Improve vp9_fdct4x4_sse2 (x1.2)

Modifications are done to reduce the total clock cycle.
Speedup: 1.2

Tested with: park_joy_420_720p50.y4m

Change-Id: Ia36b87e62e2f80a5fadaf5628729aedc80f38f3f

diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c
index fefca66..65431bd 100644 (file)
--- a/vp9/encoder/x86/vp9_dct_sse2.c
+++ b/vp9/encoder/x86/vp9_dct_sse2.c
@@ -26,24 +26,25 @@ void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) {
   //    by constructing the 32 bit constant corresponding to that pair.
   const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
   const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
   const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
   const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
   const __m128i kOne = _mm_set1_epi16(1);
-  __m128i in0, in1, in2, in3;
+  __m128i in0, in1;
   // Load inputs.
   {
     in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
-    in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
-    in2  = _mm_loadl_epi64((const __m128i *)(input +  2 * stride));
-    in3  = _mm_loadl_epi64((const __m128i *)(input +  3 * stride));
+    in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+           (input +  1 * stride)));
+    in1  = _mm_loadl_epi64((const __m128i *)(input +  2 * stride));
+    in1  = _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)
+           (input +  3 * stride)), in1);
+
     // x = x << 4
     in0 = _mm_slli_epi16(in0, 4);
     in1 = _mm_slli_epi16(in1, 4);
-    in2 = _mm_slli_epi16(in2, 4);
-    in3 = _mm_slli_epi16(in3, 4);
     // if (i == 0 && input[0]) input[0] += 1;
     {
       // The mask will only contain wether the first value is zero, all
@@ -60,18 +61,18 @@ void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) {
   // Do the two transform/transpose passes
   for (pass = 0; pass < 2; ++pass) {
     // Transform 1/2: Add/substract
-    const __m128i r0 = _mm_add_epi16(in0, in3);
-    const __m128i r1 = _mm_add_epi16(in1, in2);
-    const __m128i r2 = _mm_sub_epi16(in1, in2);
-    const __m128i r3 = _mm_sub_epi16(in0, in3);
+    const __m128i r0 = _mm_add_epi16(in0, in1);
+    const __m128i r1 = _mm_sub_epi16(in0, in1);
+    const __m128i r2 = _mm_unpacklo_epi64(r0, r1);
+    const __m128i r3 = _mm_unpackhi_epi64(r0, r1);
     // Transform 1/2: Interleave to do the multiply by constants which gets us
     //                into 32 bits.
-    const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
-    const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+    const __m128i t0 = _mm_unpacklo_epi16(r2, r3);
+    const __m128i t2 = _mm_unpackhi_epi16(r2, r3);
     const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
     const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
-    const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
-    const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+    const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p08_p24);
+    const __m128i u6 = _mm_madd_epi16(t2, k__cospi_p24_m08);
     const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
     const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
     const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
@@ -90,24 +91,21 @@ void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) {
     // 00 10 01 11 02 12 03 13
     // 20 30 21 31 22 32 23 33
     in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-    in2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+    in1 = _mm_shuffle_epi32(in1, 0x4E);
     // 00 10 20 30 01 11 21 31      in0 contains 0 followed by 1
-    // 02 12 22 32 03 13 23 33      in2 contains 2 followed by 3
-    if (0 == pass) {
-      // Extract values in the high part for second pass as transform code
-      // only uses the first four values.
-      in1 = _mm_unpackhi_epi64(in0, in0);
-      in3 = _mm_unpackhi_epi64(in2, in2);
-    } else {
-      // Post-condition output and store it (v + 1) >> 2, taking advantage
-      // of the fact 1/3 are stored just after 0/2.
-      __m128i out01 = _mm_add_epi16(in0, kOne);
-      __m128i out23 = _mm_add_epi16(in2, kOne);
-      out01 = _mm_srai_epi16(out01, 2);
-      out23 = _mm_srai_epi16(out23, 2);
-      _mm_storeu_si128((__m128i *)(output + 0 * 4), out01);
-      _mm_storeu_si128((__m128i *)(output + 2 * 4), out23);
-    }
+    // 02 12 22 32 03 13 23 33      in1 contains 2 followed by 3
+  }
+  in1 = _mm_shuffle_epi32(in1, 0x4E);
+  // Post-condition output and store it (v + 1) >> 2, taking advantage
+  // of the fact 1/3 are stored just after 0/2.
+  {
+     __m128i out01 = _mm_add_epi16(in0, kOne);
+     __m128i out23 = _mm_add_epi16(in1, kOne);
+     out01 = _mm_srai_epi16(out01, 2);
+     out23 = _mm_srai_epi16(out23, 2);
+     _mm_storeu_si128((__m128i *)(output + 0 * 4), out01);
+     _mm_storeu_si128((__m128i *)(output + 2 * 4), out23);
   }
 }