generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
dc0f457c94
vpx/test/variance_test.cc

689 lines
29 KiB
C++
Raw Normal View History

Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
/*
* Copyright (c) 2012 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 <stdlib.h>
#include <new>
#include "third_party/googletest/src/include/gtest/gtest.h"
tests: clear system state after non-API calls add ClearSystemState() to reset MMX registers avoiding corrupting subsequent tests. Change-Id: I668deb09aa7aa467709776e5819f936910698bc0
2013-06-18 07:58:40 +02:00
#include "test/clear_system_state.h"
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
#include "test/register_state_check.h"
tests: clear system state after non-API calls add ClearSystemState() to reset MMX registers avoiding corrupting subsequent tests. Change-Id: I668deb09aa7aa467709776e5819f936910698bc0
2013-06-18 07:58:40 +02:00
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#include "vpx/vpx_integer.h"
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
#include "vpx_config.h"
extern "C" {
Allocate memory using appropriate expected alignment in unit tests. Fixes crashes of test_libvpx on 32-bit Linux. Change-Id: If94e7628a86b788ca26c004861dee2f162e47ed6
2013-06-22 02:03:57 +02:00
#include "vpx_mem/vpx_mem.h"
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#if CONFIG_VP8_ENCODER
# include "vp8/common/variance.h"
# include "vp8_rtcd.h"
#endif
#if CONFIG_VP9_ENCODER
# include "vp9/encoder/vp9_variance.h"
# include "vp9_rtcd.h"
#endif
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
}
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
#include "test/acm_random.h"
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
namespace {
using ::std::tr1::get;
using ::std::tr1::make_tuple;
using ::std::tr1::tuple;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
using libvpx_test::ACMRandom;
static unsigned int variance_ref(const uint8_t *ref, const uint8_t *src,
int l2w, int l2h, unsigned int *sse_ptr) {
int se = 0;
unsigned int sse = 0;
const int w = 1 << l2w, h = 1 << l2h;
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++) {
int diff = ref[w * y + x] - src[w * y + x];
se += diff;
sse += diff * diff;
}
}
*sse_ptr = sse;
return sse - (((int64_t) se * se) >> (l2w + l2h));
}
static unsigned int subpel_variance_ref(const uint8_t *ref, const uint8_t *src,
int l2w, int l2h, int xoff, int yoff,
unsigned int *sse_ptr) {
int se = 0;
unsigned int sse = 0;
const int w = 1 << l2w, h = 1 << l2h;
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++) {
// bilinear interpolation at a 16th pel step
const int a1 = ref[(w + 1) * (y + 0) + x + 0];
const int a2 = ref[(w + 1) * (y + 0) + x + 1];
const int b1 = ref[(w + 1) * (y + 1) + x + 0];
const int b2 = ref[(w + 1) * (y + 1) + x + 1];
const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
const int r = a + (((b - a) * yoff + 8) >> 4);
int diff = r - src[w * y + x];
se += diff;
sse += diff * diff;
}
}
*sse_ptr = sse;
return sse - (((int64_t) se * se) >> (l2w + l2h));
}
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
static unsigned int subpel_avg_variance_ref(const uint8_t *ref,
const uint8_t *src,
const uint8_t *second_pred,
int l2w, int l2h,
int xoff, int yoff,
unsigned int *sse_ptr) {
int se = 0;
unsigned int sse = 0;
const int w = 1 << l2w, h = 1 << l2h;
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++) {
// bilinear interpolation at a 16th pel step
const int a1 = ref[(w + 1) * (y + 0) + x + 0];
const int a2 = ref[(w + 1) * (y + 0) + x + 1];
const int b1 = ref[(w + 1) * (y + 1) + x + 0];
const int b2 = ref[(w + 1) * (y + 1) + x + 1];
const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
const int r = a + (((b - a) * yoff + 8) >> 4);
int diff = ((r + second_pred[w * y + x] + 1) >> 1) - src[w * y + x];
se += diff;
sse += diff * diff;
}
}
*sse_ptr = sse;
return sse - (((int64_t) se * se) >> (l2w + l2h));
}
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
template<typename VarianceFunctionType>
class VarianceTest :
public ::testing::TestWithParam<tuple<int, int, VarianceFunctionType> > {
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
public:
virtual void SetUp() {
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
const tuple<int, int, VarianceFunctionType>& params = this->GetParam();
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
log2width_ = get<0>(params);
width_ = 1 << log2width_;
log2height_ = get<1>(params);
height_ = 1 << log2height_;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
variance_ = get<2>(params);
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
