HBD hybrid transform 8x8 SSE4.1 optimization
- Tx_type: DCT_DCT, DCT_ADST, ADST_DCT, ADST_ADST.
- Update bit-exact unit test against current C version.
- HBD encoder speed improves ~3.8%.
Change-Id: Ie13925ba11214eef2b5326814940638507bf68ec
diff --git a/test/vp10_fht8x8_test.cc b/test/vp10_fht8x8_test.cc
index df1826f..468b8c9 100644
--- a/test/vp10_fht8x8_test.cc
+++ b/test/vp10_fht8x8_test.cc
@@ -27,13 +27,28 @@
int tx_type);
using libvpx_test::FhtFunc;
-typedef std::tr1::tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t, int> Ht8x8Param;
+using std::tr1::tuple;
+typedef tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t, int> Ht8x8Param;
void fht8x8_ref(const int16_t *in, tran_low_t *out, int stride,
int tx_type) {
vp10_fht8x8_c(in, out, stride, tx_type);
}
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*IhighbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+ int tx_type, int bd);
+typedef void (*HBDFhtFunc)(const int16_t *input, int32_t *output, int stride,
+ int tx_type, int bd);
+// Target optimized function, tx_type, bit depth
+typedef tuple<HBDFhtFunc, int, int> HighbdHt8x8Param;
+
+void highbe_fht8x8_ref(const int16_t *in, int32_t *out, int stride,
+ int tx_type, int bd) {
+ vp10_fwd_txfm2d_8x8_c(in, out, stride, tx_type, bd);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
class VP10Trans8x8HT
: public libvpx_test::TransformTestBase,
public ::testing::TestWithParam<Ht8x8Param> {
@@ -69,60 +84,76 @@
RunCoeffCheck();
}
-#if CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
-TEST(VP10Trans8x8HTSpeedTest, 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 = 64;
- int16_t *input = new int16_t[num_coeffs];
- tran_low_t *output = new tran_low_t[num_coeffs];
- const int stride = 8;
- int tx_type;
+#if CONFIG_VP9_HIGHBITDEPTH
+class VP10HighbdTrans8x8HT : public ::testing::TestWithParam<HighbdHt8x8Param> {
+ public:
+ virtual ~VP10HighbdTrans8x8HT() {}
- 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_fht8x8_c(input, output, stride, tx_type);
- }
+ virtual void SetUp() {
+ fwd_txfm_ = GET_PARAM(0);
+ fwd_txfm_ref_ = highbe_fht8x8_ref;
+ tx_type_ = GET_PARAM(1);
+ bit_depth_ = GET_PARAM(2);
+ mask_ = (1 << bit_depth_) - 1;
+ num_coeffs_ = 64;
+
+ 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:
+ HBDFhtFunc fwd_txfm_;
+ HBDFhtFunc 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 VP10HighbdTrans8x8HT::RunBitexactCheck() {
+ ACMRandom rnd(ACMRandom::DeterministicSeed());
+ int i, j;
+ const int stride = 8;
+ const int num_tests = 200000;
+ const int num_coeffs = 64;
+
+ 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_);
+ fwd_txfm_(input_, output_, stride, tx_type_, bit_depth_);
-#if HAVE_SSE2 && CONFIG_EXT_TX && !CONFIG_VP9_HIGHBITDEPTH
-TEST(VP10Trans8x8HTSpeedTest, 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 = 64;
- 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 = 8;
- 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_fht8x8_sse2(input, output, stride, tx_type);
- }
+ for (j = 0; j < num_coeffs; ++j) {
+ EXPECT_EQ(output_[j], output_ref_[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(VP10HighbdTrans8x8HT, HighbdCoeffCheck) {
+ RunBitexactCheck();
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
using std::tr1::make_tuple;
@@ -166,4 +197,20 @@
::testing::ValuesIn(kArrayHt8x8Param_sse2));
#endif // HAVE_SSE2
+#if HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
+const HighbdHt8x8Param kArrayHBDHt8x8Param_sse4_1[] = {
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 0, 10),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 0, 12),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 1, 10),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 1, 12),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 2, 10),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 2, 12),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 3, 10),
