diff --git a/test/sad_test.cc b/test/sad_test.cc
index e0c5cfefa01398648a695bc3922c86977db07219..165e2c8f0c58182513ba2a63105f70848c5ad63f 100644
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -37,14 +37,22 @@ typedef unsigned int (*sad_m_by_n_fn_t)(const unsigned char *source_ptr,
int reference_stride,
unsigned int max_sad);
+typedef void (*sad_n_by_n_by_4_fn_t)(const uint8_t *src_ptr,
+ int src_stride,
+ const unsigned char * const ref_ptr[],
+ int ref_stride,
+ unsigned int *sad_array);
+
using libvpx_test::ACMRandom;
namespace {
-class SADTest : public PARAMS(int, int, sad_m_by_n_fn_t) {
+class SADTestBase : public ::testing::Test {
public:
+ SADTestBase(int width, int height) : width_(width), height_(height) {}
+
static void SetUpTestCase() {
source_data_ = reinterpret_cast<uint8_t*>(
- vpx_memalign(kDataAlignment, kDataBufferSize));
+ vpx_memalign(kDataAlignment, kDataBlockSize));
reference_data_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kDataBufferSize));
}
@@ -59,35 +67,29 @@ class SADTest : public PARAMS(int, int, sad_m_by_n_fn_t) {
protected:
// Handle blocks up to 4 blocks 64x64 with stride up to 128
static const int kDataAlignment = 16;
- static const int kDataBufferSize = 4 * 64 * 128;
+ static const int kDataBlockSize = 64 * 128;
+ static const int kDataBufferSize = 4 * kDataBlockSize;
virtual void SetUp() {
- sad_fn_ = GET_PARAM(2);
- height_ = GET_PARAM(1);
- width_ = GET_PARAM(0);
source_stride_ = (width_ + 31) & ~31;
reference_stride_ = width_ * 2;
rnd_.Reset(ACMRandom::DeterministicSeed());
}
- sad_m_by_n_fn_t sad_fn_;
- virtual unsigned int SAD(unsigned int max_sad) {
- unsigned int ret;
- REGISTER_STATE_CHECK(ret = sad_fn_(source_data_, source_stride_,
- reference_data_, reference_stride_,
- max_sad));
- return ret;
+ virtual uint8_t* GetReference(int block_idx) {
+ return reference_data_ + block_idx * kDataBlockSize;
}
// Sum of Absolute Differences. Given two blocks, calculate the absolute
// difference between two pixels in the same relative location; accumulate.
- unsigned int ReferenceSAD(unsigned int max_sad) {
+ unsigned int ReferenceSAD(unsigned int max_sad, int block_idx = 0) {
unsigned int sad = 0;
+ const uint8_t* const reference = GetReference(block_idx);
for (int h = 0; h < height_; ++h) {
for (int w = 0; w < width_; ++w) {
sad += abs(source_data_[h * source_stride_ + w]
- - reference_data_[h * reference_stride_ + w]);
+ - reference[h * reference_stride_ + w]);
}
if (sad > max_sad) {
break;
@@ -112,6 +114,32 @@ class SADTest : public PARAMS(int, int, sad_m_by_n_fn_t) {
}
}
+ int width_, height_;
+ static uint8_t* source_data_;
+ int source_stride_;
+ static uint8_t* reference_data_;
+ int reference_stride_;
+
+ ACMRandom rnd_;
+};
+
+class SADTest : public SADTestBase,
+ public ::testing::WithParamInterface<
+ std::tr1::tuple<int, int, sad_m_by_n_fn_t> > {
+ public:
+ SADTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+ unsigned int SAD(unsigned int max_sad, int block_idx = 0) {
+ unsigned int ret;
+ const uint8_t* const reference = GetReference(block_idx);
+
+ REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+ reference, reference_stride_,
+ max_sad));
+ return ret;
+ }
+
void CheckSad(unsigned int max_sad) {
unsigned int reference_sad, exp_sad;
@@ -125,18 +153,38 @@ class SADTest : public PARAMS(int, int, sad_m_by_n_fn_t) {
ASSERT_GE(exp_sad, reference_sad);
}
}
+};
- int height_, width_;
- static uint8_t* source_data_;
- int source_stride_;
- static uint8_t* reference_data_;
- int reference_stride_;
+class SADx4Test : public SADTestBase,
+ public ::testing::WithParamInterface<
+ std::tr1::tuple<int, int, sad_n_by_n_by_4_fn_t> > {
+ public:
+ SADx4Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
- ACMRandom rnd_;
+ protected:
+ void SADs(unsigned int *results) {
+ const uint8_t* refs[] = {GetReference(0), GetReference(1),
+ GetReference(2), GetReference(3)};
+
+ REGISTER_STATE_CHECK(GET_PARAM(2)(source_data_, source_stride_,
+ refs, reference_stride_,
+ results));
+ }
+
+ void CheckSADs() {
+ unsigned int reference_sad, exp_sad[4];
+
+ SADs(exp_sad);
+ for (int block = 0; block < 4; block++) {
+ reference_sad = ReferenceSAD(UINT_MAX, block);
+
+ EXPECT_EQ(exp_sad[block], reference_sad) << "block " << block;
+ }
+ }
};
-uint8_t* SADTest::source_data_ = NULL;
-uint8_t* SADTest::reference_data_ = NULL;
+uint8_t* SADTestBase::source_data_ = NULL;
+uint8_t* SADTestBase::reference_data_ = NULL;
TEST_P(SADTest, MaxRef) {
FillConstant(source_data_, source_stride_, 0);
@@ -144,12 +192,30 @@ TEST_P(SADTest, MaxRef) {
CheckSad(UINT_MAX);
}
+TEST_P(SADx4Test, MaxRef) {
+ FillConstant(source_data_, source_stride_, 0);
+ FillConstant(GetReference(0), reference_stride_, 255);
+ FillConstant(GetReference(1), reference_stride_, 255);
+ FillConstant(GetReference(2), reference_stride_, 255);
+ FillConstant(GetReference(3), reference_stride_, 255);
+ CheckSADs();
+}
+
TEST_P(SADTest, MaxSrc) {
FillConstant(source_data_, source_stride_, 255);
FillConstant(reference_data_, reference_stride_, 0);
CheckSad(UINT_MAX);
}
+TEST_P(SADx4Test, MaxSrc) {
+ FillConstant(source_data_, source_stride_, 255);
+ FillConstant(GetReference(0), reference_stride_, 0);
+ FillConstant(GetReference(1), reference_stride_, 0);
+ FillConstant(GetReference(2), reference_stride_, 0);
+ FillConstant(GetReference(3), reference_stride_, 0);
+ CheckSADs();
+}
+
TEST_P(SADTest, ShortRef) {
int tmp_stride = reference_stride_;
reference_stride_ >>= 1;
@@ -159,6 +225,18 @@ TEST_P(SADTest, ShortRef) {
reference_stride_ = tmp_stride;
}
+TEST_P(SADx4Test, ShortRef) {
+ int tmp_stride = reference_stride_;
+ reference_stride_ >>= 1;
+ FillRandom(source_data_, source_stride_);
+ FillRandom(GetReference(0), reference_stride_);
+ FillRandom(GetReference(1), reference_stride_);
+ FillRandom(GetReference(2), reference_stride_);
+ FillRandom(GetReference(3), reference_stride_);
+ CheckSADs();
+ reference_stride_ = tmp_stride;
+}
+
TEST_P(SADTest, UnalignedRef) {
// The reference frame, but not the source frame, may be unaligned for
// certain types of searches.
@@ -170,6 +248,20 @@ TEST_P(SADTest, UnalignedRef) {
reference_stride_ = tmp_stride;
}
+TEST_P(SADx4Test, UnalignedRef) {
+ // The reference frame, but not the source frame, may be unaligned for
+ // certain types of searches.
+ int tmp_stride = reference_stride_;
+ reference_stride_ -= 1;
+ FillRandom(source_data_, source_stride_);
+ FillRandom(GetReference(0), reference_stride_);
+ FillRandom(GetReference(1), reference_stride_);
+ FillRandom(GetReference(2), reference_stride_);
+ FillRandom(GetReference(3), reference_stride_);
+ CheckSADs();
+ reference_stride_ = tmp_stride;
+}
+
TEST_P(SADTest, ShortSrc) {
int tmp_stride = source_stride_;
source_stride_ >>= 1;
@@ -179,6 +271,18 @@ TEST_P(SADTest, ShortSrc) {
source_stride_ = tmp_stride;
}
+TEST_P(SADx4Test, ShortSrc) {
+ int tmp_stride = source_stride_;
+ source_stride_ >>= 1;
+ FillRandom(source_data_, source_stride_);
+ FillRandom(GetReference(0), reference_stride_);
+ FillRandom(GetReference(1), reference_stride_);
+ FillRandom(GetReference(2), reference_stride_);
+ FillRandom(GetReference(3), reference_stride_);
+ CheckSADs();
+ source_stride_ = tmp_stride;
+}
+
TEST_P(SADTest, MaxSAD) {
// Verify that, when max_sad is set, the implementation does not return a
// value lower than the reference.