rnd(ACMRandom::DeterministicSeed());
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
block_size_ = width_ * height_;
src_ = new uint8_t[block_size_];
ref_ = new uint8_t[block_size_];
ASSERT_TRUE(src_ != NULL);
ASSERT_TRUE(ref_ != NULL);
}
virtual void TearDown() {
delete[] src_;
delete[] ref_;
tests: clear system state after non-API calls add ClearSystemState() to reset MMX registers avoiding corrupting subsequent tests. Change-Id: I668deb09aa7aa467709776e5819f936910698bc0
2013-06-18 07:58:40 +02:00
libvpx_test::ClearSystemState();
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
}
protected:
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
void ZeroTest();
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
void RefTest();
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
void OneQuarterTest();
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
ACMRandom rnd;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
uint8_t* src_;
uint8_t* ref_;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
int width_, log2width_;
int height_, log2height_;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
int block_size_;
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
VarianceFunctionType variance_;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
};
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
template<typename VarianceFunctionType>
void VarianceTest<VarianceFunctionType>::ZeroTest() {
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
for (int i = 0; i <= 255; ++i) {
memset(src_, i, block_size_);
for (int j = 0; j <= 255; ++j) {
memset(ref_, j, block_size_);
unsigned int sse;
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
unsigned int var;
REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
EXPECT_EQ(0u, var) << "src values: " << i << "ref values: " << j;
}
}
}
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
template<typename VarianceFunctionType>
void VarianceTest<VarianceFunctionType>::RefTest() {
for (int i = 0; i < 10; ++i) {
for (int j = 0; j < block_size_; j++) {
src_[j] = rnd.Rand8();
ref_[j] = rnd.Rand8();
}
unsigned int sse1, sse2;
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
unsigned int var1;
REGISTER_STATE_CHECK(var1 = variance_(src_, width_, ref_, width_, &sse1));
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
const unsigned int var2 = variance_ref(src_, ref_, log2width_,
log2height_, &sse2);
EXPECT_EQ(sse1, sse2);
EXPECT_EQ(var1, var2);
}
}
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
template<typename VarianceFunctionType>
void VarianceTest<VarianceFunctionType>::OneQuarterTest() {
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
memset(src_, 255, block_size_);
const int half = block_size_ / 2;
memset(ref_, 255, half);
memset(ref_ + half, 0, half);
unsigned int sse;
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
unsigned int var;
REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
const unsigned int expected = block_size_ * 255 * 255 / 4;
EXPECT_EQ(expected, var);
}
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
template<typename SubpelVarianceFunctionType>
class SubpelVarianceTest :
public ::testing::TestWithParam<tuple<int, int,
SubpelVarianceFunctionType> > {
public:
virtual void SetUp() {
const tuple<int, int, SubpelVarianceFunctionType>& params =
this->GetParam();
log2width_ = get<0>(params);
width_ = 1 << log2width_;
log2height_ = get<1>(params);
height_ = 1 << log2height_;
subpel_variance_ = get<2>(params);
rnd(ACMRandom::DeterministicSeed());
block_size_ = width_ * height_;
Allocate memory using appropriate expected alignment in unit tests. Fixes crashes of test_libvpx on 32-bit Linux. Change-Id: If94e7628a86b788ca26c004861dee2f162e47ed6
2013-06-22 02:03:57 +02:00
src_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
sec_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
ref_ = new uint8_t[block_size_ + width_ + height_ + 1];
ASSERT_TRUE(src_ != NULL);
Allocate memory using appropriate expected alignment in unit tests. Fixes crashes of test_libvpx on 32-bit Linux. Change-Id: If94e7628a86b788ca26c004861dee2f162e47ed6
2013-06-22 02:03:57 +02:00
ASSERT_TRUE(sec_ != NULL);
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
ASSERT_TRUE(ref_ != NULL);
}
virtual void TearDown() {
Allocate memory using appropriate expected alignment in unit tests. Fixes crashes of test_libvpx on 32-bit Linux. Change-Id: If94e7628a86b788ca26c004861dee2f162e47ed6
2013-06-22 02:03:57 +02:00
vpx_free(src_);
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
delete[] ref_;