+ make_tuple(&vp10_fwd_txfm2d_8x8_sse4_1, 3, 12)
+};
+INSTANTIATE_TEST_CASE_P(
+ SSE4_1, VP10HighbdTrans8x8HT,
+ ::testing::ValuesIn(kArrayHBDHt8x8Param_sse4_1));
+#endif // HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
+
} // namespace
diff --git a/vp10/common/x86/idct_intrin_sse2.c b/vp10/common/x86/idct_intrin_sse2.c
index 900f091..d8b6d95 100644
--- a/vp10/common/x86/idct_intrin_sse2.c
+++ b/vp10/common/x86/idct_intrin_sse2.c
@@ -14,13 +14,6 @@
#include "vp10/common/enums.h"
#if CONFIG_EXT_TX
-// Reverse the 8 16 bit words in __m128i
-static INLINE __m128i mm_reverse_epi16(const __m128i x) {
- const __m128i a = _mm_shufflelo_epi16(x, 0x1b);
- const __m128i b = _mm_shufflehi_epi16(a, 0x1b);
- return _mm_shuffle_epi32(b, 0x4e);
-}
-
static INLINE void fliplr_4x4(__m128i in[2]) {
in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
in[0] = _mm_shufflehi_epi16(in[0], 0x1b);
diff --git a/vp10/common/x86/vp10_fwd_txfm2d_sse4.c b/vp10/common/x86/vp10_fwd_txfm2d_sse4.c
index 1b11087..daed25e 100644
--- a/vp10/common/x86/vp10_fwd_txfm2d_sse4.c
+++ b/vp10/common/x86/vp10_fwd_txfm2d_sse4.c
@@ -87,15 +87,6 @@
transpose_32(txfm_size, buf_128, out_128);
}
-void vp10_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output,
- const int stride, int tx_type,
- const int bd) {
- int32_t txfm_buf[64];
- const TXFM_2D_CFG* cfg = vp10_get_txfm_8x8_cfg(tx_type);
- (void)bd;
- fwd_txfm2d_sse4_1(input, output, stride, cfg, txfm_buf);
-}
-
void vp10_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output,
const int stride, int tx_type,
const int bd) {
diff --git a/vp10/encoder/x86/dct_sse2.c b/vp10/encoder/x86/dct_sse2.c
index 47422ad..ea0ccb8 100644
--- a/vp10/encoder/x86/dct_sse2.c
+++ b/vp10/encoder/x86/dct_sse2.c
@@ -18,13 +18,6 @@
#include "vpx_dsp/x86/txfm_common_sse2.h"
#include "vpx_ports/mem.h"
-// Reverse the 8 16 bit words in __m128i
-static INLINE __m128i mm_reverse_epi16(const __m128i x) {
- const __m128i a = _mm_shufflelo_epi16(x, 0x1b);
- const __m128i b = _mm_shufflehi_epi16(a, 0x1b);
- return _mm_shuffle_epi32(b, 0x4e);
-}
-
static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
int stride, int flipud, int fliplr) {
const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
diff --git a/vp10/encoder/x86/highbd_fwd_txfm_sse4.c b/vp10/encoder/x86/highbd_fwd_txfm_sse4.c
index 3cda783..949816c 100644
--- a/vp10/encoder/x86/highbd_fwd_txfm_sse4.c
+++ b/vp10/encoder/x86/highbd_fwd_txfm_sse4.c
@@ -16,6 +16,7 @@
#include "vp10/common/vp10_fwd_txfm2d_cfg.h"
#include "vp10/common/vp10_txfm.h"
#include "vpx_dsp/txfm_common.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
#include "vpx_ports/mem.h"
static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
@@ -244,3 +245,726 @@
}
(void)bd;
}
+
+static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
+ int stride, int flipud, int fliplr,
+ int shift) {
+ __m128i u;
+ if (!flipud) {
+ in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+ in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+ in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+ in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+ in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+ in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+ in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+ in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+ } else {
