32x32_135_add_neon, TX_32X32, 135, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_135_add_neon, TX_32X32, 135),
+ &vpx_idct32x32_34_add_neon, TX_32X32, 34, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_neon, TX_32X32, 34),
- make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1_add_neon, TX_32X32, 1),
+ &vpx_idct32x32_1_add_neon, TX_32X32, 1, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_256_add_neon, TX_16X16, 256),
+ &vpx_idct16x16_256_add_neon, TX_16X16, 256, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_10_add_neon, TX_16X16, 10),
+ &vpx_idct16x16_10_add_neon, TX_16X16, 10, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_1_add_neon, TX_16X16, 1),
+ &vpx_idct16x16_1_add_neon, TX_16X16, 1, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_neon, TX_8X8, 64),
+ &vpx_idct8x8_64_add_neon, TX_8X8, 64, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_neon, TX_8X8, 12),
+ &vpx_idct8x8_12_add_neon, TX_8X8, 12, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_1_add_neon, TX_8X8, 1),
+ &vpx_idct8x8_1_add_neon, TX_8X8, 1, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_16_add_neon, TX_4X4, 16),
+ &vpx_idct4x4_16_add_neon, TX_4X4, 16, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_1_add_neon, TX_4X4, 1)));
-#endif // CONFIG_VP9_HIGHBITDEPTH
+ &vpx_idct4x4_1_add_neon, TX_4X4, 1, 8)));
#endif // HAVE_NEON && !CONFIG_EMULATE_HARDWARE
-#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+#if HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
// 32x32_135_ is implemented using the 1024 version.
INSTANTIATE_TEST_CASE_P(
SSE2, PartialIDctTest,
::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_sse2, TX_32X32, 1024),
+ &vpx_idct32x32_1024_add_sse2, TX_32X32, 1024,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_sse2, TX_32X32, 135),
+ &vpx_idct32x32_1024_add_sse2, TX_32X32, 135,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_sse2, TX_32X32, 34),
+ &vpx_idct32x32_34_add_sse2, TX_32X32, 34, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1_add_sse2, TX_32X32, 1),
+ &vpx_idct32x32_1_add_sse2, TX_32X32, 1, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_256_add_sse2, TX_16X16, 256),
+ &vpx_idct16x16_256_add_sse2, TX_16X16, 256, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_10_add_sse2, TX_16X16, 10),
+ &vpx_idct16x16_10_add_sse2, TX_16X16, 10, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_1_add_sse2, TX_16X16, 1),
+ &vpx_idct16x16_1_add_sse2, TX_16X16, 1, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_sse2, TX_8X8, 64),
+ &vpx_idct8x8_64_add_sse2, TX_8X8, 64, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_sse2, TX_8X8, 12),
+ &vpx_idct8x8_12_add_sse2, TX_8X8, 12, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_1_add_sse2, TX_8X8, 1),
+ &vpx_idct8x8_1_add_sse2, TX_8X8, 1, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_16_add_sse2, TX_4X4, 16),
+ &vpx_idct4x4_16_add_sse2, TX_4X4, 16, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_1_add_sse2, TX_4X4, 1)));
-#endif // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+ &vpx_idct4x4_1_add_sse2, TX_4X4, 1, 8)));
+#endif // HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
-#if HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_VP9_HIGHBITDEPTH && \
- !CONFIG_EMULATE_HARDWARE
+#if HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_EMULATE_HARDWARE
INSTANTIATE_TEST_CASE_P(
SSSE3_64, PartialIDctTest,
::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_ssse3, TX_32X32, 1024),
+ &vpx_idct32x32_1024_add_ssse3, TX_32X32, 1024,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_135_add_ssse3, TX_32X32, 135),
+ &vpx_idct32x32_135_add_ssse3, TX_32X32, 135,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_ssse3, TX_32X32, 34),
+ &vpx_idct32x32_34_add_ssse3, TX_32X32, 34, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_ssse3, TX_8X8, 64),
+ &vpx_idct8x8_64_add_ssse3, TX_8X8, 64, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_ssse3, TX_8X8, 12)));
-#endif // HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_VP9_HIGHBITDEPTH &&
- // !CONFIG_EMULATE_HARDWARE
+ &vpx_idct8x8_12_add_ssse3, TX_8X8, 12, 8)));
+#endif // HAVE_SSSE3 && ARCH_X86_64 && !CONFIG_EMULATE_HARDWARE
-#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+#if HAVE_MSA && !CONFIG_EMULATE_HARDWARE
// 32x32_135_ is implemented using the 1024 version.