Allocate memory using appropriate expected alignment in unit tests. Fixes crashes of test_libvpx on 32-bit Linux. Change-Id: If94e7628a86b788ca26c004861dee2f162e47ed6
2013-06-22 02:03:57 +02:00
vpx_free(sec_);
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
}
protected:
void RefTest();
ACMRandom rnd;
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
uint8_t *src_;
uint8_t *ref_;
uint8_t *sec_;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
int width_, log2width_;
int height_, log2height_;
int block_size_;
SubpelVarianceFunctionType subpel_variance_;
};
template<typename SubpelVarianceFunctionType>
void SubpelVarianceTest<SubpelVarianceFunctionType>::RefTest() {
for (int x = 0; x < 16; ++x) {
for (int y = 0; y < 16; ++y) {
for (int j = 0; j < block_size_; j++) {
src_[j] = rnd.Rand8();
}
for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
ref_[j] = rnd.Rand8();
}
unsigned int sse1, sse2;
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
unsigned int var1;
REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
src_, width_, &sse1));
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
const unsigned int var2 = subpel_variance_ref(ref_, src_, log2width_,
log2height_, x, y, &sse2);
EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
}
}
}
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
template<>
void SubpelVarianceTest<vp9_subp_avg_variance_fn_t>::RefTest() {
for (int x = 0; x < 16; ++x) {
for (int y = 0; y < 16; ++y) {
for (int j = 0; j < block_size_; j++) {
src_[j] = rnd.Rand8();
sec_[j] = rnd.Rand8();
}
for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
ref_[j] = rnd.Rand8();
}
unsigned int sse1, sse2;
variance_test: use REGISTER_STATE_CHECK Change-Id: Id54ad9a781634f075e990d5bade5be8490959975
2013-06-19 04:15:56 +02:00
unsigned int var1;
REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
src_, width_, &sse1, sec_));
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
const unsigned int var2 = subpel_avg_variance_ref(ref_, src_, sec_,
log2width_, log2height_,
x, y, &sse2);
EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
}
}
}
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
// -----------------------------------------------------------------------------
// VP8 test cases.
namespace vp8 {
#if CONFIG_VP8_ENCODER
typedef VarianceTest<vp8_variance_fn_t> VP8VarianceTest;
TEST_P(VP8VarianceTest, Zero) { ZeroTest(); }
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
TEST_P(VP8VarianceTest, Ref) { RefTest(); }
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
TEST_P(VP8VarianceTest, OneQuarter) { OneQuarterTest(); }
const vp8_variance_fn_t variance4x4_c = vp8_variance4x4_c;
const vp8_variance_fn_t variance8x8_c = vp8_variance8x8_c;
const vp8_variance_fn_t variance8x16_c = vp8_variance8x16_c;
const vp8_variance_fn_t variance16x8_c = vp8_variance16x8_c;
const vp8_variance_fn_t variance16x16_c = vp8_variance16x16_c;
INSTANTIATE_TEST_CASE_P(
C, VP8VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_c),
make_tuple(3, 3, variance8x8_c),
make_tuple(3, 4, variance8x16_c),
make_tuple(4, 3, variance16x8_c),
make_tuple(4, 4, variance16x16_c)));
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#if HAVE_MMX
const vp8_variance_fn_t variance4x4_mmx = vp8_variance4x4_mmx;
const vp8_variance_fn_t variance8x8_mmx = vp8_variance8x8_mmx;
const vp8_variance_fn_t variance8x16_mmx = vp8_variance8x16_mmx;
const vp8_variance_fn_t variance16x8_mmx = vp8_variance16x8_mmx;
const vp8_variance_fn_t variance16x16_mmx = vp8_variance16x16_mmx;
INSTANTIATE_TEST_CASE_P(
MMX, VP8VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_mmx),
make_tuple(3, 3, variance8x8_mmx),
make_tuple(3, 4, variance8x16_mmx),
make_tuple(4, 3, variance16x8_mmx),
make_tuple(4, 4, variance16x16_mmx)));
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#endif
#if HAVE_SSE2
const vp8_variance_fn_t variance4x4_wmt = vp8_variance4x4_wmt;
const vp8_variance_fn_t variance8x8_wmt = vp8_variance8x8_wmt;
const vp8_variance_fn_t variance8x16_wmt = vp8_variance8x16_wmt;
const vp8_variance_fn_t variance16x8_wmt = vp8_variance16x8_wmt;
const vp8_variance_fn_t variance16x16_wmt = vp8_variance16x16_wmt;
INSTANTIATE_TEST_CASE_P(
SSE2, VP8VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_wmt),
make_tuple(3, 3, variance8x8_wmt),
make_tuple(3, 4, variance8x16_wmt),
make_tuple(4, 3, variance16x8_wmt),
make_tuple(4, 4, variance16x16_wmt)));
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#endif
#endif // CONFIG_VP8_ENCODER
} // namespace vp8
// -----------------------------------------------------------------------------
// VP9 test cases.
