Fix the filter tap calculation in mips optimizations

The interp filter tap calculation was not accurate to tell the
difference between 2 taps and 4 taps. This patch fixed the bug, and
resolved Jenkins test failures in mips sub-pel filter optimizations.

BUG=webm:1568

Change-Id: I51eb8adb7ed194ef2ea7dd4aa57aa9870ee38cfc

diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index a45db8eba29c37ed36fc85349a9f57124f95a9a8..c4b3922e20f708a3d7672c86a349c65effc12ffd 100644 (file)
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -789,13 +789,7 @@ TEST_P(ConvolveTest, Copy2D) {
   }
 }
 
-#if HAVE_MSA
-// TODO(any) MSA optimizations doesn't work with 4-tap interp filter. Need to be
-// fixed.
-const int kNumFilterBanks = 4;
-#else
 const int kNumFilterBanks = 5;
-#endif
 const int kNumFilters = 16;
 
 TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) {

diff --git a/vpx_dsp/mips/convolve8_avg_dspr2.c b/vpx_dsp/mips/convolve8_avg_dspr2.c
index d9c2bef69ed0c27d4a25ed7b334ad241f1d850ef..cc458c86182daf03526a7593ae5a6c965ccaf964 100644 (file)
--- a/vpx_dsp/mips/convolve8_avg_dspr2.c
+++ b/vpx_dsp/mips/convolve8_avg_dspr2.c
@@ -15,6 +15,7 @@
 #include "vpx_dsp/mips/convolve_common_dspr2.h"
 #include "vpx_dsp/vpx_convolve.h"
 #include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
 #include "vpx_ports/mem.h"
 
 #if HAVE_DSPR2
@@ -341,7 +342,7 @@ void vpx_convolve8_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
   assert(y_step_q4 == 16);
   assert(((const int32_t *)filter_y)[1] != 0x800000);
 
-  if (((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_y) == 2) {
     vpx_convolve2_avg_vert_dspr2(src, src_stride, dst, dst_stride, filter,
                                  x0_q4, x_step_q4, y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c b/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
index fb68ad881398bea4bacb42bc2ef29fdd887297e0..7a9aa49d8a15a04b1d6cf19a751f134ec7b50cbc 100644 (file)
--- a/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
+++ b/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
@@ -15,6 +15,7 @@
 #include "vpx_dsp/mips/convolve_common_dspr2.h"
 #include "vpx_dsp/vpx_convolve.h"
 #include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
 #include "vpx_ports/mem.h"
 
 #if HAVE_DSPR2
@@ -945,7 +946,7 @@ void vpx_convolve8_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
   assert(x_step_q4 == 16);
   assert(((const int32_t *)filter_x)[1] != 0x800000);
 
-  if (((const int32_t *)filter_x)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2) {
     vpx_convolve2_avg_horiz_dspr2(src, src_stride, dst, dst_stride, filter,
                                   x0_q4, x_step_q4, y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/convolve8_dspr2.c b/vpx_dsp/mips/convolve8_dspr2.c
index 89f0f41962afe70fda299efc83f7620722f20cb8..1e7052f6c58c22315390e7c2ecb8c0fafe13f805 100644 (file)
--- a/vpx_dsp/mips/convolve8_dspr2.c
+++ b/vpx_dsp/mips/convolve8_dspr2.c
@@ -1322,7 +1322,7 @@ void vpx_convolve8_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
   if (filter_x[3] == 0x80) {
     copy_horiz_transposed(src - src_stride * 3, src_stride, temp,
                           intermediate_height, w, intermediate_height);
-  } else if (((const int32_t *)filter_x)[0] == 0) {
+  } else if (vpx_get_filter_taps(filter_x) == 2) {
     vpx_convolve2_dspr2(src - src_stride * 3, src_stride, temp,
                         intermediate_height, filter_x, w, intermediate_height);
   } else {
@@ -1365,7 +1365,7 @@ void vpx_convolve8_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
   /* copy the src to dst */
   if (filter_y[3] == 0x80) {
     copy_horiz_transposed(temp + 3, intermediate_height, dst, dst_stride, h, w);
-  } else if (((const int32_t *)filter_y)[0] == 0) {
+  } else if (vpx_get_filter_taps(filter_y) == 2) {
     vpx_convolve2_dspr2(temp + 3, intermediate_height, dst, dst_stride,
                         filter_y, h, w);
   } else {

