Add AVX2 intrinsic for idct16x16 and idct32x32 functions

Added AVX2 intrinsic optimization for the following functions
1. vpx_idct16x16_256_add
2. vpx_idct32x32_1024_add
3. vpx_idct32x32_135_add

The module level scaling w.r.t C function (timer based) for
existing (SSE2) and new AVX2 intrinsics:
                            Scaling
   Function Name         SSE2      AVX2
vpx_idct32x32_1024_add  3.62x     7.49x
vpx_idct32x32_135_add   4.85x     9.41x
vpx_idct16x16_256_add   4.82x     7.70x

This is a bit-exact change.

Change-Id: Id9dda933aa1f5093bb6b35ac3b8a41846afca9d2

3 files changed, 4 insertions, 2 deletions

diff --git a/vp9/common/vp9_idct.c b/vp9/common/vp9_idct.c
index 69069042c..71be0f310 100644
--- a/vp9/common/vp9_idct.c
+++ b/vp9/common/vp9_idct.c
@@ -150,6 +150,7 @@ void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
 
 void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
                        int eob) {
+  assert(((intptr_t)input) % 32 == 0);
   /* The calculation can be simplified if there are not many non-zero dct
    * coefficients. Use eobs to separate different cases. */
   if (eob == 1) /* DC only DCT coefficient. */
@@ -164,6 +165,7 @@ void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
 
 void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
                        int eob) {
+  assert(((intptr_t)input) % 32 == 0);
   if (eob == 1)
     vpx_idct32x32_1_add(input, dest, stride);
   else if (eob <= 34)
diff --git a/vp9/decoder/vp9_decoder.c b/vp9/decoder/vp9_decoder.c
index 7db8ed72d..92cd91f1e 100644
--- a/vp9/decoder/vp9_decoder.c
+++ b/vp9/decoder/vp9_decoder.c
@@ -87,7 +87,7 @@ void vp9_dec_alloc_row_mt_mem(RowMTWorkerData *row_mt_worker_data,
   row_mt_worker_data->num_sbs = num_sbs;
   for (plane = 0; plane < 3; ++plane) {
     CHECK_MEM_ERROR(cm, row_mt_worker_data->dqcoeff[plane],
-                    vpx_memalign(16, dqcoeff_size));
+                    vpx_memalign(32, dqcoeff_size));
     memset(row_mt_worker_data->dqcoeff[plane], 0, dqcoeff_size);
     CHECK_MEM_ERROR(cm, row_mt_worker_data->eob[plane],
                     vpx_calloc(num_sbs << EOBS_PER_SB_LOG2,
diff --git a/vp9/decoder/vp9_decoder.h b/vp9/decoder/vp9_decoder.h
index b0ef83c73..2e198d552 100644
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -54,7 +54,7 @@ typedef struct TileWorkerData {
   VP9LfSync *lf_sync;
   DECLARE_ALIGNED(16, MACROBLOCKD, xd);
   /* dqcoeff are shared by all the planes. So planes must be decoded serially */
-  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[32 * 32]);
+  DECLARE_ALIGNED(32, tran_low_t, dqcoeff[32 * 32]);
   DECLARE_ALIGNED(16, uint16_t, extend_and_predict_buf[80 * 2 * 80 * 2]);
   struct vpx_internal_error_info error_info;
 } TileWorkerData;