ref: 6e44bf20f72aab0e3a930d25fd0745cfde9a149b
parent: 79fb3a013eb09c4b0ffae86188fd1fe0f701b261
author: James Zern <[email protected]>
date: Wed Jun 17 12:34:14 EDT 2015
vp9_reconintra_neon: add DC 4x4 predictors ~85-89% faster over 20M pixels Change-Id: I3812e8adfffe5255034da88dfe6546e12f4d10ee
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -208,9 +208,11 @@
#endif // HAVE_DSPR2
#if HAVE_NEON
-INTRA_PRED_TEST(NEON, TestIntraPred4, NULL, NULL, NULL, NULL,
- vp9_v_predictor_4x4_neon, vp9_h_predictor_4x4_neon, NULL, NULL,
- NULL, NULL, NULL, NULL, vp9_tm_predictor_4x4_neon)
+INTRA_PRED_TEST(NEON, TestIntraPred4, vp9_dc_predictor_4x4_neon,
+ vp9_dc_left_predictor_4x4_neon, vp9_dc_top_predictor_4x4_neon,
+ vp9_dc_128_predictor_4x4_neon, vp9_v_predictor_4x4_neon,
+ vp9_h_predictor_4x4_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+ vp9_tm_predictor_4x4_neon)
#endif // HAVE_NEON
#if HAVE_MSA
--- a/vp9/common/arm/neon/vp9_reconintra_neon.c
+++ b/vp9/common/arm/neon/vp9_reconintra_neon.c
@@ -15,6 +15,75 @@
#include "vpx/vpx_integer.h"
//------------------------------------------------------------------------------
+// DC 4x4
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride,
+ const uint8_t *above, const uint8_t *left,
+ int do_above, int do_left) {+ uint16x8_t sum_top;
+ uint16x8_t sum_left;
+ uint8x8_t dc0;
+
+ if (do_above) {+ const uint8x8_t A = vld1_u8(above); // top row
+ const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top
+ const uint16x4_t p1 = vpadd_u16(p0, p0);
+ sum_top = vcombine_u16(p1, p1);
+ }
+
+ if (do_left) {+ const uint8x8_t L = vld1_u8(left); // left border
+ const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left
+ const uint16x4_t p1 = vpadd_u16(p0, p0);
+ sum_left = vcombine_u16(p1, p1);
+ }
+
+ if (do_above && do_left) {+ const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+ dc0 = vrshrn_n_u16(sum, 3);
+ } else if (do_above) {+ dc0 = vrshrn_n_u16(sum_top, 2);
+ } else if (do_left) {+ dc0 = vrshrn_n_u16(sum_left, 2);
+ } else {+ dc0 = vdup_n_u8(0x80);
+ }
+
+ {+ const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+ int i;
+ for (i = 0; i < 4; ++i) {+ vst1_lane_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc), 0);
+ }
+ }
+}
+
+void vp9_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+ const uint8_t *above, const uint8_t *left) {+ dc_4x4(dst, stride, above, left, 1, 1);
+}
+
+void vp9_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+ const uint8_t *above, const uint8_t *left) {+ (void)above;
+ dc_4x4(dst, stride, NULL, left, 0, 1);
+}
+
+void vp9_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+ const uint8_t *above, const uint8_t *left) {+ (void)left;
+ dc_4x4(dst, stride, above, NULL, 1, 0);
+}
+
+void vp9_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+ const uint8_t *above, const uint8_t *left) {+ (void)above;
+ (void)left;
+ dc_4x4(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
// DC 8x8
// 'do_above' and 'do_left' facilitate branch removal when inlined.
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -84,16 +84,16 @@
specialize qw/vp9_tm_predictor_4x4 neon dspr2 msa/, "$sse_x86inc";
add_proto qw/void vp9_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_4x4 dspr2 msa/, "$sse_x86inc";
+specialize qw/vp9_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";
add_proto qw/void vp9_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_4x4 msa/, "$sse_x86inc";
+specialize qw/vp9_dc_top_predictor_4x4 msa neon/, "$sse_x86inc";
add_proto qw/void vp9_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_4x4 msa/, "$sse_x86inc";
+specialize qw/vp9_dc_left_predictor_4x4 msa neon/, "$sse_x86inc";
add_proto qw/void vp9_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_4x4 msa/, "$sse_x86inc";
+specialize qw/vp9_dc_128_predictor_4x4 msa neon/, "$sse_x86inc";
add_proto qw/void vp9_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
specialize qw/vp9_d207_predictor_8x8/, "$ssse3_x86inc";