+ in[0] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+ in[1] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+ in[2] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+ in[3] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+ in[4] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+ in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+ in[6] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+ in[7] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+ }
+
+ if (fliplr) {
+ in[0] = mm_reverse_epi16(in[0]);
+ in[1] = mm_reverse_epi16(in[1]);
+ in[2] = mm_reverse_epi16(in[2]);
+ in[3] = mm_reverse_epi16(in[3]);
+ in[4] = mm_reverse_epi16(in[4]);
+ in[5] = mm_reverse_epi16(in[5]);
+ in[6] = mm_reverse_epi16(in[6]);
+ in[7] = mm_reverse_epi16(in[7]);
+ }
+
+ u = _mm_unpackhi_epi64(in[4], in[4]);
+ in[8] = _mm_cvtepi16_epi32(in[4]);
+ in[9] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[5], in[5]);
+ in[10] = _mm_cvtepi16_epi32(in[5]);
+ in[11] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[6], in[6]);
+ in[12] = _mm_cvtepi16_epi32(in[6]);
+ in[13] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[7], in[7]);
+ in[14] = _mm_cvtepi16_epi32(in[7]);
+ in[15] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[3], in[3]);
+ in[6] = _mm_cvtepi16_epi32(in[3]);
+ in[7] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[2], in[2]);
+ in[4] = _mm_cvtepi16_epi32(in[2]);
+ in[5] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[1], in[1]);
+ in[2] = _mm_cvtepi16_epi32(in[1]);
+ in[3] = _mm_cvtepi16_epi32(u);
+
+ u = _mm_unpackhi_epi64(in[0], in[0]);
+ in[0] = _mm_cvtepi16_epi32(in[0]);
+ in[1] = _mm_cvtepi16_epi32(u);
+
+ in[0] = _mm_slli_epi32(in[0], shift);
+ in[1] = _mm_slli_epi32(in[1], shift);
+ in[2] = _mm_slli_epi32(in[2], shift);
+ in[3] = _mm_slli_epi32(in[3], shift);
+ in[4] = _mm_slli_epi32(in[4], shift);
+ in[5] = _mm_slli_epi32(in[5], shift);
+ in[6] = _mm_slli_epi32(in[6], shift);
+ in[7] = _mm_slli_epi32(in[7], shift);
+
+ in[8] = _mm_slli_epi32(in[8], shift);
+ in[9] = _mm_slli_epi32(in[9], shift);
+ in[10] = _mm_slli_epi32(in[10], shift);
+ in[11] = _mm_slli_epi32(in[11], shift);
+ in[12] = _mm_slli_epi32(in[12], shift);
+ in[13] = _mm_slli_epi32(in[13], shift);
+ in[14] = _mm_slli_epi32(in[14], shift);
+ in[15] = _mm_slli_epi32(in[15], shift);
+}
+
+static INLINE void col_txfm_8x8_rounding(__m128i *in, int shift) {
+ const __m128i rounding = _mm_set1_epi32(1 << (shift - 1));
+
+ in[0] = _mm_add_epi32(in[0], rounding);
+ in[1] = _mm_add_epi32(in[1], rounding);
+ in[2] = _mm_add_epi32(in[2], rounding);
+ in[3] = _mm_add_epi32(in[3], rounding);
+ in[4] = _mm_add_epi32(in[4], rounding);
+ in[5] = _mm_add_epi32(in[5], rounding);
+ in[6] = _mm_add_epi32(in[6], rounding);
+ in[7] = _mm_add_epi32(in[7], rounding);
+ in[8] = _mm_add_epi32(in[8], rounding);
+ in[9] = _mm_add_epi32(in[9], rounding);
+ in[10] = _mm_add_epi32(in[10], rounding);
+ in[11] = _mm_add_epi32(in[11], rounding);
+ in[12] = _mm_add_epi32(in[12], rounding);
+ in[13] = _mm_add_epi32(in[13], rounding);
+ in[14] = _mm_add_epi32(in[14], rounding);
+ in[15] = _mm_add_epi32(in[15], rounding);
+
+ in[0] = _mm_srai_epi32(in[0], shift);
+ in[1] = _mm_srai_epi32(in[1], shift);
+ in[2] = _mm_srai_epi32(in[2], shift);
+ in[3] = _mm_srai_epi32(in[3], shift);
+ in[4] = _mm_srai_epi32(in[4], shift);
+ in[5] = _mm_srai_epi32(in[5], shift);
+ in[6] = _mm_srai_epi32(in[6], shift);
+ in[7] = _mm_srai_epi32(in[7], shift);
+ in[8] = _mm_srai_epi32(in[8], shift);
+ in[9] = _mm_srai_epi32(in[9], shift);
+ in[10] = _mm_srai_epi32(in[10], shift);
+ in[11] = _mm_srai_epi32(in[11], shift);
+ in[12] = _mm_srai_epi32(in[12], shift);
+ in[13] = _mm_srai_epi32(in[13], shift);
+ in[14] = _mm_srai_epi32(in[14], shift);
+ in[15] = _mm_srai_epi32(in[15], shift);
+}
+
+#define TRANSPOSE_4X4(x0, x1, x2, x3, y0, y1, y2, y3) \
+ do { \
+ __m128i u0, u1, u2, u3; \
