ref: 25e8514bde59b78d465966ef8ecc96b74a24f479
parent: ed5b3db6c5f6dd981c7dbcf4c71c508d8e858ba8
parent: f9dc411d89eed99d7def7de1e9dddba782c1212c
author: Luc Trudeau <luc@trud.ca>
date: Thu Jun 14 12:13:02 EDT 2018
Merge changes I51776f0e,I843f3b34 * changes: [VSX] Optimize PROCESS16 macro VSX Version of SAD8xN
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -1021,6 +1021,9 @@
SadMxNParam(16, 32, &vpx_sad16x32_vsx),
SadMxNParam(16, 16, &vpx_sad16x16_vsx),
SadMxNParam(16, 8, &vpx_sad16x8_vsx),
+ SadMxNParam(8, 16, &vpx_sad8x16_vsx),
+ SadMxNParam(8, 8, &vpx_sad8x8_vsx),
+ SadMxNParam(8, 4, &vpx_sad8x4_vsx),
};
INSTANTIATE_TEST_CASE_P(VSX, SADTest, ::testing::ValuesIn(vsx_tests));
--- a/vpx_dsp/ppc/sad_vsx.c
+++ b/vpx_dsp/ppc/sad_vsx.c
@@ -17,71 +17,75 @@
#include "vpx/vpx_integer.h"
#include "vpx_ports/mem.h"
-#define PROCESS16(offset) \
- v_a = vec_vsx_ld(offset, a); \
- v_b = vec_vsx_ld(offset, b); \
- v_ah = unpack_to_s16_h(v_a); \
- v_al = unpack_to_s16_l(v_a); \
- v_bh = unpack_to_s16_h(v_b); \
- v_bl = unpack_to_s16_l(v_b); \
- v_subh = vec_sub(v_ah, v_bh); \
- v_subl = vec_sub(v_al, v_bl); \
- v_absh = vec_abs(v_subh); \
- v_absl = vec_abs(v_subl); \
- v_sad = vec_sum4s(v_absh, v_sad); \
- v_sad = vec_sum4s(v_absl, v_sad);
+#define PROCESS16(offset) \
+ v_a = vec_vsx_ld(offset, a); \
+ v_b = vec_vsx_ld(offset, b); \
+ v_abs = vec_sub(vec_max(v_a, v_b), vec_min(v_a, v_b)); \
+ v_sad = vec_sum4s(v_abs, v_sad);
+#define SAD8(height) \
+ unsigned int vpx_sad8x##height##_vsx(const uint8_t *a, int a_stride, \
+ const uint8_t *b, int b_stride) { \+ int y = 0; \
+ uint8x16_t v_a, v_b, v_abs; \
+ uint32x4_t v_sad = vec_zeros_u32; \
+ \
+ do { \+ PROCESS16(0) \
+ \
+ a += a_stride; \
+ b += b_stride; \
+ y++; \
+ } while (y < height); \
+ \
+ return v_sad[1] + v_sad[0]; \
+ }
+
#define SAD16(height) \
unsigned int vpx_sad16x##height##_vsx(const uint8_t *a, int a_stride, \
const uint8_t *b, int b_stride) { \- int y; \
- unsigned int sad[4]; \
- uint8x16_t v_a, v_b; \
- int16x8_t v_ah, v_al, v_bh, v_bl, v_absh, v_absl, v_subh, v_subl; \
- int32x4_t v_sad = vec_splat_s32(0); \
+ int y = 0; \
+ uint8x16_t v_a, v_b, v_abs; \
+ uint32x4_t v_sad = vec_zeros_u32; \
\
- for (y = 0; y < height; y++) { \+ do { \PROCESS16(0); \
\
a += a_stride; \
b += b_stride; \
- } \
- vec_vsx_st((uint32x4_t)v_sad, 0, sad); \
+ y++; \
+ } while (y < height); \
\
- return sad[3] + sad[2] + sad[1] + sad[0]; \
+ return v_sad[3] + v_sad[2] + v_sad[1] + v_sad[0]; \
}
#define SAD32(height) \