INSTANTIATE_TEST_CASE_P(
MSA, PartialIDctTest,
::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_msa, TX_32X32, 1024),
+ &vpx_idct32x32_1024_add_msa, TX_32X32, 1024,
+ 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1024_add_msa, TX_32X32, 135),
+ &vpx_idct32x32_1024_add_msa, TX_32X32, 135, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_34_add_msa, TX_32X32, 34),
+ &vpx_idct32x32_34_add_msa, TX_32X32, 34, 8),
make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,
- &vpx_idct32x32_1_add_msa, TX_32X32, 1),
+ &vpx_idct32x32_1_add_msa, TX_32X32, 1, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_256_add_msa, TX_16X16, 256),
+ &vpx_idct16x16_256_add_msa, TX_16X16, 256, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_10_add_msa, TX_16X16, 10),
+ &vpx_idct16x16_10_add_msa, TX_16X16, 10, 8),
make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,
- &vpx_idct16x16_1_add_msa, TX_16X16, 1),
+ &vpx_idct16x16_1_add_msa, TX_16X16, 1, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_64_add_msa, TX_8X8, 64),
+ &vpx_idct8x8_64_add_msa, TX_8X8, 64, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_12_add_msa, TX_8X8, 10),
+ &vpx_idct8x8_12_add_msa, TX_8X8, 10, 8),
make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,
- &vpx_idct8x8_1_add_msa, TX_8X8, 1),
+ &vpx_idct8x8_1_add_msa, TX_8X8, 1, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_16_add_msa, TX_4X4, 16),
+ &vpx_idct4x4_16_add_msa, TX_4X4, 16, 8),
make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c,
- &vpx_idct4x4_1_add_msa, TX_4X4, 1)));
-#endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+ &vpx_idct4x4_1_add_msa, TX_4X4, 1, 8)));
+#endif // HAVE_MSA && !CONFIG_EMULATE_HARDWARE
+
+#endif // CONFIG_VP9_HIGHBITDEPTH
} // namespace
--- a/vpx_dsp/arm/idct4x4_1_add_neon.c
+++ b/vpx_dsp/arm/idct4x4_1_add_neon.c
@@ -9,39 +9,33 @@
*/
#include <arm_neon.h>
+#include <assert.h>
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/inv_txfm.h"
-#include "vpx_ports/mem.h"
void vpx_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride) {- uint8x8_t d6u8;
- uint32x2_t d2u32 = vdup_n_u32(0);
- uint16x8_t q8u16;
- int16x8_t q0s16;
- uint8_t *d1, *d2;
- int16_t i, a1;
- int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
- out = dct_const_round_shift(out * cospi_16_64);
- a1 = ROUND_POWER_OF_TWO(out, 4);
+ int i;
+ const int16_t out0 = dct_const_round_shift((int16_t)input[0] * cospi_16_64);
+ const int16_t out1 = dct_const_round_shift(out0 * cospi_16_64);
+ const int16_t a1 = ROUND_POWER_OF_TWO(out1, 4);
+ const int16x8_t dc = vdupq_n_s16(a1);
+ uint32x2_t d = vdup_n_u32(0);
+ uint16x8_t a;
+ uint8x8_t b;
- q0s16 = vdupq_n_s16(a1);
+ assert(!((intptr_t)dest % sizeof(uint32_t)));
+ assert(!(dest_stride % sizeof(uint32_t)));
- // dc_only_idct_add
- d1 = d2 = dest;
for (i = 0; i < 2; i++) {- d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
- d1 += dest_stride;
- d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
- d1 += dest_stride;
-
- q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16), vreinterpret_u8_u32(d2u32));
- d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-
- vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
- d2 += dest_stride;
- vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
- d2 += dest_stride;
+ d = vld1_lane_u32((const uint32_t *)dest, d, 0);
+ d = vld1_lane_u32((const uint32_t *)(dest + dest_stride), d, 1);
+ a = vaddw_u8(vreinterpretq_u16_s16(dc), vreinterpret_u8_u32(d));
+ b = vqmovun_s16(vreinterpretq_s16_u16(a));
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(b), 0);
+ dest += dest_stride;
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(b), 1);
+ dest += dest_stride;
}
}
--- a/vpx_dsp/arm/idct4x4_add_neon.asm
+++ b/vpx_dsp/arm/idct4x4_add_neon.asm
@@ -72,16 +72,15 @@
; do the transform on transposed rows
; stage 1
- vadd.s16 d23, d16, d18 ; (input[0] + input[2])
- vsub.s16 d24, d16, d18 ; (input[0] - input[2])
-
vmull.s16 q15, d17, d22 ; input[1] * cospi_24_64
vmull.s16 q1, d17, d20 ; input[1] * cospi_8_64
; (input[0] + input[2]) * cospi_16_64;
; (input[0] - input[2]) * cospi_16_64;
- vmull.s16 q13, d23, d21
- vmull.s16 q14, d24, d21
+ vmull.s16 q8, d16, d21
+ vmull.s16 q14, d18, d21
+ vadd.s32 q13, q8, q14
+ vsub.s32 q14, q8, q14
; input[1] * cospi_24_64 - input[3] * cospi_8_64;
; input[1] * cospi_8_64 + input[3] * cospi_24_64;
--- a/vpx_dsp/arm/idct4x4_add_neon.c
+++ b/vpx_dsp/arm/idct4x4_add_neon.c
@@ -9,139 +9,89 @@
*/
#include <arm_neon.h>
+#include <assert.h>
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/txfm_common.h"
+static INLINE void idct4x4_16_kernel(const int16x4_t cospis, int16x8_t *a0,
+ int16x8_t *a1) {+ int16x4_t b0, b1, b2, b3;
+ int32x4_t c0, c1, c2, c3;
+ int16x8_t d0, d1;
+
+ transpose_s16_4x4q(a0, a1);
+ b0 = vget_low_s16(*a0);
+ b1 = vget_high_s16(*a0);
+ b2 = vget_low_s16(*a1);
+ b3 = vget_high_s16(*a1);
+ c0 = vmull_lane_s16(b0, cospis, 2);
+ c2 = vmull_lane_s16(b1, cospis, 2);
+ c1 = vsubq_s32(c0, c2);
+ c0 = vaddq_s32(c0, c2);
+ c2 = vmull_lane_s16(b2, cospis, 3);
+ c3 = vmull_lane_s16(b2, cospis, 1);
+ c2 = vmlsl_lane_s16(c2, b3, cospis, 1);
+ c3 = vmlal_lane_s16(c3, b3, cospis, 3);
+ b0 = vrshrn_n_s32(c0, 14);
+ b1 = vrshrn_n_s32(c1, 14);
+ b2 = vrshrn_n_s32(c2, 14);
+ b3 = vrshrn_n_s32(c3, 14);
+ d0 = vcombine_s16(b0, b1);
+ d1 = vcombine_s16(b3, b2);
+ *a0 = vaddq_s16(d0, d1);
+ *a1 = vsubq_s16(d0, d1);
+}
+
void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride) {- uint8x8_t d26u8, d27u8;
- uint32x2_t d26u32, d27u32;
- uint16x8_t q8u16, q9u16;
- int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
- int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
- int16x8_t q8s16, q9s16, q13s16, q14s16;
- int32x4_t q1s32, q13s32, q14s32, q15s32;
- int16x4x2_t d0x2s16, d1x2s16;
- int32x4x2_t q0x2s32;
- uint8_t *d;
+ DECLARE_ALIGNED(16, static const int16_t, cospi[4]) = {+ 0, (int16_t)cospi_8_64, (int16_t)cospi_16_64, (int16_t)cospi_24_64
+ };
+ const uint8_t *dst = dest;
+ const int16x4_t cospis = vld1_s16(cospi);
+ uint32x2_t dest01_u32 = vdup_n_u32(0);
+ uint32x2_t dest32_u32 = vdup_n_u32(0);
+ int16x8_t a0, a1;
+ uint8x8_t d01, d32;
+ uint16x8_t d01_u16, d32_u16;
- d26u32 = d27u32 = vdup_n_u32(0);