+ u0 = _mm_unpacklo_epi32(x0, x1); \
+ u1 = _mm_unpackhi_epi32(x0, x1); \
+ u2 = _mm_unpacklo_epi32(x2, x3); \
+ u3 = _mm_unpackhi_epi32(x2, x3); \
+ y0 = _mm_unpacklo_epi64(u0, u2); \
+ y1 = _mm_unpackhi_epi64(u0, u2); \
+ y2 = _mm_unpacklo_epi64(u1, u3); \
+ y3 = _mm_unpackhi_epi64(u1, u3); \
+ } while (0)
+
+static INLINE void transpose_8x8(__m128i *in) {
+ __m128i t[4];
+
+ TRANSPOSE_4X4(in[0], in[2], in[4], in[6], in[0], in[2], in[4], in[6]);
+ TRANSPOSE_4X4(in[1], in[3], in[5], in[7], t[0], t[1], t[2], t[3]);
+ TRANSPOSE_4X4(in[8], in[10], in[12], in[14], in[1], in[3], in[5], in[7]);
+ in[8] = t[0];
+ in[10] = t[1];
+ in[12] = t[2];
+ in[14] = t[3];
+ TRANSPOSE_4X4(in[9], in[11], in[13], in[15], in[9], in[11], in[13], in[15]);
+}
+
+static INLINE void write_buffer_8x8(__m128i *res, tran_low_t *output) {
+ _mm_store_si128((__m128i *)(output + 0 * 4), res[0]);
+ _mm_store_si128((__m128i *)(output + 1 * 4), res[1]);
+ _mm_store_si128((__m128i *)(output + 2 * 4), res[2]);
+ _mm_store_si128((__m128i *)(output + 3 * 4), res[3]);
+
+ _mm_store_si128((__m128i *)(output + 4 * 4), res[4]);
+ _mm_store_si128((__m128i *)(output + 5 * 4), res[5]);
+ _mm_store_si128((__m128i *)(output + 6 * 4), res[6]);
+ _mm_store_si128((__m128i *)(output + 7 * 4), res[7]);
+
+ _mm_store_si128((__m128i *)(output + 8 * 4), res[8]);
+ _mm_store_si128((__m128i *)(output + 9 * 4), res[9]);
+ _mm_store_si128((__m128i *)(output + 10 * 4), res[10]);
+ _mm_store_si128((__m128i *)(output + 11 * 4), res[11]);
+
+ _mm_store_si128((__m128i *)(output + 12 * 4), res[12]);
+ _mm_store_si128((__m128i *)(output + 13 * 4), res[13]);
+ _mm_store_si128((__m128i *)(output + 14 * 4), res[14]);
+ _mm_store_si128((__m128i *)(output + 15 * 4), res[15]);
+}
+
+static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+ const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+ const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+ const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+ const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+ const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+ const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+ const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+ const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+ const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+ const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+ __m128i u[8], v[8];
+
+ // Even 8 points 0, 2, ..., 14
+ // stage 0
+ // stage 1
+ u[0] = _mm_add_epi32(in[0], in[14]);
+ v[7] = _mm_sub_epi32(in[0], in[14]); // v[7]
+ u[1] = _mm_add_epi32(in[2], in[12]);
+ u[6] = _mm_sub_epi32(in[2], in[12]);
+ u[2] = _mm_add_epi32(in[4], in[10]);
+ u[5] = _mm_sub_epi32(in[4], in[10]);
+ u[3] = _mm_add_epi32(in[6], in[8]);
+ v[4] = _mm_sub_epi32(in[6], in[8]); // v[4]
+
+ // stage 2
+ v[0] = _mm_add_epi32(u[0], u[3]);
+ v[3] = _mm_sub_epi32(u[0], u[3]);
+ v[1] = _mm_add_epi32(u[1], u[2]);
+ v[2] = _mm_sub_epi32(u[1], u[2]);
+
+ v[5] = _mm_mullo_epi32(u[5], cospim32);
+ v[6] = _mm_mullo_epi32(u[6], cospi32);
+ v[5] = _mm_add_epi32(v[5], v[6]);
+ v[5] = _mm_add_epi32(v[5], rnding);
+ v[5] = _mm_srai_epi32(v[5], bit);
+
+ u[0] = _mm_mullo_epi32(u[5], cospi32);
+ v[6] = _mm_mullo_epi32(u[6], cospim32);
+ v[6] = _mm_sub_epi32(u[0], v[6]);
+ v[6] = _mm_add_epi32(v[6], rnding);
+ v[6] = _mm_srai_epi32(v[6], bit);
+
+ // stage 3
+ // type 0
+ v[0] = _mm_mullo_epi32(v[0], cospi32);
+ v[1] = _mm_mullo_epi32(v[1], cospi32);
+ u[0] = _mm_add_epi32(v[0], v[1]);
+ u[0] = _mm_add_epi32(u[0], rnding);
+ u[0] = _mm_srai_epi32(u[0], bit);
+
+ u[1] = _mm_sub_epi32(v[0], v[1]);
+ u[1] = _mm_add_epi32(u[1], rnding);
+ u[1] = _mm_srai_epi32(u[1], bit);
+
+ // type 1