unsigned int vpx_sad32x##height##_vsx(const uint8_t *a, int a_stride, \
const uint8_t *b, int b_stride) { \- int y; \
- unsigned int sad[4]; \
- uint8x16_t v_a, v_b; \
- int16x8_t v_ah, v_al, v_bh, v_bl, v_absh, v_absl, v_subh, v_subl; \
- int32x4_t v_sad = vec_splat_s32(0); \
+ int y = 0; \
+ uint8x16_t v_a, v_b, v_abs; \
+ uint32x4_t v_sad = vec_zeros_u32; \
\
- for (y = 0; y < height; y++) { \+ do { \PROCESS16(0); \
PROCESS16(16); \
\
a += a_stride; \
b += b_stride; \
- } \
- vec_vsx_st((uint32x4_t)v_sad, 0, sad); \
+ y++; \
+ } while (y < height); \
\
- return sad[3] + sad[2] + sad[1] + sad[0]; \
+ return v_sad[3] + v_sad[2] + v_sad[1] + v_sad[0]; \
}
#define SAD64(height) \
unsigned int vpx_sad64x##height##_vsx(const uint8_t *a, int a_stride, \
const uint8_t *b, int b_stride) { \- int y; \
- unsigned int sad[4]; \
- uint8x16_t v_a, v_b; \
- int16x8_t v_ah, v_al, v_bh, v_bl, v_absh, v_absl, v_subh, v_subl; \
- int32x4_t v_sad = vec_splat_s32(0); \
+ int y = 0; \
+ uint8x16_t v_a, v_b, v_abs; \
+ uint32x4_t v_sad = vec_zeros_u32; \
\
- for (y = 0; y < height; y++) { \+ do { \PROCESS16(0); \
PROCESS16(16); \
PROCESS16(32); \
@@ -89,12 +93,15 @@
\
a += a_stride; \
b += b_stride; \
- } \
- vec_vsx_st((uint32x4_t)v_sad, 0, sad); \
+ y++; \
+ } while (y < height); \
\
- return sad[3] + sad[2] + sad[1] + sad[0]; \
+ return v_sad[3] + v_sad[2] + v_sad[1] + v_sad[0]; \
}
+SAD8(4);
+SAD8(8);
+SAD8(16);
SAD16(8);
SAD16(16);
SAD16(32);
--- a/vpx_dsp/ppc/types_vsx.h
+++ b/vpx_dsp/ppc/types_vsx.h
@@ -82,6 +82,7 @@
static const uint16x8_t vec_ones_u16 = { 1, 1, 1, 1, 1, 1, 1, 1 }; static const uint32x4_t vec_ones_u32 = { 1, 1, 1, 1 }; static const int32x4_t vec_zeros_s32 = { 0, 0, 0, 0 };+static const uint32x4_t vec_zeros_u32 = { 0, 0, 0, 0 }; static const uint16x8_t vec_shift_sign_s16 = { 15, 15, 15, 15, 15, 15, 15, 15 }; static const uint32x4_t vec_shift_sign_s32 = { 31, 31, 31, 31 }; static const uint8x16_t vec_perm64 = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -748,13 +748,13 @@
specialize qw/vpx_sad16x8 neon msa sse2 vsx mmi/;
add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x16 neon msa sse2 mmi/;
+specialize qw/vpx_sad8x16 neon msa sse2 vsx mmi/;
add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x8 neon msa sse2 mmi/;
+specialize qw/vpx_sad8x8 neon msa sse2 vsx mmi/;
add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x4 neon msa sse2 mmi/;
+specialize qw/vpx_sad8x4 neon msa sse2 vsx mmi/;
add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
specialize qw/vpx_sad4x8 neon msa sse2 mmi/;