+ assert(!((intptr_t)dest % sizeof(uint32_t)));
+ assert(!(dest_stride % sizeof(uint32_t)));
- q8s16 = load_tran_low_to_s16(input);
- q9s16 = load_tran_low_to_s16(input + 8);
+ // Rows
+ a0 = load_tran_low_to_s16(input);
+ a1 = load_tran_low_to_s16(input + 8);
+ idct4x4_16_kernel(cospis, &a0, &a1);
- d16s16 = vget_low_s16(q8s16);
- d17s16 = vget_high_s16(q8s16);
- d18s16 = vget_low_s16(q9s16);
- d19s16 = vget_high_s16(q9s16);
+ // Columns
+ a1 = vcombine_s16(vget_high_s16(a1), vget_low_s16(a1));
+ idct4x4_16_kernel(cospis, &a0, &a1);
+ a0 = vrshrq_n_s16(a0, 4);
+ a1 = vrshrq_n_s16(a1, 4);
- d0x2s16 = vtrn_s16(d16s16, d17s16);
- d1x2s16 = vtrn_s16(d18s16, d19s16);
- q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
- q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+ dest01_u32 = vld1_lane_u32((const uint32_t *)dst, dest01_u32, 0);
+ dst += dest_stride;
+ dest01_u32 = vld1_lane_u32((const uint32_t *)dst, dest01_u32, 1);
+ dst += dest_stride;
+ dest32_u32 = vld1_lane_u32((const uint32_t *)dst, dest32_u32, 1);
+ dst += dest_stride;
+ dest32_u32 = vld1_lane_u32((const uint32_t *)dst, dest32_u32, 0);
- d20s16 = vdup_n_s16((int16_t)cospi_8_64);
- d21s16 = vdup_n_s16((int16_t)cospi_16_64);
+ d01_u16 =
+ vaddw_u8(vreinterpretq_u16_s16(a0), vreinterpret_u8_u32(dest01_u32));
+ d32_u16 =
+ vaddw_u8(vreinterpretq_u16_s16(a1), vreinterpret_u8_u32(dest32_u32));
+ d01 = vqmovun_s16(vreinterpretq_s16_u16(d01_u16));
+ d32 = vqmovun_s16(vreinterpretq_s16_u16(d32_u16));
- q0x2s32 =
- vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
- d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
- d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
- d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
- d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
-
- d22s16 = vdup_n_s16((int16_t)cospi_24_64);
-
- // stage 1
- d23s16 = vadd_s16(d16s16, d18s16);
- d24s16 = vsub_s16(d16s16, d18s16);
-
- q15s32 = vmull_s16(d17s16, d22s16);
- q1s32 = vmull_s16(d17s16, d20s16);
- q13s32 = vmull_s16(d23s16, d21s16);
- q14s32 = vmull_s16(d24s16, d21s16);
-
- q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
- q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
-
- d26s16 = vqrshrn_n_s32(q13s32, 14);
- d27s16 = vqrshrn_n_s32(q14s32, 14);
- d29s16 = vqrshrn_n_s32(q15s32, 14);
- d28s16 = vqrshrn_n_s32(q1s32, 14);
- q13s16 = vcombine_s16(d26s16, d27s16);
- q14s16 = vcombine_s16(d28s16, d29s16);
-
- // stage 2
- q8s16 = vaddq_s16(q13s16, q14s16);
- q9s16 = vsubq_s16(q13s16, q14s16);
-
- d16s16 = vget_low_s16(q8s16);
- d17s16 = vget_high_s16(q8s16);
- d18s16 = vget_high_s16(q9s16); // vswp d18 d19
- d19s16 = vget_low_s16(q9s16);
-
- d0x2s16 = vtrn_s16(d16s16, d17s16);
- d1x2s16 = vtrn_s16(d18s16, d19s16);
- q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
- q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
-
- q0x2s32 =
- vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
- d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
- d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
- d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