namespace vp9 {
#if CONFIG_VP9_ENCODER
typedef VarianceTest<vp9_variance_fn_t> VP9VarianceTest;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
typedef SubpelVarianceTest<vp9_subpixvariance_fn_t> VP9SubpelVarianceTest;
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
typedef SubpelVarianceTest<vp9_subp_avg_variance_fn_t> VP9SubpelAvgVarianceTest;
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
TEST_P(VP9VarianceTest, Zero) { ZeroTest(); }
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
TEST_P(VP9VarianceTest, Ref) { RefTest(); }
TEST_P(VP9SubpelVarianceTest, Ref) { RefTest(); }
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
TEST_P(VP9SubpelAvgVarianceTest, Ref) { RefTest(); }
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
TEST_P(VP9VarianceTest, OneQuarter) { OneQuarterTest(); }
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
const vp9_variance_fn_t variance4x4_c = vp9_variance4x4_c;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
const vp9_variance_fn_t variance4x8_c = vp9_variance4x8_c;
const vp9_variance_fn_t variance8x4_c = vp9_variance8x4_c;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
const vp9_variance_fn_t variance8x8_c = vp9_variance8x8_c;
const vp9_variance_fn_t variance8x16_c = vp9_variance8x16_c;
const vp9_variance_fn_t variance16x8_c = vp9_variance16x8_c;
const vp9_variance_fn_t variance16x16_c = vp9_variance16x16_c;
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
const vp9_variance_fn_t variance16x32_c = vp9_variance16x32_c;
const vp9_variance_fn_t variance32x16_c = vp9_variance32x16_c;
const vp9_variance_fn_t variance32x32_c = vp9_variance32x32_c;
const vp9_variance_fn_t variance32x64_c = vp9_variance32x64_c;
const vp9_variance_fn_t variance64x32_c = vp9_variance64x32_c;
const vp9_variance_fn_t variance64x64_c = vp9_variance64x64_c;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
INSTANTIATE_TEST_CASE_P(
C, VP9VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_c),
make_tuple(2, 3, variance4x8_c),
make_tuple(3, 2, variance8x4_c),
make_tuple(3, 3, variance8x8_c),
make_tuple(3, 4, variance8x16_c),
make_tuple(4, 3, variance16x8_c),
make_tuple(4, 4, variance16x16_c),
make_tuple(4, 5, variance16x32_c),
make_tuple(5, 4, variance32x16_c),
make_tuple(5, 5, variance32x32_c),
make_tuple(5, 6, variance32x64_c),
make_tuple(6, 5, variance64x32_c),
make_tuple(6, 6, variance64x64_c)));
const vp9_subpixvariance_fn_t subpel_variance4x4_c =
vp9_sub_pixel_variance4x4_c;
const vp9_subpixvariance_fn_t subpel_variance4x8_c =
vp9_sub_pixel_variance4x8_c;
const vp9_subpixvariance_fn_t subpel_variance8x4_c =
vp9_sub_pixel_variance8x4_c;
const vp9_subpixvariance_fn_t subpel_variance8x8_c =
vp9_sub_pixel_variance8x8_c;
const vp9_subpixvariance_fn_t subpel_variance8x16_c =
vp9_sub_pixel_variance8x16_c;
const vp9_subpixvariance_fn_t subpel_variance16x8_c =
vp9_sub_pixel_variance16x8_c;
const vp9_subpixvariance_fn_t subpel_variance16x16_c =
vp9_sub_pixel_variance16x16_c;
const vp9_subpixvariance_fn_t subpel_variance16x32_c =
vp9_sub_pixel_variance16x32_c;
const vp9_subpixvariance_fn_t subpel_variance32x16_c =
vp9_sub_pixel_variance32x16_c;
const vp9_subpixvariance_fn_t subpel_variance32x32_c =
vp9_sub_pixel_variance32x32_c;
const vp9_subpixvariance_fn_t subpel_variance32x64_c =
vp9_sub_pixel_variance32x64_c;
const vp9_subpixvariance_fn_t subpel_variance64x32_c =
vp9_sub_pixel_variance64x32_c;
const vp9_subpixvariance_fn_t subpel_variance64x64_c =
vp9_sub_pixel_variance64x64_c;
INSTANTIATE_TEST_CASE_P(
C, VP9SubpelVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_variance4x4_c),