+ v[0] = _mm_mullo_epi32(v[2], cospi48);
+ v[1] = _mm_mullo_epi32(v[3], cospi16);
+ u[2] = _mm_add_epi32(v[0], v[1]);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ v[0] = _mm_mullo_epi32(v[2], cospi16);
+ v[1] = _mm_mullo_epi32(v[3], cospi48);
+ u[3] = _mm_sub_epi32(v[1], v[0]);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ u[4] = _mm_add_epi32(v[4], v[5]);
+ u[5] = _mm_sub_epi32(v[4], v[5]);
+ u[6] = _mm_sub_epi32(v[7], v[6]);
+ u[7] = _mm_add_epi32(v[7], v[6]);
+
+ // stage 4
+ // stage 5
+ v[0] = _mm_mullo_epi32(u[4], cospi56);
+ v[1] = _mm_mullo_epi32(u[7], cospi8);
+ v[0] = _mm_add_epi32(v[0], v[1]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[2] = _mm_srai_epi32(v[0], bit); // buf0[4]
+
+ v[0] = _mm_mullo_epi32(u[4], cospi8);
+ v[1] = _mm_mullo_epi32(u[7], cospi56);
+ v[0] = _mm_sub_epi32(v[1], v[0]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[14] = _mm_srai_epi32(v[0], bit); // buf0[7]
+
+ v[0] = _mm_mullo_epi32(u[5], cospi24);
+ v[1] = _mm_mullo_epi32(u[6], cospi40);
+ v[0] = _mm_add_epi32(v[0], v[1]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[10] = _mm_srai_epi32(v[0], bit); // buf0[5]
+
+ v[0] = _mm_mullo_epi32(u[5], cospi40);
+ v[1] = _mm_mullo_epi32(u[6], cospi24);
+ v[0] = _mm_sub_epi32(v[1], v[0]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[6] = _mm_srai_epi32(v[0], bit); // buf0[6]
+
+ out[0] = u[0]; // buf0[0]
+ out[8] = u[1]; // buf0[1]
+ out[4] = u[2]; // buf0[2]
+ out[12] = u[3]; // buf0[3]
+
+ // Odd 8 points: 1, 3, ..., 15
+ // stage 0
+ // stage 1
+ u[0] = _mm_add_epi32(in[1], in[15]);
+ v[7] = _mm_sub_epi32(in[1], in[15]); // v[7]
+ u[1] = _mm_add_epi32(in[3], in[13]);
+ u[6] = _mm_sub_epi32(in[3], in[13]);
+ u[2] = _mm_add_epi32(in[5], in[11]);
+ u[5] = _mm_sub_epi32(in[5], in[11]);
+ u[3] = _mm_add_epi32(in[7], in[9]);
+ v[4] = _mm_sub_epi32(in[7], in[9]); // v[4]
+
+ // stage 2
+ v[0] = _mm_add_epi32(u[0], u[3]);
+ v[3] = _mm_sub_epi32(u[0], u[3]);
+ v[1] = _mm_add_epi32(u[1], u[2]);
+ v[2] = _mm_sub_epi32(u[1], u[2]);
+
+ v[5] = _mm_mullo_epi32(u[5], cospim32);
+ v[6] = _mm_mullo_epi32(u[6], cospi32);
+ v[5] = _mm_add_epi32(v[5], v[6]);
+ v[5] = _mm_add_epi32(v[5], rnding);
+ v[5] = _mm_srai_epi32(v[5], bit);
+
+ u[0] = _mm_mullo_epi32(u[5], cospi32);
+ v[6] = _mm_mullo_epi32(u[6], cospim32);
+ v[6] = _mm_sub_epi32(u[0], v[6]);
+ v[6] = _mm_add_epi32(v[6], rnding);
+ v[6] = _mm_srai_epi32(v[6], bit);
+
+ // stage 3
+ // type 0
+ v[0] = _mm_mullo_epi32(v[0], cospi32);
+ v[1] = _mm_mullo_epi32(v[1], cospi32);
+ u[0] = _mm_add_epi32(v[0], v[1]);
+ u[0] = _mm_add_epi32(u[0], rnding);
+ u[0] = _mm_srai_epi32(u[0], bit);
+
+ u[1] = _mm_sub_epi32(v[0], v[1]);
+ u[1] = _mm_add_epi32(u[1], rnding);
+ u[1] = _mm_srai_epi32(u[1], bit);
+
+ // type 1
+ v[0] = _mm_mullo_epi32(v[2], cospi48);
+ v[1] = _mm_mullo_epi32(v[3], cospi16);
+ u[2] = _mm_add_epi32(v[0], v[1]);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ v[0] = _mm_mullo_epi32(v[2], cospi16);
+ v[1] = _mm_mullo_epi32(v[3], cospi48);
+ u[3] = _mm_sub_epi32(v[1], v[0]);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ u[4] = _mm_add_epi32(v[4], v[5]);
+ u[5] = _mm_sub_epi32(v[4], v[5]);
+ u[6] = _mm_sub_epi32(v[7], v[6]);
+ u[7] = _mm_add_epi32(v[7], v[6]);
+
+ // stage 4
+ // stage 5
+ v[0] = _mm_mullo_epi32(u[4], cospi56);
+ v[1] = _mm_mullo_epi32(u[7], cospi8);
+ v[0] = _mm_add_epi32(v[0], v[1]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[3] = _mm_srai_epi32(v[0], bit); // buf0[4]
+
+ v[0] = _mm_mullo_epi32(u[4], cospi8);
+ v[1] = _mm_mullo_epi32(u[7], cospi56);
+ v[0] = _mm_sub_epi32(v[1], v[0]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[15] = _mm_srai_epi32(v[0], bit); // buf0[7]
+
+ v[0] = _mm_mullo_epi32(u[5], cospi24);
