Add vp9_sad32x32x4d_neon Neon intrinsic function.
On Nexus 7 speed -6 saw ~18% increase in perf.
Tested on Nexus 7, built with ndk r10d, gcc 4.9.
BUG=https://code.google.com/p/webm/issues/detail?id=908
Change-Id: I70ccdea0326750552ed946fb004507d6efe02d5c
diff --git a/test/sad_test.cc b/test/sad_test.cc
index 780f425..65e9561 100644
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -1245,9 +1245,11 @@
#if HAVE_NEON
const SadMxNx4Func sad_16x16x4d_neon = vp9_sad16x16x4d_neon;
+const SadMxNx4Func sad_32x32x4d_neon = vp9_sad32x32x4d_neon;
const SadMxNx4Func sad_64x64x4d_neon = vp9_sad64x64x4d_neon;
INSTANTIATE_TEST_CASE_P(NEON, SADx4Test, ::testing::Values(
make_tuple(16, 16, sad_16x16x4d_neon, -1),
+ make_tuple(32, 32, sad_32x32x4d_neon, -1),
make_tuple(64, 64, sad_64x64x4d_neon, -1)));
#endif // HAVE_NEON
#endif // CONFIG_VP9_ENCODER
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 271b55a..a1b15e8 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1064,7 +1064,7 @@
specialize qw/vp9_sad16x32x4d sse2/;
add_proto qw/void vp9_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad32x32x4d sse2 avx2/;
+specialize qw/vp9_sad32x32x4d sse2 avx2 neon/;
add_proto qw/void vp9_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad16x16x4d sse2 neon/;
diff --git a/vp9/encoder/arm/neon/vp9_sad4d_neon.c b/vp9/encoder/arm/neon/vp9_sad4d_neon.c
index e77852c..cec1689 100644
--- a/vp9/encoder/arm/neon/vp9_sad4d_neon.c
+++ b/vp9/encoder/arm/neon/vp9_sad4d_neon.c
@@ -60,6 +60,26 @@
vget_high_u8(vec_ref_48));
}
+// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16,
+// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi.
+static void sad_neon_32(const uint8x16_t vec_src_00,
+ const uint8x16_t vec_src_16,
+ const uint8_t *ref,
+ uint16x8_t *vec_sum_ref_lo,
+ uint16x8_t *vec_sum_ref_hi) {
+ const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+ const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+
+ *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+ vget_low_u8(vec_ref_00));
+ *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+ vget_high_u8(vec_ref_00));
+ *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+ vget_low_u8(vec_ref_16));
+ *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+ vget_high_u8(vec_ref_16));
+}
+
void vp9_sad64x64x4d_neon(const uint8_t *src, int src_stride,
const uint8_t* const ref[4], int ref_stride,
unsigned int *res) {
@@ -106,6 +126,50 @@
res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
}
+void vp9_sad32x32x4d_neon(const uint8_t *src, int src_stride,
+ const uint8_t* const ref[4], int ref_stride,
+ unsigned int *res) {
+ int i;
+ uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+ uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+ const uint8_t *ref0, *ref1, *ref2, *ref3;
+ ref0 = ref[0];
+ ref1 = ref[1];
+ ref2 = ref[2];
+ ref3 = ref[3];
+
+ for (i = 0; i < 32; ++i) {
+ const uint8x16_t vec_src_00 = vld1q_u8(src);
+ const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+
+ sad_neon_32(vec_src_00, vec_src_16, ref0,
+ &vec_sum_ref0_lo, &vec_sum_ref0_hi);
+ sad_neon_32(vec_src_00, vec_src_16, ref1,
+ &vec_sum_ref1_lo, &vec_sum_ref1_hi);
+ sad_neon_32(vec_src_00, vec_src_16, ref2,
+ &vec_sum_ref2_lo, &vec_sum_ref2_hi);
+ sad_neon_32(vec_src_00, vec_src_16, ref3,
+ &vec_sum_ref3_lo, &vec_sum_ref3_hi);
+
+ src += src_stride;
+ ref0 += ref_stride;
+ ref1 += ref_stride;
+ ref2 += ref_stride;
+ ref3 += ref_stride;
+ }
+
+ res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+ res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+ res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+ res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
void vp9_sad16x16x4d_neon(const uint8_t *src, int src_stride,
const uint8_t* const ref[4], int ref_stride,
unsigned int *res) {