@@ -231,6 +335,20 @@ INSTANTIATE_TEST_CASE_P(C, SADTest, ::testing::Values(
#endif
));
+#if CONFIG_VP9_ENCODER
+const sad_n_by_n_by_4_fn_t sad_64x64x4d_c = vp9_sad64x64x4d_c;
+const sad_n_by_n_by_4_fn_t sad_32x32x4d_c = vp9_sad32x32x4d_c;
+const sad_n_by_n_by_4_fn_t sad_16x16x4d_c = vp9_sad16x16x4d_c;
+const sad_n_by_n_by_4_fn_t sad_8x8x4d_c = vp9_sad8x8x4d_c;
+const sad_n_by_n_by_4_fn_t sad_4x4x4d_c = vp9_sad4x4x4d_c;
+INSTANTIATE_TEST_CASE_P(C, SADx4Test, ::testing::Values(
+ make_tuple(64, 64, sad_64x64x4d_c),
+ make_tuple(32, 32, sad_32x32x4d_c),
+ make_tuple(16, 16, sad_16x16x4d_c),
+ make_tuple(8, 8, sad_8x8x4d_c),
+ make_tuple(4, 4, sad_4x4x4d_c)));
+#endif
+
// ARM tests
#if HAVE_MEDIA
const sad_m_by_n_fn_t sad_16x16_armv6 = vp8_sad16x16_armv6;
@@ -293,6 +411,10 @@ INSTANTIATE_TEST_CASE_P(MMX, SADTest, ::testing::Values(
const sad_m_by_n_fn_t sad_4x4_sse_vp9 = vp9_sad4x4_sse;
INSTANTIATE_TEST_CASE_P(SSE, SADTest, ::testing::Values(
make_tuple(4, 4, sad_4x4_sse_vp9)));
+
+const sad_n_by_n_by_4_fn_t sad_4x4x4d_sse = vp9_sad4x4x4d_sse;
+INSTANTIATE_TEST_CASE_P(SSE, SADx4Test, ::testing::Values(
+ make_tuple(4, 4, sad_4x4x4d_sse)));
#endif
#endif
@@ -330,6 +452,38 @@ INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::Values(
make_tuple(8, 8, sad_8x8_sse2_vp9)
#endif
));
+
+#if CONFIG_VP9_ENCODER
+const sad_n_by_n_by_4_fn_t sad_64x64x4d_sse2 = vp9_sad64x64x4d_sse2;
+const sad_n_by_n_by_4_fn_t sad_32x32x4d_sse2 = vp9_sad32x32x4d_sse2;
+const sad_n_by_n_by_4_fn_t sad_16x16x4d_sse2 = vp9_sad16x16x4d_sse2;
+const sad_n_by_n_by_4_fn_t sad_16x8x4d_sse2 = vp9_sad16x8x4d_sse2;
+const sad_n_by_n_by_4_fn_t sad_8x16x4d_sse2 = vp9_sad8x16x4d_sse2;
+const sad_n_by_n_by_4_fn_t sad_8x8x4d_sse2 = vp9_sad8x8x4d_sse2;
+INSTANTIATE_TEST_CASE_P(SSE2, SADx4Test, ::testing::Values(
+ make_tuple(64, 64, sad_64x64x4d_sse2),
+ make_tuple(32, 32, sad_32x32x4d_sse2),
+ make_tuple(16, 16, sad_16x16x4d_sse2),
+ make_tuple(16, 8, sad_16x8x4d_sse2),
+ make_tuple(8, 16, sad_8x16x4d_sse2),
+ make_tuple(8, 8, sad_8x8x4d_sse2)));
+#endif
+#endif
+
+#if HAVE_SSE3
+#if CONFIG_VP8_ENCODER
+const sad_n_by_n_by_4_fn_t sad_16x16x4d_sse3 = vp8_sad16x16x4d_sse3;
+const sad_n_by_n_by_4_fn_t sad_16x8x4d_sse3 = vp8_sad16x8x4d_sse3;
+const sad_n_by_n_by_4_fn_t sad_8x16x4d_sse3 = vp8_sad8x16x4d_sse3;
+const sad_n_by_n_by_4_fn_t sad_8x8x4d_sse3 = vp8_sad8x8x4d_sse3;
+const sad_n_by_n_by_4_fn_t sad_4x4x4d_sse3 = vp8_sad4x4x4d_sse3;
+INSTANTIATE_TEST_CASE_P(SSE3, SADx4Test, ::testing::Values(
+ make_tuple(16, 16, sad_16x16x4d_sse3),
+ make_tuple(16, 8, sad_16x8x4d_sse3),
+ make_tuple(8, 16, sad_8x16x4d_sse3),
+ make_tuple(8, 8, sad_8x8x4d_sse3),
+ make_tuple(4, 4, sad_4x4x4d_sse3)));
+#endif
#endif
#if HAVE_SSSE3
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index b9acadea7f0c6b869ccda0c01dbc76a79aa3c146..e6dcff4d1a16c9a2cbefbd8af7b2d8e34b3c6eef 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -470,25 +470,25 @@ specialize vp9_sad8x8x8 sse4
prototype void vp9_sad4x4x8 "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"
specialize vp9_sad4x4x8 sse4
-prototype void vp9_sad64x64x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad64x64x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad64x64x4d sse2