+ v[1] = _mm_mullo_epi32(u[6], cospi40);
+ v[0] = _mm_add_epi32(v[0], v[1]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[11] = _mm_srai_epi32(v[0], bit); // buf0[5]
+
+ v[0] = _mm_mullo_epi32(u[5], cospi40);
+ v[1] = _mm_mullo_epi32(u[6], cospi24);
+ v[0] = _mm_sub_epi32(v[1], v[0]);
+ v[0] = _mm_add_epi32(v[0], rnding);
+ out[7] = _mm_srai_epi32(v[0], bit); // buf0[6]
+
+ out[1] = u[0]; // buf0[0]
+ out[9] = u[1]; // buf0[1]
+ out[5] = u[2]; // buf0[2]
+ out[13] = u[3]; // buf0[3]
+}
+
+static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+ const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+ const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+ const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+ const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+ const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+ const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+ const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+ const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+ const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+ const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+ const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+ const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+ const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+ const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+ const __m128i kZero = _mm_setzero_si128();
+ __m128i u[8], v[8], x;
+
+ // Even 8 points: 0, 2, ..., 14
+ // stage 0
+ // stage 1
+ // stage 2
+ // (1)
+ u[0] = _mm_mullo_epi32(in[14], cospi4);
+ x = _mm_mullo_epi32(in[0], cospi60);
+ u[0] = _mm_add_epi32(u[0], x);
+ u[0] = _mm_add_epi32(u[0], rnding);
+ u[0] = _mm_srai_epi32(u[0], bit);
+
+ u[1] = _mm_mullo_epi32(in[14], cospi60);
+ x = _mm_mullo_epi32(in[0], cospi4);
+ u[1] = _mm_sub_epi32(u[1], x);
+ u[1] = _mm_add_epi32(u[1], rnding);
+ u[1] = _mm_srai_epi32(u[1], bit);
+
+ // (2)
+ u[2] = _mm_mullo_epi32(in[10], cospi20);
+ x = _mm_mullo_epi32(in[4], cospi44);
+ u[2] = _mm_add_epi32(u[2], x);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ u[3] = _mm_mullo_epi32(in[10], cospi44);
+ x = _mm_mullo_epi32(in[4], cospi20);
+ u[3] = _mm_sub_epi32(u[3], x);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ // (3)
+ u[4] = _mm_mullo_epi32(in[6], cospi36);
+ x = _mm_mullo_epi32(in[8], cospi28);
+ u[4] = _mm_add_epi32(u[4], x);
+ u[4] = _mm_add_epi32(u[4], rnding);
+ u[4] = _mm_srai_epi32(u[4], bit);
+
+ u[5] = _mm_mullo_epi32(in[6], cospi28);
+ x = _mm_mullo_epi32(in[8], cospi36);
+ u[5] = _mm_sub_epi32(u[5], x);
+ u[5] = _mm_add_epi32(u[5], rnding);
+ u[5] = _mm_srai_epi32(u[5], bit);
+
+ // (4)
+ u[6] = _mm_mullo_epi32(in[2], cospi52);
+ x = _mm_mullo_epi32(in[12], cospi12);
+ u[6] = _mm_add_epi32(u[6], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_mullo_epi32(in[2], cospi12);
+ x = _mm_mullo_epi32(in[12], cospi52);
+ u[7] = _mm_sub_epi32(u[7], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 3
+ v[0] = _mm_add_epi32(u[0], u[4]);
+ v[4] = _mm_sub_epi32(u[0], u[4]);
+ v[1] = _mm_add_epi32(u[1], u[5]);
+ v[5] = _mm_sub_epi32(u[1], u[5]);
+ v[2] = _mm_add_epi32(u[2], u[6]);
+ v[6] = _mm_sub_epi32(u[2], u[6]);
+ v[3] = _mm_add_epi32(u[3], u[7]);
+ v[7] = _mm_sub_epi32(u[3], u[7]);
+
+ // stage 4
+ u[0] = v[0];
+ u[1] = v[1];
+ u[2] = v[2];
+ u[3] = v[3];
+
+ u[4] = _mm_mullo_epi32(v[4], cospi16);
+ x = _mm_mullo_epi32(v[5], cospi48);
+ u[4] = _mm_add_epi32(u[4], x);
+ u[4] = _mm_add_epi32(u[4], rnding);
+ u[4] = _mm_srai_epi32(u[4], bit);
+
+ u[5] = _mm_mullo_epi32(v[4], cospi48);
+ x = _mm_mullo_epi32(v[5], cospi16);
+ u[5] = _mm_sub_epi32(u[5], x);
+ u[5] = _mm_add_epi32(u[5], rnding);
+ u[5] = _mm_srai_epi32(u[5], bit);
+