- d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
-
- // do the transform on columns
- // stage 1
- d23s16 = vadd_s16(d16s16, d18s16);
- d24s16 = vsub_s16(d16s16, d18s16);
-
- q15s32 = vmull_s16(d17s16, d22s16);
- q1s32 = vmull_s16(d17s16, d20s16);
- q13s32 = vmull_s16(d23s16, d21s16);
- q14s32 = vmull_s16(d24s16, d21s16);
-
- q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
- q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
-
- d26s16 = vqrshrn_n_s32(q13s32, 14);
- d27s16 = vqrshrn_n_s32(q14s32, 14);
- d29s16 = vqrshrn_n_s32(q15s32, 14);
- d28s16 = vqrshrn_n_s32(q1s32, 14);
- q13s16 = vcombine_s16(d26s16, d27s16);
- q14s16 = vcombine_s16(d28s16, d29s16);
-
- // stage 2
- q8s16 = vaddq_s16(q13s16, q14s16);
- q9s16 = vsubq_s16(q13s16, q14s16);
-
- q8s16 = vrshrq_n_s16(q8s16, 4);
- q9s16 = vrshrq_n_s16(q9s16, 4);
-
- d = dest;
- d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
- d += dest_stride;
- d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
- d += dest_stride;
- d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
- d += dest_stride;
- d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
-
- q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
- q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
-
- d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
- d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-
- d = dest;
- vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
- d += dest_stride;
- vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
- d += dest_stride;
- vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
- d += dest_stride;
- vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d01), 0);
+ dest += dest_stride;
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d01), 1);
+ dest += dest_stride;
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d32), 1);
+ dest += dest_stride;
+ vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d32), 0);
}
--- a/vpx_dsp/arm/transpose_neon.h
+++ b/vpx_dsp/arm/transpose_neon.h
@@ -110,6 +110,37 @@
*a3 = vreinterpret_s16_s32(c1.val[1]);
}
+static INLINE void transpose_s16_4x4q(int16x8_t *a0, int16x8_t *a1) {+ // Swap 32 bit elements. Goes from:
+ // a0: 00 01 02 03 10 11 12 13
+ // a1: 20 21 22 23 30 31 32 33
+ // to:
+ // b0.val[0]: 00 01 20 21 10 11 30 31
+ // b0.val[1]: 02 03 22 23 12 13 32 33
+
+ const int32x4x2_t b0 =
+ vtrnq_s32(vreinterpretq_s32_s16(*a0), vreinterpretq_s32_s16(*a1));
+
+ // Swap 64 bit elements resulting in:
+ // c0.val[0]: 00 01 20 21 02 03 22 23
+ // c0.val[1]: 10 11 30 31 12 13 32 33
+
+ const int32x4_t c0 =
+ vcombine_s32(vget_low_s32(b0.val[0]), vget_low_s32(b0.val[1]));
+ const int32x4_t c1 =
+ vcombine_s32(vget_high_s32(b0.val[0]), vget_high_s32(b0.val[1]));
+
+ // Swap 16 bit elements resulting in:
+ // d0.val[0]: 00 10 20 30 02 12 22 32
+ // d0.val[1]: 01 11 21 31 03 13 23 33
+
+ const int16x8x2_t d0 =
+ vtrnq_s16(vreinterpretq_s16_s32(c0), vreinterpretq_s16_s32(c1));
+
+ *a0 = d0.val[0];
+ *a1 = d0.val[1];
+}
+
static INLINE void transpose_u16_4x4q(uint16x8_t *a0, uint16x8_t *a1) {// Swap 32 bit elements. Goes from:
// a0: 00 01 02 03 10 11 12 13