Improved vp9_ihtllm_c

As suggested by Yaowu, we can use eob to reduce the complexity
of the vp9_ihtllm_c function.  For the 1080p test clip used, the decoder
performance improved by 17%.

Change-Id: I32486f2f06f9b8f60467d2a574209aa3a3daa435

3 files changed, 25 insertions, 9 deletions

diff --git a/vp9/common/vp9_idctllm.c b/vp9/common/vp9_idctllm.c
index 9622dfdee..897514ee1 100644
--- a/vp9/common/vp9_idctllm.c
+++ b/vp9/common/vp9_idctllm.c
@@ -404,8 +404,9 @@ void vp9_ihtllm_float_c(const int16_t *input, int16_t *output, int pitch,
 #define HORIZONTAL_SHIFT 17  // 15
 #define HORIZONTAL_ROUNDING ((1 << (HORIZONTAL_SHIFT - 1)) - 1)
 void vp9_ihtllm_c(const int16_t *input, int16_t *output, int pitch,
-                      TX_TYPE tx_type, int tx_dim) {
+                      TX_TYPE tx_type, int tx_dim, uint16_t eobs) {
   int i, j, k;
+  int nz_dim;
   int16_t imbuf[256];
 
   const int16_t *ip = input;
@@ -444,12 +445,25 @@ void vp9_ihtllm_c(const int16_t *input, int16_t *output, int pitch,
       break;
   }
 
+  nz_dim = tx_dim;
+  if(tx_dim > 4) {
+    if(eobs < 36) {
+      vpx_memset(im, 0, 512);
+      nz_dim = 8;
+      if(eobs < 3) {
+        nz_dim = 2;
+      } else if(eobs < 10) {
+        nz_dim = 4;
+      }
+    }
+  }
+
   /* vertical transformation */
   for (j = 0; j < tx_dim; j++) {
-    for (i = 0; i < tx_dim; i++) {
+    for (i = 0; i < nz_dim; i++) {
       int temp = 0;
 
-      for (k = 0; k < tx_dim; k++) {
+      for (k = 0; k < nz_dim; k++) {
         temp += ptv[k] * ip[(k * tx_dim)];
       }
 
@@ -470,7 +484,7 @@ void vp9_ihtllm_c(const int16_t *input, int16_t *output, int pitch,
     for (i = 0; i < tx_dim; i++) {
       int temp = 0;
 
-      for (k = 0; k < tx_dim; k++) {
+      for (k = 0; k < nz_dim; k++) {
         temp += im[k] * pthc[k];
       }
 
diff --git a/vp9/common/vp9_invtrans.c b/vp9/common/vp9_invtrans.c
index c78f1ad3c..eff919865 100644
--- a/vp9/common/vp9_invtrans.c
+++ b/vp9/common/vp9_invtrans.c
@@ -52,7 +52,7 @@ void vp9_inverse_transform_mby_4x4(MACROBLOCKD *xd) {
     TX_TYPE tx_type = get_tx_type_4x4(xd, &xd->block[i]);
     if (tx_type != DCT_DCT) {
       vp9_ihtllm(xd->block[i].dqcoeff, xd->block[i].diff, 32,
-                   tx_type, 4);
+                   tx_type, 4, xd->block[i].eob);
     } else {
       vp9_inverse_transform_b_4x4(xd, i, 32);
     }
@@ -91,7 +91,8 @@ void vp9_inverse_transform_mby_8x8(MACROBLOCKD *xd) {
   for (i = 0; i < 9; i += 8) {
     TX_TYPE tx_type = get_tx_type_8x8(xd, &xd->block[i]);
     if (tx_type != DCT_DCT) {
-      vp9_ihtllm(xd->block[i].dqcoeff, xd->block[i].diff, 32, tx_type, 8);
+      vp9_ihtllm(xd->block[i].dqcoeff, xd->block[i].diff, 32, tx_type, 8,
+                 xd->block[i].eob);
     } else {
       vp9_inverse_transform_b_8x8(&blockd[i].dqcoeff[0],
                                   &blockd[i].diff[0], 32);
@@ -100,7 +101,8 @@ void vp9_inverse_transform_mby_8x8(MACROBLOCKD *xd) {
   for (i = 2; i < 11; i += 8) {
     TX_TYPE tx_type = get_tx_type_8x8(xd, &xd->block[i]);
     if (tx_type != DCT_DCT) {
-      vp9_ihtllm(xd->block[i + 2].dqcoeff, xd->block[i].diff, 32, tx_type, 8);
+      vp9_ihtllm(xd->block[i + 2].dqcoeff, xd->block[i].diff, 32, tx_type, 8,
+                 xd->block[i + 2].eob);
     } else {
       vp9_inverse_transform_b_8x8(&blockd[i + 2].dqcoeff[0],
                                   &blockd[i].diff[0], 32);
@@ -132,7 +134,7 @@ void vp9_inverse_transform_mby_16x16(MACROBLOCKD *xd) {
   BLOCKD *bd = &xd->block[0];
   TX_TYPE tx_type = get_tx_type_16x16(xd, bd);
   if (tx_type != DCT_DCT) {
-    vp9_ihtllm(bd->dqcoeff, bd->diff, 32, tx_type, 16);
+    vp9_ihtllm(bd->dqcoeff, bd->diff, 32, tx_type, 16, bd->eob);
   } else {
     vp9_inverse_transform_b_16x16(&xd->block[0].dqcoeff[0],
                                   &xd->block[0].diff[0], 32);
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 5b7af100b..e8981ce5e 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -361,7 +361,7 @@ specialize vp9_short_idct16x16
 prototype void vp9_short_idct10_16x16 "short *input, short *output, int pitch"
 specialize vp9_short_idct10_16x16
 
-prototype void vp9_ihtllm "const short *input, short *output, int pitch, int tx_type, int tx_dim"
+prototype void vp9_ihtllm "const short *input, short *output, int pitch, int tx_type, int tx_dim, short eobs"
 specialize vp9_ihtllm
 
 #