+ u[6] = _mm_mullo_epi32(v[6], cospim48);
+ x = _mm_mullo_epi32(v[7], cospi16);
+ u[6] = _mm_add_epi32(u[6], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_mullo_epi32(v[6], cospi16);
+ x = _mm_mullo_epi32(v[7], cospim48);
+ u[7] = _mm_sub_epi32(u[7], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 5
+ v[0] = _mm_add_epi32(u[0], u[2]);
+ v[2] = _mm_sub_epi32(u[0], u[2]);
+ v[1] = _mm_add_epi32(u[1], u[3]);
+ v[3] = _mm_sub_epi32(u[1], u[3]);
+ v[4] = _mm_add_epi32(u[4], u[6]);
+ v[6] = _mm_sub_epi32(u[4], u[6]);
+ v[5] = _mm_add_epi32(u[5], u[7]);
+ v[7] = _mm_sub_epi32(u[5], u[7]);
+
+ // stage 6
+ u[0] = v[0];
+ u[1] = v[1];
+ u[4] = v[4];
+ u[5] = v[5];
+
+ v[0] = _mm_mullo_epi32(v[2], cospi32);
+ x = _mm_mullo_epi32(v[3], cospi32);
+ u[2] = _mm_add_epi32(v[0], x);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ u[3] = _mm_sub_epi32(v[0], x);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ v[0] = _mm_mullo_epi32(v[6], cospi32);
+ x = _mm_mullo_epi32(v[7], cospi32);
+ u[6] = _mm_add_epi32(v[0], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_sub_epi32(v[0], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 7
+ out[0] = u[0];
+ out[2] = _mm_sub_epi32(kZero, u[4]);
+ out[4] = u[6];
+ out[6] = _mm_sub_epi32(kZero, u[2]);
+ out[8] = u[3];
+ out[10] = _mm_sub_epi32(kZero, u[7]);
+ out[12] = u[5];
+ out[14] = _mm_sub_epi32(kZero, u[1]);
+
+ // Odd 8 points: 1, 3, ..., 15
+ // stage 0
+ // stage 1
+ // stage 2
+ // (1)
+ u[0] = _mm_mullo_epi32(in[15], cospi4);
+ x = _mm_mullo_epi32(in[1], cospi60);
+ u[0] = _mm_add_epi32(u[0], x);
+ u[0] = _mm_add_epi32(u[0], rnding);
+ u[0] = _mm_srai_epi32(u[0], bit);
+
+ u[1] = _mm_mullo_epi32(in[15], cospi60);
+ x = _mm_mullo_epi32(in[1], cospi4);
+ u[1] = _mm_sub_epi32(u[1], x);
+ u[1] = _mm_add_epi32(u[1], rnding);
+ u[1] = _mm_srai_epi32(u[1], bit);
+
+ // (2)
+ u[2] = _mm_mullo_epi32(in[11], cospi20);
+ x = _mm_mullo_epi32(in[5], cospi44);
+ u[2] = _mm_add_epi32(u[2], x);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ u[3] = _mm_mullo_epi32(in[11], cospi44);
+ x = _mm_mullo_epi32(in[5], cospi20);
+ u[3] = _mm_sub_epi32(u[3], x);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ // (3)
+ u[4] = _mm_mullo_epi32(in[7], cospi36);
+ x = _mm_mullo_epi32(in[9], cospi28);
+ u[4] = _mm_add_epi32(u[4], x);
+ u[4] = _mm_add_epi32(u[4], rnding);
+ u[4] = _mm_srai_epi32(u[4], bit);
+
+ u[5] = _mm_mullo_epi32(in[7], cospi28);
+ x = _mm_mullo_epi32(in[9], cospi36);
+ u[5] = _mm_sub_epi32(u[5], x);
+ u[5] = _mm_add_epi32(u[5], rnding);
+ u[5] = _mm_srai_epi32(u[5], bit);
+
+ // (4)
+ u[6] = _mm_mullo_epi32(in[3], cospi52);
+ x = _mm_mullo_epi32(in[13], cospi12);
+ u[6] = _mm_add_epi32(u[6], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_mullo_epi32(in[3], cospi12);
+ x = _mm_mullo_epi32(in[13], cospi52);
+ u[7] = _mm_sub_epi32(u[7], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 3
+ v[0] = _mm_add_epi32(u[0], u[4]);
+ v[4] = _mm_sub_epi32(u[0], u[4]);
+ v[1] = _mm_add_epi32(u[1], u[5]);
+ v[5] = _mm_sub_epi32(u[1], u[5]);
+ v[2] = _mm_add_epi32(u[2], u[6]);
+ v[6] = _mm_sub_epi32(u[2], u[6]);
+ v[3] = _mm_add_epi32(u[3], u[7]);
+ v[7] = _mm_sub_epi32(u[3], u[7]);
+
+ // stage 4
+ u[0] = v[0];
+ u[1] = v[1];
+ u[2] = v[2];
+ u[3] = v[3];
+
+ u[4] = _mm_mullo_epi32(v[4], cospi16);
+ x = _mm_mullo_epi32(v[5], cospi48);
+ u[4] = _mm_add_epi32(u[4], x);
+ u[4] = _mm_add_epi32(u[4], rnding);
+ u[4] = _mm_srai_epi32(u[4], bit);
+
+ u[5] = _mm_mullo_epi32(v[4], cospi48);
+ x = _mm_mullo_epi32(v[5], cospi16);
+ u[5] = _mm_sub_epi32(u[5], x);
+ u[5] = _mm_add_epi32(u[5], rnding);
+ u[5] = _mm_srai_epi32(u[5], bit);
+
+ u[6] = _mm_mullo_epi32(v[6], cospim48);
+ x = _mm_mullo_epi32(v[7], cospi16);