make_tuple(2, 3, subpel_variance4x8_c),
make_tuple(3, 2, subpel_variance8x4_c),
make_tuple(3, 3, subpel_variance8x8_c),
make_tuple(3, 4, subpel_variance8x16_c),
make_tuple(4, 3, subpel_variance16x8_c),
make_tuple(4, 4, subpel_variance16x16_c),
make_tuple(4, 5, subpel_variance16x32_c),
make_tuple(5, 4, subpel_variance32x16_c),
make_tuple(5, 5, subpel_variance32x32_c),
make_tuple(5, 6, subpel_variance32x64_c),
make_tuple(6, 5, subpel_variance64x32_c),
make_tuple(6, 6, subpel_variance64x64_c)));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x4_c =
vp9_sub_pixel_avg_variance4x4_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x8_c =
vp9_sub_pixel_avg_variance4x8_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x4_c =
vp9_sub_pixel_avg_variance8x4_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x8_c =
vp9_sub_pixel_avg_variance8x8_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x16_c =
vp9_sub_pixel_avg_variance8x16_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x8_c =
vp9_sub_pixel_avg_variance16x8_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x16_c =
vp9_sub_pixel_avg_variance16x16_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x32_c =
vp9_sub_pixel_avg_variance16x32_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x16_c =
vp9_sub_pixel_avg_variance32x16_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x32_c =
vp9_sub_pixel_avg_variance32x32_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x64_c =
vp9_sub_pixel_avg_variance32x64_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x32_c =
vp9_sub_pixel_avg_variance64x32_c;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x64_c =
vp9_sub_pixel_avg_variance64x64_c;
INSTANTIATE_TEST_CASE_P(
C, VP9SubpelAvgVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_avg_variance4x4_c),
make_tuple(2, 3, subpel_avg_variance4x8_c),
make_tuple(3, 2, subpel_avg_variance8x4_c),
make_tuple(3, 3, subpel_avg_variance8x8_c),
make_tuple(3, 4, subpel_avg_variance8x16_c),
make_tuple(4, 3, subpel_avg_variance16x8_c),
make_tuple(4, 4, subpel_avg_variance16x16_c),
make_tuple(4, 5, subpel_avg_variance16x32_c),
make_tuple(5, 4, subpel_avg_variance32x16_c),
make_tuple(5, 5, subpel_avg_variance32x32_c),
make_tuple(5, 6, subpel_avg_variance32x64_c),
make_tuple(6, 5, subpel_avg_variance64x32_c),
make_tuple(6, 6, subpel_avg_variance64x64_c)));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
#if HAVE_MMX
const vp9_variance_fn_t variance4x4_mmx = vp9_variance4x4_mmx;
const vp9_variance_fn_t variance8x8_mmx = vp9_variance8x8_mmx;
const vp9_variance_fn_t variance8x16_mmx = vp9_variance8x16_mmx;
const vp9_variance_fn_t variance16x8_mmx = vp9_variance16x8_mmx;
const vp9_variance_fn_t variance16x16_mmx = vp9_variance16x16_mmx;
INSTANTIATE_TEST_CASE_P(
MMX, VP9VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_mmx),
make_tuple(3, 3, variance8x8_mmx),
make_tuple(3, 4, variance8x16_mmx),
make_tuple(4, 3, variance16x8_mmx),
make_tuple(4, 4, variance16x16_mmx)));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
#endif
#if HAVE_SSE2
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
const vp9_variance_fn_t variance4x4_sse2 = vp9_variance4x4_sse2;
const vp9_variance_fn_t variance4x8_sse2 = vp9_variance4x8_sse2;
const vp9_variance_fn_t variance8x4_sse2 = vp9_variance8x4_sse2;
const vp9_variance_fn_t variance8x8_sse2 = vp9_variance8x8_sse2;
const vp9_variance_fn_t variance8x16_sse2 = vp9_variance8x16_sse2;
const vp9_variance_fn_t variance16x8_sse2 = vp9_variance16x8_sse2;
const vp9_variance_fn_t variance16x16_sse2 = vp9_variance16x16_sse2;
const vp9_variance_fn_t variance16x32_sse2 = vp9_variance16x32_sse2;
const vp9_variance_fn_t variance32x16_sse2 = vp9_variance32x16_sse2;
const vp9_variance_fn_t variance32x32_sse2 = vp9_variance32x32_sse2;
const vp9_variance_fn_t variance32x64_sse2 = vp9_variance32x64_sse2;
const vp9_variance_fn_t variance64x32_sse2 = vp9_variance64x32_sse2;
const vp9_variance_fn_t variance64x64_sse2 = vp9_variance64x64_sse2;
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
INSTANTIATE_TEST_CASE_P(
SSE2, VP9VarianceTest,
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
::testing::Values(make_tuple(2, 2, variance4x4_sse2),
make_tuple(2, 3, variance4x8_sse2),
make_tuple(3, 2, variance8x4_sse2),
make_tuple(3, 3, variance8x8_sse2),
make_tuple(3, 4, variance8x16_sse2),
make_tuple(4, 3, variance16x8_sse2),
make_tuple(4, 4, variance16x16_sse2),
make_tuple(4, 5, variance16x32_sse2),
make_tuple(5, 4, variance32x16_sse2),
make_tuple(5, 5, variance32x32_sse2),
make_tuple(5, 6, variance32x64_sse2),
make_tuple(6, 5, variance64x32_sse2),
make_tuple(6, 6, variance64x64_sse2)));
const vp9_subpixvariance_fn_t subpel_variance4x4_sse =
vp9_sub_pixel_variance4x4_sse;
const vp9_subpixvariance_fn_t subpel_variance4x8_sse =
vp9_sub_pixel_variance4x8_sse;
const vp9_subpixvariance_fn_t subpel_variance8x4_sse2 =
vp9_sub_pixel_variance8x4_sse2;
const vp9_subpixvariance_fn_t subpel_variance8x8_sse2 =
vp9_sub_pixel_variance8x8_sse2;
const vp9_subpixvariance_fn_t subpel_variance8x16_sse2 =
vp9_sub_pixel_variance8x16_sse2;
const vp9_subpixvariance_fn_t subpel_variance16x8_sse2 =
vp9_sub_pixel_variance16x8_sse2;
const vp9_subpixvariance_fn_t subpel_variance16x16_sse2 =
vp9_sub_pixel_variance16x16_sse2;
const vp9_subpixvariance_fn_t subpel_variance16x32_sse2 =
vp9_sub_pixel_variance16x32_sse2;
const vp9_subpixvariance_fn_t subpel_variance32x16_sse2 =
vp9_sub_pixel_variance32x16_sse2;
const vp9_subpixvariance_fn_t subpel_variance32x32_sse2 =
vp9_sub_pixel_variance32x32_sse2;
const vp9_subpixvariance_fn_t subpel_variance32x64_sse2 =
vp9_sub_pixel_variance32x64_sse2;
const vp9_subpixvariance_fn_t subpel_variance64x32_sse2 =
vp9_sub_pixel_variance64x32_sse2;
const vp9_subpixvariance_fn_t subpel_variance64x64_sse2 =
vp9_sub_pixel_variance64x64_sse2;
INSTANTIATE_TEST_CASE_P(
SSE2, VP9SubpelVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_variance4x4_sse),
make_tuple(2, 3, subpel_variance4x8_sse),
make_tuple(3, 2, subpel_variance8x4_sse2),
make_tuple(3, 3, subpel_variance8x8_sse2),
make_tuple(3, 4, subpel_variance8x16_sse2),
make_tuple(4, 3, subpel_variance16x8_sse2),
make_tuple(4, 4, subpel_variance16x16_sse2),
make_tuple(4, 5, subpel_variance16x32_sse2),
make_tuple(5, 4, subpel_variance32x16_sse2),
make_tuple(5, 5, subpel_variance32x32_sse2),
make_tuple(5, 6, subpel_variance32x64_sse2),
make_tuple(6, 5, subpel_variance64x32_sse2),
make_tuple(6, 6, subpel_variance64x64_sse2)));
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x4_sse =
vp9_sub_pixel_avg_variance4x4_sse;
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x8_sse =
vp9_sub_pixel_avg_variance4x8_sse;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x4_sse2 =
vp9_sub_pixel_avg_variance8x4_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x8_sse2 =
vp9_sub_pixel_avg_variance8x8_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x16_sse2 =
vp9_sub_pixel_avg_variance8x16_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x8_sse2 =
vp9_sub_pixel_avg_variance16x8_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x16_sse2 =
vp9_sub_pixel_avg_variance16x16_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x32_sse2 =
vp9_sub_pixel_avg_variance16x32_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x16_sse2 =
vp9_sub_pixel_avg_variance32x16_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x32_sse2 =
vp9_sub_pixel_avg_variance32x32_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x64_sse2 =
vp9_sub_pixel_avg_variance32x64_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x32_sse2 =
vp9_sub_pixel_avg_variance64x32_sse2;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x64_sse2 =
vp9_sub_pixel_avg_variance64x64_sse2;
INSTANTIATE_TEST_CASE_P(
SSE2, VP9SubpelAvgVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_avg_variance4x4_sse),
make_tuple(2, 3, subpel_avg_variance4x8_sse),
make_tuple(3, 2, subpel_avg_variance8x4_sse2),
make_tuple(3, 3, subpel_avg_variance8x8_sse2),
make_tuple(3, 4, subpel_avg_variance8x16_sse2),
make_tuple(4, 3, subpel_avg_variance16x8_sse2),
make_tuple(4, 4, subpel_avg_variance16x16_sse2),
make_tuple(4, 5, subpel_avg_variance16x32_sse2),
make_tuple(5, 4, subpel_avg_variance32x16_sse2),
make_tuple(5, 5, subpel_avg_variance32x32_sse2),
make_tuple(5, 6, subpel_avg_variance32x64_sse2),
make_tuple(6, 5, subpel_avg_variance64x32_sse2),
make_tuple(6, 6, subpel_avg_variance64x64_sse2)));
Implement sse2 and ssse3 versions for all sub_pixel_variance sizes. Overall speedup around 5% (bus @ 1500kbps first 50 frames 4min10 -> 3min58). Specific changes to timings for each function compared to original assembly-optimized versions (or just new version timings if no previous assembly-optimized version was available): sse2 4x4: 99 -> 82 cycles sse2 4x8: 128 cycles sse2 8x4: 121 cycles sse2 8x8: 149 -> 129 cycles sse2 8x16: 235 -> 245 cycles (?) sse2 16x8: 269 -> 203 cycles sse2 16x16: 441 -> 349 cycles sse2 16x32: 641 cycles sse2 32x16: 643 cycles sse2 32x32: 1733 -> 1154 cycles sse2 32x64: 2247 cycles sse2 64x32: 2323 cycles sse2 64x64: 6984 -> 4442 cycles ssse3 4x4: 100 cycles (?) ssse3 4x8: 103 cycles ssse3 8x4: 71 cycles ssse3 8x8: 147 cycles ssse3 8x16: 158 cycles ssse3 16x8: 188 -> 162 cycles ssse3 16x16: 316 -> 273 cycles ssse3 16x32: 535 cycles ssse3 32x16: 564 cycles ssse3 32x32: 973 cycles ssse3 32x64: 1930 cycles ssse3 64x32: 1922 cycles ssse3 64x64: 3760 cycles Change-Id: I81ff6fe51daf35a40d19785167004664d7e0c59d
2013-06-20 18:34:25 +02:00
#endif
#if HAVE_SSSE3
const vp9_subpixvariance_fn_t subpel_variance4x4_ssse3 =
vp9_sub_pixel_variance4x4_ssse3;
const vp9_subpixvariance_fn_t subpel_variance4x8_ssse3 =
vp9_sub_pixel_variance4x8_ssse3;
const vp9_subpixvariance_fn_t subpel_variance8x4_ssse3 =
vp9_sub_pixel_variance8x4_ssse3;
const vp9_subpixvariance_fn_t subpel_variance8x8_ssse3 =
vp9_sub_pixel_variance8x8_ssse3;
const vp9_subpixvariance_fn_t subpel_variance8x16_ssse3 =
vp9_sub_pixel_variance8x16_ssse3;
const vp9_subpixvariance_fn_t subpel_variance16x8_ssse3 =
vp9_sub_pixel_variance16x8_ssse3;
const vp9_subpixvariance_fn_t subpel_variance16x16_ssse3 =
vp9_sub_pixel_variance16x16_ssse3;
const vp9_subpixvariance_fn_t subpel_variance16x32_ssse3 =
vp9_sub_pixel_variance16x32_ssse3;
const vp9_subpixvariance_fn_t subpel_variance32x16_ssse3 =
vp9_sub_pixel_variance32x16_ssse3;
const vp9_subpixvariance_fn_t subpel_variance32x32_ssse3 =
vp9_sub_pixel_variance32x32_ssse3;
const vp9_subpixvariance_fn_t subpel_variance32x64_ssse3 =
vp9_sub_pixel_variance32x64_ssse3;
const vp9_subpixvariance_fn_t subpel_variance64x32_ssse3 =
vp9_sub_pixel_variance64x32_ssse3;
const vp9_subpixvariance_fn_t subpel_variance64x64_ssse3 =
vp9_sub_pixel_variance64x64_ssse3;
INSTANTIATE_TEST_CASE_P(