-prototype void vp9_sad32x32x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad32x32x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad32x32x4d sse2
-prototype void vp9_sad16x16x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad16x16x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad16x16x4d sse2
-prototype void vp9_sad16x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad16x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad16x8x4d sse2
-prototype void vp9_sad8x16x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad8x16x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad8x16x4d sse2
-prototype void vp9_sad8x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad8x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad8x8x4d sse2
-prototype void vp9_sad4x4x4d "const uint8_t *src_ptr, int src_stride, const uint8_t **ref_ptr, int ref_stride, unsigned int *sad_array"
+prototype void vp9_sad4x4x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad4x4x4d sse
prototype unsigned int vp9_sub_pixel_mse16x16 "const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse"
specialize vp9_sub_pixel_mse16x16 sse2 mmx
diff --git a/vp9/encoder/vp9_sad_c.c b/vp9/encoder/vp9_sad_c.c
index dc21f02f638b804ead98c58fb9b2889ee9fc2791..af5526dce333bf133eff44b957ea3c63951ab766 100644
--- a/vp9/encoder/vp9_sad_c.c
+++ b/vp9/encoder/vp9_sad_c.c
@@ -383,7 +383,7 @@ void vp9_sad4x4x8_c(const uint8_t *src_ptr,
void vp9_sad64x64x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad64x64(src_ptr, src_stride,
@@ -398,7 +398,7 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr,
void vp9_sad32x32x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad32x32(src_ptr, src_stride,
@@ -413,7 +413,7 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr,
void vp9_sad16x16x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad16x16(src_ptr, src_stride,
@@ -428,7 +428,7 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr,
void vp9_sad16x8x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad16x8(src_ptr, src_stride,
@@ -443,7 +443,7 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr,
void vp9_sad8x8x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad8x8(src_ptr, src_stride,
@@ -458,7 +458,7 @@ void vp9_sad8x8x4d_c(const uint8_t *src_ptr,
void vp9_sad8x16x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad8x16(src_ptr, src_stride,
@@ -473,7 +473,7 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr,
void vp9_sad4x4x4d_c(const uint8_t *src_ptr,
int src_stride,
- const uint8_t *ref_ptr[],
+ const uint8_t* const ref_ptr[],
int ref_stride,
unsigned int *sad_array) {
sad_array[0] = vp9_sad4x4(src_ptr, src_stride,
diff --git a/vp9/encoder/vp9_variance.h b/vp9/encoder/vp9_variance.h
index 7120c5fe7e1fb56140e2bdd9a7eeac1c80d917ef..13dabbda41d711a8dce115cf2892182450083a14 100644
--- a/vp9/encoder/vp9_variance.h
+++ b/vp9/encoder/vp9_variance.h
@@ -33,7 +33,7 @@ typedef void (*vp9_sad_multi1_fn_t)(const uint8_t *src_ptr,
typedef void (*vp9_sad_multi_d_fn_t)(const uint8_t *src_ptr,
int source_stride,
- const uint8_t ** ref_ptr,
+ const uint8_t* const ref_ptr[],
int ref_stride, unsigned int *sad_array);
typedef unsigned int (*vp9_variance_fn_t)(const uint8_t *src_ptr,