+ u[6] = _mm_add_epi32(u[6], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_mullo_epi32(v[6], cospi16);
+ x = _mm_mullo_epi32(v[7], cospim48);
+ u[7] = _mm_sub_epi32(u[7], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 5
+ v[0] = _mm_add_epi32(u[0], u[2]);
+ v[2] = _mm_sub_epi32(u[0], u[2]);
+ v[1] = _mm_add_epi32(u[1], u[3]);
+ v[3] = _mm_sub_epi32(u[1], u[3]);
+ v[4] = _mm_add_epi32(u[4], u[6]);
+ v[6] = _mm_sub_epi32(u[4], u[6]);
+ v[5] = _mm_add_epi32(u[5], u[7]);
+ v[7] = _mm_sub_epi32(u[5], u[7]);
+
+ // stage 6
+ u[0] = v[0];
+ u[1] = v[1];
+ u[4] = v[4];
+ u[5] = v[5];
+
+ v[0] = _mm_mullo_epi32(v[2], cospi32);
+ x = _mm_mullo_epi32(v[3], cospi32);
+ u[2] = _mm_add_epi32(v[0], x);
+ u[2] = _mm_add_epi32(u[2], rnding);
+ u[2] = _mm_srai_epi32(u[2], bit);
+
+ u[3] = _mm_sub_epi32(v[0], x);
+ u[3] = _mm_add_epi32(u[3], rnding);
+ u[3] = _mm_srai_epi32(u[3], bit);
+
+ v[0] = _mm_mullo_epi32(v[6], cospi32);
+ x = _mm_mullo_epi32(v[7], cospi32);
+ u[6] = _mm_add_epi32(v[0], x);
+ u[6] = _mm_add_epi32(u[6], rnding);
+ u[6] = _mm_srai_epi32(u[6], bit);
+
+ u[7] = _mm_sub_epi32(v[0], x);
+ u[7] = _mm_add_epi32(u[7], rnding);
+ u[7] = _mm_srai_epi32(u[7], bit);
+
+ // stage 7
+ out[1] = u[0];
+ out[3] = _mm_sub_epi32(kZero, u[4]);
+ out[5] = u[6];
+ out[7] = _mm_sub_epi32(kZero, u[2]);
+ out[9] = u[3];
+ out[11] = _mm_sub_epi32(kZero, u[7]);
+ out[13] = u[5];
+ out[15] = _mm_sub_epi32(kZero, u[1]);
+}
+
+void vp10_fwd_txfm2d_8x8_sse4_1(const int16_t *input, tran_low_t *coeff,
+ int stride, int tx_type, int bd) {
+ __m128i in[16], out[16];
+ const TXFM_2D_CFG *cfg = NULL;
+
+ switch (tx_type) {
+ case DCT_DCT:
+ cfg = &fwd_txfm_2d_cfg_dct_dct_8;
+ load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+ fdct8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+ col_txfm_8x8_rounding(out, -cfg->shift[1]);
+ transpose_8x8(out);
+ fdct8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+ transpose_8x8(in);
+ write_buffer_8x8(in, coeff);
+ break;
+ case ADST_DCT:
+ cfg = &fwd_txfm_2d_cfg_adst_dct_8;
+ load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+ fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+ col_txfm_8x8_rounding(out, -cfg->shift[1]);
+ transpose_8x8(out);
+ fdct8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+ transpose_8x8(in);
+ write_buffer_8x8(in, coeff);
+ break;
+ case DCT_ADST:
+ cfg = &fwd_txfm_2d_cfg_dct_adst_8;
+ load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+ fdct8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+ col_txfm_8x8_rounding(out, -cfg->shift[1]);
+ transpose_8x8(out);
+ fadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+ transpose_8x8(in);
+ write_buffer_8x8(in, coeff);
+ break;
+ case ADST_ADST:
+ cfg = &fwd_txfm_2d_cfg_adst_adst_8;
+ load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+ fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+ col_txfm_8x8_rounding(out, -cfg->shift[1]);
+ transpose_8x8(out);
+ fadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+ transpose_8x8(in);
+ write_buffer_8x8(in, coeff);
+ break;
+ default:
+ assert(0);
+ }
+ (void)bd;
+}
diff --git a/vpx_dsp/x86/txfm_common_sse2.h b/vpx_dsp/x86/txfm_common_sse2.h
index 536b206..f886d30 100644
--- a/vpx_dsp/x86/txfm_common_sse2.h
+++ b/vpx_dsp/x86/txfm_common_sse2.h
@@ -26,4 +26,11 @@
_mm_setr_epi16((int16_t)(a), (int16_t)(b), (int16_t)(c), (int16_t)(d), \
(int16_t)(e), (int16_t)(f), (int16_t)(g), (int16_t)(h))
+// Reverse the 8 16 bit words in __m128i
+static INLINE __m128i mm_reverse_epi16(const __m128i x) {
+ const __m128i a = _mm_shufflelo_epi16(x, 0x1b);
+ const __m128i b = _mm_shufflehi_epi16(a, 0x1b);
+ return _mm_shuffle_epi32(b, 0x4e);
+}
+
#endif // VPX_DSP_X86_TXFM_COMMON_SSE2_H_
