ref: d468fd90e05ba7f5173d849c63f6a50115c9769b
parent: 034f94c127b4d608200d94889c52e7e42b2f62e4
author: Luca Barbato <lu_zero@gentoo.org>
date: Wed Jun 6 17:10:18 EDT 2018
Implement subtract_block for VSX ~2x speedup or better. [ RUN ] C/VP9SubtractBlockTest.Speed/0 [ BENCH ] 4x4 365.1 ms ( ±2.2 ms ) [ BENCH ] 8x4 258.5 ms ( ±0.3 ms ) [ BENCH ] 4x8 202.7 ms ( ±0.2 ms ) [ BENCH ] 8x8 162.2 ms ( ±0.5 ms ) [ BENCH ] 16x8 138.8 ms ( ±0.3 ms ) [ BENCH ] 8x16 121.5 ms ( ±0.4 ms ) [ BENCH ] 16x16 110.2 ms ( ±0.5 ms ) [ BENCH ] 32x16 104.8 ms ( ±0.1 ms ) [ BENCH ] 16x32 32.7 ms ( ±0.1 ms ) [ BENCH ] 32x32 30.0 ms ( ±0.0 ms ) [ BENCH ] 64x32 28.7 ms ( ±0.0 ms ) [ BENCH ] 32x64 20.1 ms ( ±0.0 ms ) [ BENCH ] 64x64 19.3 ms ( ±0.0 ms ) [ RUN ] VSX/VP9SubtractBlockTest.Speed/0 [ BENCH ] 4x4 155.3 ms ( ±0.9 ms ) [ BENCH ] 8x4 99.3 ms ( ±0.4 ms ) [ BENCH ] 4x8 77.2 ms ( ±0.1 ms ) [ BENCH ] 8x8 45.7 ms ( ±0.0 ms ) [ BENCH ] 16x8 34.1 ms ( ±0.0 ms ) [ BENCH ] 8x16 29.5 ms ( ±0.0 ms ) [ BENCH ] 16x16 19.9 ms ( ±0.0 ms ) [ BENCH ] 32x16 15.1 ms ( ±0.0 ms ) [ BENCH ] 16x32 16.7 ms ( ±0.0 ms ) [ BENCH ] 32x32 14.1 ms ( ±0.0 ms ) [ BENCH ] 64x32 12.6 ms ( ±0.0 ms ) [ BENCH ] 32x64 12.0 ms ( ±0.0 ms ) [ BENCH ] 64x64 11.2 ms ( ±0.0 ms ) Change-Id: I89ce12b6475871dc9e8fde84d0b6fe5c420c28c7
--- a/test/vp9_subtract_test.cc
+++ b/test/vp9_subtract_test.cc
@@ -145,4 +145,9 @@
::testing::Values(vpx_subtract_block_mmi));
#endif
+#if HAVE_VSX
+INSTANTIATE_TEST_CASE_P(VSX, VP9SubtractBlockTest,
+ ::testing::Values(vpx_subtract_block_vsx));
+#endif
+
} // namespace vp9
--- /dev/null
+++ b/vpx_dsp/ppc/subtract_vsx.c
@@ -1,0 +1,117 @@
+/*
+ * Copyright (c) 2018 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 <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/ppc/types_vsx.h"
+
+static VPX_FORCE_INLINE void subtract_block4x4(
+ int16_t *diff, ptrdiff_t diff_stride, const uint8_t *src,
+ ptrdiff_t src_stride, const uint8_t *pred, ptrdiff_t pred_stride) {+ int16_t *diff1 = diff + 2 * diff_stride;
+ const uint8_t *src1 = src + 2 * src_stride;
+ const uint8_t *pred1 = pred + 2 * pred_stride;
+
+ const int16x8_t d0 = vec_vsx_ld(0, diff);
+ const int16x8_t d1 = vec_vsx_ld(0, diff + diff_stride);
+ const int16x8_t d2 = vec_vsx_ld(0, diff1);
+ const int16x8_t d3 = vec_vsx_ld(0, diff1 + diff_stride);
+
+ const uint8x16_t s0 = read4x2(src, (int)src_stride);
+ const uint8x16_t p0 = read4x2(pred, (int)pred_stride);
+ const uint8x16_t s1 = read4x2(src1, (int)src_stride);
+ const uint8x16_t p1 = read4x2(pred1, (int)pred_stride);
+
+ const int16x8_t da = vec_sub(unpack_to_s16_h(s0), unpack_to_s16_h(p0));
+ const int16x8_t db = vec_sub(unpack_to_s16_h(s1), unpack_to_s16_h(p1));
+
+ vec_vsx_st(xxpermdi(da, d0, 1), 0, diff);
+ vec_vsx_st(xxpermdi(da, d1, 3), 0, diff + diff_stride);
+ vec_vsx_st(xxpermdi(db, d2, 1), 0, diff1);
+ vec_vsx_st(xxpermdi(db, d3, 3), 0, diff1 + diff_stride);
+}
+
+void vpx_subtract_block_vsx(int rows, int cols, int16_t *diff,
+ ptrdiff_t diff_stride, const uint8_t *src,
+ ptrdiff_t src_stride, const uint8_t *pred,
+ ptrdiff_t pred_stride) {+ int r = rows, c;
+
+ switch (cols) {+ case 64:
+ case 32:
+ do {+ for (c = 0; c < cols; c += 32) {+ const uint8x16_t s0 = vec_vsx_ld(0, src + c);
+ const uint8x16_t s1 = vec_vsx_ld(16, src + c);
+ const uint8x16_t p0 = vec_vsx_ld(0, pred + c);
+ const uint8x16_t p1 = vec_vsx_ld(16, pred + c);
+ const int16x8_t d0l =
+ vec_sub(unpack_to_s16_l(s0), unpack_to_s16_l(p0));
+ const int16x8_t d0h =
+ vec_sub(unpack_to_s16_h(s0), unpack_to_s16_h(p0));
+ const int16x8_t d1l =
+ vec_sub(unpack_to_s16_l(s1), unpack_to_s16_l(p1));
+ const int16x8_t d1h =
+ vec_sub(unpack_to_s16_h(s1), unpack_to_s16_h(p1));
+ vec_vsx_st(d0h, 0, diff + c);
+ vec_vsx_st(d0l, 16, diff + c);
+ vec_vsx_st(d1h, 0, diff + c + 16);
+ vec_vsx_st(d1l, 16, diff + c + 16);
+ }
+ diff += diff_stride;
+ pred += pred_stride;
+ src += src_stride;
+ } while (--r);
+ break;
+ case 16:
+ do {+ const uint8x16_t s0 = vec_vsx_ld(0, src);
+ const uint8x16_t p0 = vec_vsx_ld(0, pred);
+ const int16x8_t d0l = vec_sub(unpack_to_s16_l(s0), unpack_to_s16_l(p0));
+ const int16x8_t d0h = vec_sub(unpack_to_s16_h(s0), unpack_to_s16_h(p0));
+ vec_vsx_st(d0h, 0, diff);
+ vec_vsx_st(d0l, 16, diff);
+ diff += diff_stride;
+ pred += pred_stride;
+ src += src_stride;
+ } while (--r);
+ break;
+ case 8:
+ do {+ const uint8x16_t s0 = vec_vsx_ld(0, src);
+ const uint8x16_t p0 = vec_vsx_ld(0, pred);
+ const int16x8_t d0h = vec_sub(unpack_to_s16_h(s0), unpack_to_s16_h(p0));
+ vec_vsx_st(d0h, 0, diff);
+ diff += diff_stride;
+ pred += pred_stride;
+ src += src_stride;
+ } while (--r);
+ break;
+ case 4:
+ subtract_block4x4(diff, diff_stride, src, src_stride, pred, pred_stride);
+ if (r > 4) {+ diff += 4 * diff_stride;
+ pred += 4 * pred_stride;
+ src += 4 * src_stride;
+
+ subtract_block4x4(diff, diff_stride,
+
+ src, src_stride,
+
+ pred, pred_stride);
+ }
+ break;
+ default:
+ assert(0); // unreachable
+ }
+}
--- a/vpx_dsp/ppc/types_vsx.h
+++ b/vpx_dsp/ppc/types_vsx.h
@@ -68,6 +68,13 @@
#endif
#endif
+static INLINE uint8x16_t read4x2(const uint8_t *a, int stride) {+ const uint32x4_t a0 = (uint32x4_t)vec_vsx_ld(0, a);
+ const uint32x4_t a1 = (uint32x4_t)vec_vsx_ld(0, a + stride);
+
+ return (uint8x16_t)vec_mergeh(a0, a1);
+}
+
static const uint8x16_t vec_zeros_u8 = { 0, 0, 0, 0, 0, 0, 0, 0,0, 0, 0, 0, 0, 0, 0, 0 };
static const int16x8_t vec_zeros_s16 = { 0, 0, 0, 0, 0, 0, 0, 0 };--- a/vpx_dsp/ppc/variance_vsx.c
+++ b/vpx_dsp/ppc/variance_vsx.c
@@ -14,13 +14,6 @@
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/ppc/types_vsx.h"
-static INLINE uint8x16_t read4x2(const uint8_t *a, int stride) {- const uint32x4_t a0 = (uint32x4_t)vec_vsx_ld(0, a);
- const uint32x4_t a1 = (uint32x4_t)vec_vsx_ld(0, a + stride);
-
- return (uint8x16_t)vec_mergeh(a0, a1);
-}
-
uint32_t vpx_get4x4sse_cs_vsx(const uint8_t *a, int a_stride, const uint8_t *b,
int b_stride) {int distortion;
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -342,6 +342,7 @@
DSP_SRCS-$(HAVE_SSE2) += x86/subtract_sse2.asm
DSP_SRCS-$(HAVE_VSX) += ppc/sad_vsx.c
+DSP_SRCS-$(HAVE_VSX) += ppc/subtract_vsx.c
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -718,7 +718,7 @@
# Block subtraction
#
add_proto qw/void vpx_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-specialize qw/vpx_subtract_block neon msa mmi sse2/;
+specialize qw/vpx_subtract_block neon msa mmi sse2 vsx/;
#
# Single block SAD