SSSE3, VP9SubpelVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_variance4x4_ssse3),
make_tuple(2, 3, subpel_variance4x8_ssse3),
make_tuple(3, 2, subpel_variance8x4_ssse3),
make_tuple(3, 3, subpel_variance8x8_ssse3),
make_tuple(3, 4, subpel_variance8x16_ssse3),
make_tuple(4, 3, subpel_variance16x8_ssse3),
make_tuple(4, 4, subpel_variance16x16_ssse3),
make_tuple(4, 5, subpel_variance16x32_ssse3),
make_tuple(5, 4, subpel_variance32x16_ssse3),
make_tuple(5, 5, subpel_variance32x32_ssse3),
make_tuple(5, 6, subpel_variance32x64_ssse3),
make_tuple(6, 5, subpel_variance64x32_ssse3),
make_tuple(6, 6, subpel_variance64x64_ssse3)));
SSE2/SSSE3 optimizations and unit test for sub_pixel_avg_variance(). Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to 3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't perfectly interleaved, and can probably be improved further in the future. I've marked this with a few TODOs/FIXMEs in the code. Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
2013-06-21 00:59:48 +02:00
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x4_ssse3 =
vp9_sub_pixel_avg_variance4x4_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance4x8_ssse3 =
vp9_sub_pixel_avg_variance4x8_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x4_ssse3 =
vp9_sub_pixel_avg_variance8x4_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x8_ssse3 =
vp9_sub_pixel_avg_variance8x8_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance8x16_ssse3 =
vp9_sub_pixel_avg_variance8x16_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x8_ssse3 =
vp9_sub_pixel_avg_variance16x8_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x16_ssse3 =
vp9_sub_pixel_avg_variance16x16_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance16x32_ssse3 =
vp9_sub_pixel_avg_variance16x32_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x16_ssse3 =
vp9_sub_pixel_avg_variance32x16_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x32_ssse3 =
vp9_sub_pixel_avg_variance32x32_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance32x64_ssse3 =
vp9_sub_pixel_avg_variance32x64_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x32_ssse3 =
vp9_sub_pixel_avg_variance64x32_ssse3;
const vp9_subp_avg_variance_fn_t subpel_avg_variance64x64_ssse3 =
vp9_sub_pixel_avg_variance64x64_ssse3;
INSTANTIATE_TEST_CASE_P(
SSSE3, VP9SubpelAvgVarianceTest,
::testing::Values(make_tuple(2, 2, subpel_avg_variance4x4_ssse3),
make_tuple(2, 3, subpel_avg_variance4x8_ssse3),
make_tuple(3, 2, subpel_avg_variance8x4_ssse3),
make_tuple(3, 3, subpel_avg_variance8x8_ssse3),
make_tuple(3, 4, subpel_avg_variance8x16_ssse3),
make_tuple(4, 3, subpel_avg_variance16x8_ssse3),
make_tuple(4, 4, subpel_avg_variance16x16_ssse3),
make_tuple(4, 5, subpel_avg_variance16x32_ssse3),
make_tuple(5, 4, subpel_avg_variance32x16_ssse3),
make_tuple(5, 5, subpel_avg_variance32x32_ssse3),
make_tuple(5, 6, subpel_avg_variance32x64_ssse3),
make_tuple(6, 5, subpel_avg_variance64x32_ssse3),
make_tuple(6, 6, subpel_avg_variance64x64_ssse3)));
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
#endif
add vp8 variance test Change-Id: I4e94ee2c4e2360d6a11a454c323f2899c1bb6f72
2013-02-23 01:23:36 +01:00
#endif // CONFIG_VP9_ENCODER
} // namespace vp9
Fix variance (signed integer) overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we force the multiplication to be unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). This change is based on: 1698234 Missed some variance casts fea3556 Fix variance overflow Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
2012-11-07 01:58:11 +01:00
} // namespace
Reference in New Issue Copy Permalink