diff --git a/vpx_dsp/mips/convolve8_horiz_dspr2.c b/vpx_dsp/mips/convolve8_horiz_dspr2.c
index 77e95c844491aeb1de44c7750689e37a2a118daf..09d6f36e56150b372b05aa1e34bd85ad59a0e630 100644 (file)
--- a/vpx_dsp/mips/convolve8_horiz_dspr2.c
+++ b/vpx_dsp/mips/convolve8_horiz_dspr2.c
@@ -825,7 +825,7 @@ void vpx_convolve8_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
   assert(x_step_q4 == 16);
   assert(((const int32_t *)filter_x)[1] != 0x800000);
 
-  if (((const int32_t *)filter_x)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2) {
     vpx_convolve2_horiz_dspr2(src, src_stride, dst, dst_stride, filter, x0_q4,
                               x_step_q4, y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/convolve8_vert_dspr2.c b/vpx_dsp/mips/convolve8_vert_dspr2.c
index c329f71ccf6d80dbcca95a2e609c12ee38dc519f..fd977b53360f0016d17980d4587f2c8f23c25eb2 100644 (file)
--- a/vpx_dsp/mips/convolve8_vert_dspr2.c
+++ b/vpx_dsp/mips/convolve8_vert_dspr2.c
@@ -325,7 +325,7 @@ void vpx_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
   assert(y_step_q4 == 16);
   assert(((const int32_t *)filter_y)[1] != 0x800000);
 
-  if (((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_y) == 2) {
     vpx_convolve2_vert_dspr2(src, src_stride, dst, dst_stride, filter, x0_q4,
                              x_step_q4, y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c b/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
index 187a013421a6ef4a0f2d8b8b5a040296f87f8af7..5b5a1cbc3a5136a9f834756b57e274c3a1f25861 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
@@ -658,7 +658,7 @@ void vpx_convolve8_avg_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
     filt_hor[cnt] = filter_x[cnt];
   }
 
-  if (((const int32_t *)filter_x)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2) {
     switch (w) {
       case 4:
         common_hz_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,

diff --git a/vpx_dsp/mips/vpx_convolve8_avg_msa.c b/vpx_dsp/mips/vpx_convolve8_avg_msa.c
index 5187cea21c98401ae617a2268c9ddec1a6e94fe3..ba816192a1faade82a22f05e4d3c7b8b47c538c4 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_avg_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_avg_msa.c
@@ -538,8 +538,8 @@ void vpx_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
     filt_ver[cnt] = filter_y[cnt];
   }
 
-  if (((const int32_t *)filter_x)[0] == 0 &&
-      ((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2 &&
+      vpx_get_filter_taps(filter_y) == 2) {
     switch (w) {
       case 4:
         common_hv_2ht_2vt_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
@@ -571,8 +571,8 @@ void vpx_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
                             x_step_q4, y0_q4, y_step_q4, w, h);
         break;
     }
-  } else if (((const int32_t *)filter_x)[0] == 0 ||
-             ((const int32_t *)filter_y)[0] == 0) {
+  } else if (vpx_get_filter_taps(filter_x) == 2 ||
+             vpx_get_filter_taps(filter_y) == 2) {
     vpx_convolve8_avg_c(src, src_stride, dst, dst_stride, filter, x0_q4,
                         x_step_q4, y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c b/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
index ef8c90114060204ac0d9f9c565b5121ddbb83209..e6a790dfc6d19db30448e4c5c713b5b6f1096b11 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
@@ -625,7 +625,7 @@ void vpx_convolve8_avg_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
     filt_ver[cnt] = filter_y[cnt];
   }
 
