Commit 0d3c3d3c authored by Deb Mukherjee's avatar Deb Mukherjee
Browse files

Adds high bitdepth convolve, interpred & scaling

Change-Id: Ie51c352a6b250547207cbc1ebba833a01ed053e3
parent d3a7e677
......@@ -21,6 +21,9 @@
#include "vpx_ports/mem.h"
namespace {
static const int kMaxDimension = 64;
typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int filter_x_stride,
......@@ -30,9 +33,10 @@ typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
struct ConvolveFunctions {
ConvolveFunctions(ConvolveFunc h8, ConvolveFunc h8_avg,
ConvolveFunc v8, ConvolveFunc v8_avg,
ConvolveFunc hv8, ConvolveFunc hv8_avg)
ConvolveFunc hv8, ConvolveFunc hv8_avg,
int bd)
: h8_(h8), v8_(v8), hv8_(hv8), h8_avg_(h8_avg), v8_avg_(v8_avg),
hv8_avg_(hv8_avg) {}
hv8_avg_(hv8_avg), use_high_bd_(bd) {}
ConvolveFunc h8_;
ConvolveFunc v8_;
......@@ -40,6 +44,7 @@ struct ConvolveFunctions {
ConvolveFunc h8_avg_;
ConvolveFunc v8_avg_;
ConvolveFunc hv8_avg_;
int use_high_bd_; // 0 if high bitdepth not used, else the actual bit depth.
};
typedef std::tr1::tuple<int, int, const ConvolveFunctions *> ConvolveParam;
......@@ -66,6 +71,119 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
// This buffer is allocated to be big enough for the largest block type we
// support.
const int kInterp_Extend = 4;
const unsigned int intermediate_height =
(kInterp_Extend - 1) + output_height + kInterp_Extend;
unsigned int i, j;
// Size of intermediate_buffer is max_intermediate_height * filter_max_width,
// where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
// + kInterp_Extend
// = 3 + 16 + 4
// = 23
// and filter_max_width = 16
//
uint8_t intermediate_buffer[71 * kMaxDimension];
const int intermediate_next_stride = 1 - intermediate_height * output_width;
// Horizontal pass (src -> transposed intermediate).
uint8_t *output_ptr = intermediate_buffer;
const int src_next_row_stride = src_stride - output_width;
src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
for (i = 0; i < intermediate_height; ++i) {
for (j = 0; j < output_width; ++j) {
// Apply filter...
const int temp = (src_ptr[0] * HFilter[0]) +
(src_ptr[1] * HFilter[1]) +
(src_ptr[2] * HFilter[2]) +
(src_ptr[3] * HFilter[3]) +
(src_ptr[4] * HFilter[4]) +
(src_ptr[5] * HFilter[5]) +
(src_ptr[6] * HFilter[6]) +
(src_ptr[7] * HFilter[7]) +
(VP9_FILTER_WEIGHT >> 1); // Rounding
// Normalize back to 0-255...
*output_ptr = clip_pixel(temp >> VP9_FILTER_SHIFT);
++src_ptr;
output_ptr += intermediate_height;
}
src_ptr += src_next_row_stride;
output_ptr += intermediate_next_stride;
}
// Vertical pass (transposed intermediate -> dst).
src_ptr = intermediate_buffer;
const int dst_next_row_stride = dst_stride - output_width;
for (i = 0; i < output_height; ++i) {
for (j = 0; j < output_width; ++j) {
// Apply filter...
const int temp = (src_ptr[0] * VFilter[0]) +
(src_ptr[1] * VFilter[1]) +
(src_ptr[2] * VFilter[2]) +
(src_ptr[3] * VFilter[3]) +
(src_ptr[4] * VFilter[4]) +
(src_ptr[5] * VFilter[5]) +
(src_ptr[6] * VFilter[6]) +
(src_ptr[7] * VFilter[7]) +
(VP9_FILTER_WEIGHT >> 1); // Rounding
// Normalize back to 0-255...
*dst_ptr++ = clip_pixel(temp >> VP9_FILTER_SHIFT);
src_ptr += intermediate_height;
}
src_ptr += intermediate_next_stride;
dst_ptr += dst_next_row_stride;
}
}
void block2d_average_c(uint8_t *src,
unsigned int src_stride,
uint8_t *output_ptr,
unsigned int output_stride,
unsigned int output_width,
unsigned int output_height) {
unsigned int i, j;
for (i = 0; i < output_height; ++i) {
for (j = 0; j < output_width; ++j) {
output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
}
output_ptr += output_stride;
}
}
void filter_average_block2d_8_c(const uint8_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint8_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height) {
uint8_t tmp[kMaxDimension * kMaxDimension];
assert(output_width <= kMaxDimension);
assert(output_height <= kMaxDimension);
filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
output_width, output_height);
block2d_average_c(tmp, 64, dst_ptr, dst_stride,
output_width, output_height);
}
#if CONFIG_VP9_HIGHBITDEPTH
void high_filter_block2d_8_c(const uint16_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint16_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height,
int bd) {
// Between passes, we use an intermediate buffer whose height is extended to
// have enough horizontally filtered values as input for the vertical pass.
// This buffer is allocated to be big enough for the largest block type we
// support.
const int kInterp_Extend = 4;
const unsigned int intermediate_height =
(kInterp_Extend - 1) + output_height + kInterp_Extend;
......@@ -76,12 +194,12 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
* = 23
* and filter_max_width = 16
*/
uint8_t intermediate_buffer[71 * 64];
uint16_t intermediate_buffer[71 * kMaxDimension];
const int intermediate_next_stride = 1 - intermediate_height * output_width;
// Horizontal pass (src -> transposed intermediate).
{
uint8_t *output_ptr = intermediate_buffer;
uint16_t *output_ptr = intermediate_buffer;
const int src_next_row_stride = src_stride - output_width;
unsigned int i, j;
src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
......@@ -99,7 +217,7 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
(VP9_FILTER_WEIGHT >> 1); // Rounding
// Normalize back to 0-255...
*output_ptr = clip_pixel(temp >> VP9_FILTER_SHIFT);
*output_ptr = clip_pixel_high(temp >> VP9_FILTER_SHIFT, bd);
++src_ptr;
output_ptr += intermediate_height;
}
......@@ -110,7 +228,7 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
// Vertical pass (transposed intermediate -> dst).
{
uint8_t *src_ptr = intermediate_buffer;
uint16_t *src_ptr = intermediate_buffer;
const int dst_next_row_stride = dst_stride - output_width;
unsigned int i, j;
for (i = 0; i < output_height; ++i) {
......@@ -127,7 +245,7 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
(VP9_FILTER_WEIGHT >> 1); // Rounding
// Normalize back to 0-255...
*dst_ptr++ = clip_pixel(temp >> VP9_FILTER_SHIFT);
*dst_ptr++ = clip_pixel_high(temp >> VP9_FILTER_SHIFT, bd);
src_ptr += intermediate_height;
}
src_ptr += intermediate_next_stride;
......@@ -136,12 +254,13 @@ void filter_block2d_8_c(const uint8_t *src_ptr,
}
}
void block2d_average_c(uint8_t *src,
unsigned int src_stride,
uint8_t *output_ptr,
unsigned int output_stride,
unsigned int output_width,
unsigned int output_height) {
void high_block2d_average_c(uint16_t *src,
unsigned int src_stride,
uint16_t *output_ptr,
unsigned int output_stride,
unsigned int output_width,
unsigned int output_height,
int bd) {
unsigned int i, j;
for (i = 0; i < output_height; ++i) {
for (j = 0; j < output_width; ++j) {
......@@ -151,23 +270,25 @@ void block2d_average_c(uint8_t *src,
}
}
void filter_average_block2d_8_c(const uint8_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint8_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height) {
uint8_t tmp[64 * 64];
assert(output_width <= 64);
assert(output_height <= 64);
filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
output_width, output_height);
block2d_average_c(tmp, 64, dst_ptr, dst_stride,
output_width, output_height);
void high_filter_average_block2d_8_c(const uint16_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint16_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height,
int bd) {
uint16_t tmp[kMaxDimension * kMaxDimension];
assert(output_width <= kMaxDimension);
assert(output_height <= kMaxDimension);
high_filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
output_width, output_height, bd);
high_block2d_average_c(tmp, 64, dst_ptr, dst_stride,
output_width, output_height, bd);
}
#endif // CONFIG_VP9_HIGHBITDEPTH
class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
public:
......@@ -177,6 +298,13 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
vpx_memalign(kDataAlignment, kInputBufferSize + 1)) + 1;
output_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kOutputBufferSize));
#if CONFIG_VP9_HIGHBITDEPTH
input16_ = reinterpret_cast<uint16_t*>(
vpx_memalign(kDataAlignment,
(kInputBufferSize + 1) * sizeof(uint16_t))) + 1;
output16_ = reinterpret_cast<uint16_t*>(
vpx_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
#endif
}
static void TearDownTestCase() {
......@@ -184,6 +312,12 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
input_ = NULL;
vpx_free(output_);
output_ = NULL;
#if CONFIG_VP9_HIGHBITDEPTH
vpx_free(input16_ - 1);
input16_ = NULL;
vpx_free(output16_);
output16_ = NULL;
#endif
}
protected:
......@@ -191,7 +325,6 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
static const int kOuterBlockSize = 256;
