ref: a12ea1d5e90e460d68e38817978fe9d635fbd9bc
parent: 629279a45c1be8b5daa22b39e9c8f85503831e24
parent: cc7f0c0f3e168b31d4b5e5f42e0a368fa24ff62a
author: James Zern <jzern@google.com>
date: Fri May 12 09:33:50 EDT 2017
Merge "ppc: Add vpx_sad16x8/16/32_vsx"
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -920,4 +920,13 @@
INSTANTIATE_TEST_CASE_P(MSA, SADx4Test, ::testing::ValuesIn(x4d_msa_tests));
#endif // HAVE_MSA
+//------------------------------------------------------------------------------
+// VSX functions
+#if HAVE_VSX
+const SadMxNParam vsx_tests[] = {+ SadMxNParam(16, 32, &vpx_sad16x32_vsx),
+ SadMxNParam(16, 16, &vpx_sad16x16_vsx), SadMxNParam(16, 8, &vpx_sad16x8_vsx),
+};
+INSTANTIATE_TEST_CASE_P(VSX, SADTest, ::testing::ValuesIn(vsx_tests));
+#endif // HAVE_VSX
} // namespace
--- /dev/null
+++ b/vpx_dsp/ppc/sad_vsx.c
@@ -1,0 +1,53 @@
+/*
+ * Copyright (c) 2017 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "vpx_dsp/ppc/types_vsx.h"
+
+#include "vpx/vpx_integer.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 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); \
+ \
+ for (y = 0; y < height; y++) { \+ PROCESS16(0); \
+ \
+ a += a_stride; \
+ b += b_stride; \
+ } \
+ vec_vsx_st((uint32x4_t)v_sad, 0, sad); \
+ \
+ return sad[3] + sad[2] + sad[1] + sad[0]; \
+ }
+
+SAD16(8);
+SAD16(16);
+SAD16(32);
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -311,6 +311,8 @@
DSP_SRCS-$(HAVE_SSE2) += x86/sad_sse2.asm
DSP_SRCS-$(HAVE_SSE2) += x86/subtract_sse2.asm
+DSP_SRCS-$(HAVE_VSX) += ppc/sad_vsx.c
+
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad_sse2.asm
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -711,13 +711,13 @@
specialize qw/vpx_sad32x16 avx2 msa sse2/;
add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x32 msa sse2/;
+specialize qw/vpx_sad16x32 msa sse2 vsx/;
add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x16 neon msa sse2/;
+specialize qw/vpx_sad16x16 neon msa sse2 vsx/;
add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x8 neon msa sse2/;
+specialize qw/vpx_sad16x8 neon msa sse2 vsx/;
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/;