SSE2 4x4 invserse ADST/DCT transform

Enable SSE2 4x4 inverse ADST/DCT transform. The runtime goes from
292 cycles down to 89 cycles. Running bus_cif at 2000 kbps, the
overall runtime of speed 0 goes from 301s to 295s (2% speed-up).

Change-Id: I24098136e7fee7ab2fbf1c11755bdf2ca37f3628

diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 459336e4ddd78d5ce0c83528b21de675e16a11ee..c3800d827baddfcaac02296e6eaded7edfa2daf3 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -325,7 +325,7 @@ prototype void vp9_short_idct10_32x32_add "int16_t *input, uint8_t *dest, int de
 specialize vp9_short_idct10_32x32_add
 
 prototype void vp9_short_iht4x4_add "int16_t *input, uint8_t *dest, int dest_stride, int tx_type"
-specialize vp9_short_iht4x4_add
+specialize vp9_short_iht4x4_add sse2
 
 prototype void vp9_short_iht8x8_add "int16_t *input, uint8_t *dest, int dest_stride, int tx_type"
 specialize vp9_short_iht8x8_add

diff --git a/vp9/common/x86/vp9_idct_intrin_sse2.c b/vp9/common/x86/vp9_idct_intrin_sse2.c
index 599dcff93ea7c2d06e14949684fe61f1b0998df0..4495b15eda526cf843e559d03a9344198ddbc84f 100644 (file)
--- a/vp9/common/x86/vp9_idct_intrin_sse2.c
+++ b/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -241,6 +241,155 @@ void vp9_idct4_1d_sse2(int16_t *input, int16_t *output) {
   _mm_storel_epi64((__m128i *)output, in);
 }
 
+static INLINE void transpose_4x4(__m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(res[2], res[3]);
+  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
+
+  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
+  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
+}
+
+void idct4_1d_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+
+  transpose_4x4(in);
+  // stage 1
+  u[0] = _mm_unpacklo_epi16(in[0], in[2]);
+  u[1] = _mm_unpacklo_epi16(in[1], in[3]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[2]);
+  u[1] = _mm_packs_epi32(v[1], v[3]);
+  u[2] = _mm_unpackhi_epi64(u[0], u[0]);
+  u[3] = _mm_unpackhi_epi64(u[1], u[1]);
+
+  // stage 2
+  in[0] = _mm_add_epi16(u[0], u[3]);
+  in[1] = _mm_add_epi16(u[1], u[2]);
+  in[2] = _mm_sub_epi16(u[1], u[2]);
+  in[3] = _mm_sub_epi16(u[0], u[3]);
+}
+
+void iadst4_1d_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
+  const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16(sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8], in7;
+
+  transpose_4x4(in);
+  in7 = _mm_add_epi16(in[0], in[3]);
+  in7 = _mm_sub_epi16(in7, in[2]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[2]);
+  u[1] = _mm_unpacklo_epi16(in[1], in[3]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpacklo_epi16(in[1], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(v[3], v[4]);
+  u[2] = v[2];
+  u[3] = _mm_add_epi32(u[0], u[1]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_add_epi32(u[3], v[5]);
+  u[6] = _mm_sub_epi32(u[5], u[4]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[2]);
+  in[1] = _mm_packs_epi32(u[1], u[3]);
+  in[2] = _mm_unpackhi_epi64(in[0], in[0]);
+  in[3] = _mm_unpackhi_epi64(in[1], in[1]);
+}
+
+void vp9_short_iht4x4_add_sse2(int16_t *input, uint8_t *dest, int stride,
+                               int tx_type) {
+  __m128i in[4];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+
+  in[0] = _mm_loadl_epi64((__m128i *)input);
+  in[1] = _mm_loadl_epi64((__m128i *)(input + 4));
+  in[2] = _mm_loadl_epi64((__m128i *)(input + 8));
+  in[3] = _mm_loadl_epi64((__m128i *)(input + 12));
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct4_1d_sse2(in);
+      idct4_1d_sse2(in);
+      break;
+    case 1:  // ADST_DCT
+      idct4_1d_sse2(in);
+      iadst4_1d_sse2(in);
+      break;
+    case 2:  // DCT_ADST
+      iadst4_1d_sse2(in);
+      idct4_1d_sse2(in);
+      break;
+    case 3:  // ADST_ADST
+      iadst4_1d_sse2(in);
+      iadst4_1d_sse2(in);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  // Final round and shift
+  in[0] = _mm_add_epi16(in[0], eight);
+  in[1] = _mm_add_epi16(in[1], eight);
+  in[2] = _mm_add_epi16(in[2], eight);
+  in[3] = _mm_add_epi16(in[3], eight);
+
+  in[0] = _mm_srai_epi16(in[0], 4);
+  in[1] = _mm_srai_epi16(in[1], 4);
+  in[2] = _mm_srai_epi16(in[2], 4);
+  in[3] = _mm_srai_epi16(in[3], 4);
+
+  RECON_AND_STORE4X4(dest, in[0]);
+  RECON_AND_STORE4X4(dest, in[1]);
+  RECON_AND_STORE4X4(dest, in[2]);
+  RECON_AND_STORE4X4(dest, in[3]);
+}
+
 #define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \
                       out0, out1, out2, out3, out4, out5, out6, out7) \
   {                                                     \