static const int kInputStride = kOuterBlockSize;
static const int kOutputStride = kOuterBlockSize;
static const int kMaxDimension = 64;
static const int kInputBufferSize = kOuterBlockSize * kOuterBlockSize;
static const int kOutputBufferSize = kOuterBlockSize * kOuterBlockSize;
......@@ -212,6 +345,12 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
virtual void SetUp() {
UUT_ = GET_PARAM(2);
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ != 0)
mask_ = (1 << UUT_->use_high_bd_) - 1;
else
mask_ = 255;
#endif
/* Set up guard blocks for an inner block centered in the outer block */
for (int i = 0; i < kOutputBufferSize; ++i) {
if (IsIndexInBorder(i))
......@@ -222,15 +361,25 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
::libvpx_test::ACMRandom prng;
for (int i = 0; i < kInputBufferSize; ++i) {
if (i & 1)
if (i & 1) {
input_[i] = 255;
else
#if CONFIG_VP9_HIGHBITDEPTH
input16_[i] = mask_;
#endif
} else {
input_[i] = prng.Rand8Extremes();
#if CONFIG_VP9_HIGHBITDEPTH
input16_[i] = prng.Rand16() & mask_;
#endif
}
}
}
void SetConstantInput(int value) {
memset(input_, value, kInputBufferSize);
#if CONFIG_VP9_HIGHBITDEPTH
vpx_memset16(input16_, value, kInputBufferSize);
#endif
}
void CheckGuardBlocks() {
......@@ -240,20 +389,123 @@ class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
}
}
uint8_t* input() const {
uint8_t *input() const {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
} else {
return CONVERT_TO_BYTEPTR(input16_ + BorderTop() * kOuterBlockSize +
BorderLeft());
}
#else
return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
#endif
}
uint8_t* output() const {
uint8_t *output() const {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
} else {
return CONVERT_TO_BYTEPTR(output16_ + BorderTop() * kOuterBlockSize +
BorderLeft());
}
#else
return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
#endif
}
uint16_t lookup(uint8_t *list, int index) const {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
return list[index];
} else {
return CONVERT_TO_SHORTPTR(list)[index];
}
#else
return list[index];
#endif
}
void assign_val(uint8_t *list, int index, uint16_t val) const {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
list[index] = (uint8_t) val;
} else {
CONVERT_TO_SHORTPTR(list)[index] = val;
}
#else
list[index] = (uint8_t) val;
#endif
}
void wrapper_filter_average_block2d_8_c(const uint8_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint8_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height) {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
dst_ptr, dst_stride, output_width,
output_height);
} else {
high_filter_average_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
HFilter, VFilter,
CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
output_width, output_height,
UUT_->use_high_bd_);
}
#else
filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
dst_ptr, dst_stride, output_width,
output_height);
#endif
}
void wrapper_filter_block2d_8_c(const uint8_t *src_ptr,
const unsigned int src_stride,
const int16_t *HFilter,
const int16_t *VFilter,
uint8_t *dst_ptr,
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height) {
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0) {
filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
dst_ptr, dst_stride, output_width, output_height);
} else {
high_filter_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
HFilter, VFilter,
CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
output_width, output_height, UUT_->use_high_bd_);
}
#else
filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
dst_ptr, dst_stride, output_width, output_height);
#endif
}
const ConvolveFunctions* UUT_;
static uint8_t* input_;
static uint8_t* output_;
#if CONFIG_VP9_HIGHBITDEPTH
static uint16_t* input16_;
static uint16_t* output16_;
int mask_;
#endif
};
uint8_t* ConvolveTest::input_ = NULL;
uint8_t* ConvolveTest::output_ = NULL;
#if CONFIG_VP9_HIGHBITDEPTH
uint16_t* ConvolveTest::input16_ = NULL;
uint16_t* ConvolveTest::output16_ = NULL;
#endif
TEST_P(ConvolveTest, GuardBlocks) {
CheckGuardBlocks();
......@@ -272,7 +524,8 @@ TEST_P(ConvolveTest, CopyHoriz) {
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
ASSERT_EQ(lookup(out, y * kOutputStride + x),
lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
......@@ -289,7 +542,8 @@ TEST_P(ConvolveTest, CopyVert) {
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
ASSERT_EQ(lookup(out, y * kOutputStride + x),
lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
......@@ -306,7 +560,8 @@ TEST_P(ConvolveTest, Copy2D) {
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
ASSERT_EQ(lookup(out, y * kOutputStride + x),
lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
......@@ -339,8 +594,18 @@ const int16_t kInvalidFilter[8] = { 0 };
TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
uint8_t* const in = input();
uint8_t* const out = output();
#if CONFIG_VP9_HIGHBITDEPTH
uint8_t ref8[kOutputStride * kMaxDimension];
uint16_t ref16[kOutputStride * kMaxDimension];
uint8_t* ref;
if (UUT_->use_high_bd_ == 0) {
ref = ref8;
} else {
ref = CONVERT_TO_BYTEPTR(ref16);
}
#else
uint8_t ref[kOutputStride * kMaxDimension];
#endif
for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
const InterpKernel *filters =
......@@ -350,10 +615,10 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
filter_block2d_8_c(in, kInputStride,
filters[filter_x], filters[filter_y],
ref, kOutputStride,
Width(), Height());
wrapper_filter_block2d_8_c(in, kInputStride,
filters[filter_x], filters[filter_y],
ref, kOutputStride,
Width(), Height());
if (filters == eighttap_smooth || (filter_x && filter_y))
ASM_REGISTER_STATE_CHECK(
......@@ -375,7 +640,8 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
ASSERT_EQ(ref[y * kOutputStride + x], out[y * kOutputStride + x])
ASSERT_EQ(lookup(ref, y * kOutputStride + x),
lookup(out, y * kOutputStride + x))
<< "mismatch at (" << x << "," << y << "), "
<< "filters (" << filter_bank << ","
<< filter_x << "," << filter_y << ")";
......@@ -387,16 +653,36 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
uint8_t* const in = input();
uint8_t* const out = output();
#if CONFIG_VP9_HIGHBITDEPTH
uint8_t ref8[kOutputStride * kMaxDimension];
uint16_t ref16[kOutputStride * kMaxDimension];
uint8_t* ref;
if (UUT_->use_high_bd_ == 0) {
ref = ref8;
} else {
ref = CONVERT_TO_BYTEPTR(ref16);
}
#else
uint8_t ref[kOutputStride * kMaxDimension];
#endif
// Populate ref and out with some random data
::libvpx_test::ACMRandom prng;
for (int y = 0; y < Height(); ++y) {
for (int x = 0; x < Width(); ++x) {
const uint8_t r = prng.Rand8Extremes();
uint16_t r;
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0 || UUT_->use_high_bd_ == 8) {
r = prng.Rand8Extremes();
} else {
r = prng.Rand16() & mask_;
}
#else
r = prng.Rand8Extremes();
#endif
out[y * kOutputStride + x] = r;
ref[y * kOutputStride + x] = r;
assign_val(out, y * kOutputStride + x, r);
assign_val(ref, y * kOutputStride + x, r);
}
}
......@@ -408,10 +694,10 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
filter_average_block2d_8_c(in, kInputStride,
filters[filter_x], filters[filter_y],
ref, kOutputStride,
Width(), Height());
wrapper_filter_average_block2d_8_c(in, kInputStride,
filters[filter_x], filters[filter_y],
ref, kOutputStride,
Width(), Height());
if (filters == eighttap_smooth || (filter_x && filter_y))
ASM_REGISTER_STATE_CHECK(
......@@ -433,7 +719,8 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
ASSERT_EQ(ref[y * kOutputStride + x], out[y * kOutputStride + x])
ASSERT_EQ(lookup(ref, y * kOutputStride + x),
lookup(out, y * kOutputStride + x))
<< "mismatch at (" << x << "," << y << "), "
<< "filters (" << filter_bank << ","
<< filter_x << "," << filter_y << ")";
......@@ -442,6 +729,103 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
}
}
TEST_P(ConvolveTest, FilterExtremes) {
uint8_t *const in = input();
uint8_t *const out = output();
#if CONFIG_VP9_HIGHBITDEPTH
uint8_t ref8[kOutputStride * kMaxDimension];
uint16_t ref16[kOutputStride * kMaxDimension];
uint8_t *ref;
if (UUT_->use_high_bd_ == 0) {
ref = ref8;
} else {
ref = CONVERT_TO_BYTEPTR(ref16);
}
#else
uint8_t ref[kOutputStride * kMaxDimension];
#endif
// Populate ref and out with some random data
::libvpx_test::ACMRandom prng;
for (int y = 0; y < Height(); ++y) {
for (int x = 0; x < Width(); ++x) {
uint16_t r;
#if CONFIG_VP9_HIGHBITDEPTH
if (UUT_->use_high_bd_ == 0 || UUT_->use_high_bd_ == 8) {
r = prng.Rand8Extremes();