-  if (((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_y) == 2) {
     switch (w) {
       case 4:
         common_vt_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,

diff --git a/vpx_dsp/mips/vpx_convolve8_horiz_msa.c b/vpx_dsp/mips/vpx_convolve8_horiz_msa.c
index 152dc26104e976b3305429aaeb79db2ff1870013..792c0f709c4ca84fde3c6b4c1b0f4b501177e8ad 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_horiz_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_horiz_msa.c
@@ -634,7 +634,7 @@ void vpx_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
     filt_hor[cnt] = filter_x[cnt];
   }
 
-  if (((const int32_t *)filter_x)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2) {
     switch (w) {
       case 4:
         common_hz_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,

diff --git a/vpx_dsp/mips/vpx_convolve8_msa.c b/vpx_dsp/mips/vpx_convolve8_msa.c
index d35a5a7a639a313fe76c3c1389b3bc754b146d3b..c942167587bf8963df86458f8e8f619dd6f2865e 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_msa.c
@@ -558,8 +558,8 @@ void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
     filt_ver[cnt] = filter_y[cnt];
   }
 
-  if (((const int32_t *)filter_x)[0] == 0 &&
-      ((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_x) == 2 &&
+      vpx_get_filter_taps(filter_y) == 2) {
     switch (w) {
       case 4:
         common_hv_2ht_2vt_4w_msa(src, (int32_t)src_stride, dst,
@@ -591,8 +591,8 @@ void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
                         x_step_q4, y0_q4, y_step_q4, w, h);
         break;
     }
-  } else if (((const int32_t *)filter_x)[0] == 0 ||
-             ((const int32_t *)filter_y)[0] == 0) {
+  } else if (vpx_get_filter_taps(filter_x) == 2 ||
+             vpx_get_filter_taps(filter_y) == 2) {
     vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, x_step_q4,
                     y0_q4, y_step_q4, w, h);
   } else {

diff --git a/vpx_dsp/mips/vpx_convolve8_vert_msa.c b/vpx_dsp/mips/vpx_convolve8_vert_msa.c
index 13fce0077c9b91e0f0ac0ab29bcc9f3ed6165e21..195228689e0bfdbc583d05be3f647a9762955fce 100644 (file)
--- a/vpx_dsp/mips/vpx_convolve8_vert_msa.c
+++ b/vpx_dsp/mips/vpx_convolve8_vert_msa.c
@@ -641,7 +641,7 @@ void vpx_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
     filt_ver[cnt] = filter_y[cnt];
   }
 
-  if (((const int32_t *)filter_y)[0] == 0) {
+  if (vpx_get_filter_taps(filter_y) == 2) {
     switch (w) {
       case 4:
         common_vt_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,

diff --git a/vpx_dsp/vpx_filter.h b/vpx_dsp/vpx_filter.h
index 05eb572651be6859c7f1a5b37270b1a826bba1d5..54357ee6caec3051113227eb6d8900ff1210b0ee 100644 (file)
--- a/vpx_dsp/vpx_filter.h
+++ b/vpx_dsp/vpx_filter.h
@@ -11,6 +11,7 @@
 #ifndef VPX_VPX_DSP_VPX_FILTER_H_
 #define VPX_VPX_DSP_VPX_FILTER_H_
 
+#include <assert.h>
 #include "vpx/vpx_integer.h"
 
 #ifdef __cplusplus
@@ -26,6 +27,14 @@ extern "C" {
 
 typedef int16_t InterpKernel[SUBPEL_TAPS];
 
+static INLINE int vpx_get_filter_taps(const int16_t *const filter) {
+  assert(filter[3] != 128);
+  if (!filter[0] && !filter[1] && !filter[2])
+    return 2;
+  else
+    return 8;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif