HBD hybrid transform 16x16 SSE4.1 optimization
- Tx_type: DCT_DCT, DCT_ADST, ADST_DCT, ADST_ADST.
- Update vp10_fht16x16_test.cc to do bit-exact test against
latest C version.
- HBD encoder speed improves ~1.8%.
Change-Id: Icfc799a212e5289bcf6cedcae3722032133a2bc6
diff --git a/test/vp10_fht16x16_test.cc b/test/vp10_fht16x16_test.cc
index 0e1bfa2..8833250 100644
--- a/test/vp10_fht16x16_test.cc
+++ b/test/vp10_fht16x16_test.cc
@@ -25,16 +25,30 @@
namespace {
typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
int tx_type);
-
+using std::tr1::tuple;
using libvpx_test::FhtFunc;
-typedef std::tr1::tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t, int>
-Ht16x16Param;
+typedef tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t, int> Ht16x16Param;
void fht16x16_ref(const int16_t *in, tran_low_t *out, int stride,
int tx_type) {
vp10_fht16x16_c(in, out, stride, tx_type);
}
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*IHbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+ int tx_type, int bd);
+typedef void (*HbdHtFunc)(const int16_t *input, int32_t *output, int stride,
+ int tx_type, int bd);
+
+// Target optimized function, tx_type, bit depth
+typedef tuple<HbdHtFunc, int, int> HighbdHt16x16Param;
+
+void highbd_fht16x16_ref(const int16_t *in, int32_t *out, int stride,
+ int tx_type, int bd) {
+ vp10_fwd_txfm2d_16x16_c(in, out, stride, tx_type, bd);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
class VP10Trans16x16HT
: public libvpx_test::TransformTestBase,
public ::testing::TestWithParam<Ht16x16Param> {
@@ -70,60 +84,77 @@
RunCoeffCheck();
}
-#if CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
-TEST(VP10Trans16x16HTSpeedTest, C_version) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = 20000;
- int bit_depth = 8;
- int mask = (1 << bit_depth) - 1;
- const int num_coeffs = 256;
- int16_t *input = new int16_t[num_coeffs];
- tran_low_t *output = new tran_low_t[num_coeffs];
- const int stride = 16;
- int tx_type;
+#if CONFIG_VP9_HIGHBITDEPTH
+class VP10HighbdTrans16x16HT
+ : public ::testing::TestWithParam<HighbdHt16x16Param> {
+ public:
+ virtual ~VP10HighbdTrans16x16HT() {}
- for (int i = 0; i < count_test_block; ++i) {
- for (int j = 0; j < num_coeffs; ++j) {
- input[j] = (rnd.Rand8() & mask) - (rnd.Rand8() & mask);
- }
- for (tx_type = V_DCT; tx_type <= H_FLIPADST; ++tx_type) {
- vp10_fht16x16_c(input, output, stride, tx_type);
- }
+ virtual void SetUp() {
+ fwd_txfm_ = GET_PARAM(0);
+ fwd_txfm_ref_ = highbd_fht16x16_ref;
+ tx_type_ = GET_PARAM(1);
+ bit_depth_ = GET_PARAM(2);
+ mask_ = (1 << bit_depth_) - 1;
+ num_coeffs_ = 256;
+
+ input_ = reinterpret_cast<int16_t *>(
+ vpx_memalign(16, sizeof(int16_t) * num_coeffs_));
+ output_ = reinterpret_cast<int32_t *>(
+ vpx_memalign(16, sizeof(int32_t) * num_coeffs_));
+ output_ref_ = reinterpret_cast<int32_t *>(
+ vpx_memalign(16, sizeof(int32_t) * num_coeffs_));
+ }
+
+ virtual void TearDown() {
+ vpx_free(input_);
+ vpx_free(output_);
+ vpx_free(output_ref_);
+ libvpx_test::ClearSystemState();
+ }
+
+ protected:
+ void RunBitexactCheck();
+
+ private:
+ HbdHtFunc fwd_txfm_;
+ HbdHtFunc fwd_txfm_ref_;
+ int tx_type_;
+ int bit_depth_;
+ int mask_;
+ int num_coeffs_;
+ int16_t *input_;
+ int32_t *output_;
+ int32_t *output_ref_;
+};
+
+void VP10HighbdTrans16x16HT::RunBitexactCheck() {
+ ACMRandom rnd(ACMRandom::DeterministicSeed());
+ int i, j;
+ const int stride = 16;
+ const int num_tests = 200000;
+
+ for (i = 0; i < num_tests; ++i) {
+ for (j = 0; j < num_coeffs_; ++j) {
+ input_[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
}
- delete[] input;
- delete[] output;
-}
-#endif // CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
+ fwd_txfm_ref_(input_, output_ref_, stride, tx_type_, bit_depth_);
+ ASM_REGISTER_STATE_CHECK(fwd_txfm_(input_, output_, stride, tx_type_,
+ bit_depth_));
-#if HAVE_SSE2 && CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
-TEST(VP10Trans16x16HTSpeedTest, SSE2_version) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = 20000;
- int bit_depth = 8;
- int mask = (1 << bit_depth) - 1;
- const int num_coeffs = 256;
- int16_t *input = reinterpret_cast<int16_t *>
- (vpx_memalign(16, sizeof(int16_t) * num_coeffs));
- tran_low_t *output = reinterpret_cast<tran_low_t *>
- (vpx_memalign(16, sizeof(tran_low_t) * num_coeffs));
-
- const int stride = 16;
- int tx_type;
-
- for (int i = 0; i < count_test_block; ++i) {
- for (int j = 0; j < num_coeffs; ++j) {
- input[j] = (rnd.Rand8() & mask) - (rnd.Rand8() & mask);
- }
- for (tx_type = V_DCT; tx_type <= H_FLIPADST; ++tx_type) {
- vp10_fht16x16_sse2(input, output, stride, tx_type);
- }
+ for (j = 0; j < num_coeffs_; ++j) {
+ EXPECT_EQ(output_ref_[j], output_[j])
+ << "Not bit-exact result at index: " << j
+ << " at test block: " << i;
}
-
- vpx_free(input);
- vpx_free(output);
+ }
}
-#endif // HAVE_SSE2 && CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
+
+TEST_P(VP10HighbdTrans16x16HT, HighbdCoeffCheck) {
+ RunBitexactCheck();
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
using std::tr1::make_tuple;
@@ -167,4 +198,20 @@
::testing::ValuesIn(kArrayHt16x16Param_sse2));
#endif // HAVE_SSE2
+#if HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
+const HighbdHt16x16Param kArrayHBDHt16x16Param_sse4_1[] = {
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 0, 10),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 0, 12),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 1, 10),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 1, 12),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 2, 10),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 2, 12),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 3, 10),
+ make_tuple(&vp10_fwd_txfm2d_16x16_sse4_1, 3, 12)
+};
+INSTANTIATE_TEST_CASE_P(
+ SSE4_1, VP10HighbdTrans16x16HT,
+ ::testing::ValuesIn(kArrayHBDHt16x16Param_sse4_1));
+#endif // HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
+
} // namespace
