ref: 7b742da63e4b829ba013670dd838d263f5df8956
parent: 2057d3ef757a18e6bb005812a9912748ae4c7610
author: Johann <johannkoenig@google.com>
date: Tue May 2 03:31:05 EDT 2017
neon variance: process 4x blocks Continue processing sets of 16 values. Plenty of improvement for 4x8 (doubles the speed) but only about 30% for 4x4. BUG=webm:1422 Change-Id: Ib8dd96f75d474f0348800271d11e58356b620905
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1259,7 +1259,9 @@
VarianceParams(4, 3, &vpx_variance16x8_neon),
VarianceParams(3, 4, &vpx_variance8x16_neon),
VarianceParams(3, 3, &vpx_variance8x8_neon),
- VarianceParams(3, 2, &vpx_variance8x4_neon)));
+ VarianceParams(3, 2, &vpx_variance8x4_neon),
+ VarianceParams(2, 3, &vpx_variance4x8_neon),
+ VarianceParams(2, 2, &vpx_variance4x4_neon)));
INSTANTIATE_TEST_CASE_P(
NEON, VpxSubpelVarianceTest,
--- a/vpx_dsp/arm/mem_neon.h
+++ b/vpx_dsp/arm/mem_neon.h
@@ -79,6 +79,25 @@
memcpy(buf, &a, 4);
}
+// Load 4 sets of 4 bytes when alignment is not guaranteed.
+static INLINE uint8x16_t load_unaligned_u8q(const uint8_t *buf, int stride) {+ uint32_t a;
+ uint32x4_t a_u32 = vdupq_n_u32(0);
+ memcpy(&a, buf, 4);
+ buf += stride;
+ a_u32 = vld1q_lane_u32(&a, a_u32, 0);
+ memcpy(&a, buf, 4);
+ buf += stride;
+ a_u32 = vld1q_lane_u32(&a, a_u32, 1);
+ memcpy(&a, buf, 4);
+ buf += stride;
+ a_u32 = vld1q_lane_u32(&a, a_u32, 2);
+ memcpy(&a, buf, 4);
+ buf += stride;
+ a_u32 = vld1q_lane_u32(&a, a_u32, 3);
+ return vreinterpretq_u8_u32(a_u32);
+}
+
// Store 4 sets of 4 bytes when alignment is not guaranteed.
static INLINE void store_unaligned_u8q(uint8_t *buf, int stride,
const uint8x16_t a) {--- a/vpx_dsp/arm/variance_neon.c
+++ b/vpx_dsp/arm/variance_neon.c
@@ -14,6 +14,7 @@
#include "./vpx_config.h"
#include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_ports/mem.h"
static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {@@ -32,6 +33,47 @@
}
// w * h must be less than 2048 or sum_s16 may overflow.
+// Process a block of width 4 four rows at a time.
+static void variance_neon_w4x4(const uint8_t *a, int a_stride, const uint8_t *b,
+ int b_stride, int h, uint32_t *sse, int *sum) {+ int i;
+ int16x8_t sum_s16 = vdupq_n_s16(0);
+ int32x4_t sse_lo_s32 = vdupq_n_s32(0);
+ int32x4_t sse_hi_s32 = vdupq_n_s32(0);
+
+ for (i = 0; i < h; i += 4) {+ const uint8x16_t a_u8 = load_unaligned_u8q(a, a_stride);
+ const uint8x16_t b_u8 = load_unaligned_u8q(b, b_stride);
+ const uint16x8_t diff_lo_u16 =
+ vsubl_u8(vget_low_u8(a_u8), vget_low_u8(b_u8));
+ const uint16x8_t diff_hi_u16 =
+ vsubl_u8(vget_high_u8(a_u8), vget_high_u8(b_u8));
+
+ const int16x8_t diff_lo_s16 = vreinterpretq_s16_u16(diff_lo_u16);
+ const int16x8_t diff_hi_s16 = vreinterpretq_s16_u16(diff_hi_u16);
+
+ sum_s16 = vaddq_s16(sum_s16, diff_lo_s16);
+ sum_s16 = vaddq_s16(sum_s16, diff_hi_s16);
+
+ sse_lo_s32 = vmlal_s16(sse_lo_s32, vget_low_s16(diff_lo_s16),
+ vget_low_s16(diff_lo_s16));
+ sse_lo_s32 = vmlal_s16(sse_lo_s32, vget_high_s16(diff_lo_s16),
+ vget_high_s16(diff_lo_s16));
+
+ sse_hi_s32 = vmlal_s16(sse_hi_s32, vget_low_s16(diff_hi_s16),
+ vget_low_s16(diff_hi_s16));
+ sse_hi_s32 = vmlal_s16(sse_hi_s32, vget_high_s16(diff_hi_s16),
+ vget_high_s16(diff_hi_s16));
+
+ a += 4 * a_stride;
+ b += 4 * b_stride;
+ }
+
+ *sum = horizontal_add_s16x8(sum_s16);
+ *sse = (uint32_t)horizontal_add_s32x4(vaddq_s32(sse_lo_s32, sse_hi_s32));
+}
+
+// w * h must be less than 2048 or sum_s16 may overflow.
// Process a block of any size where the width is divisible by 16.
static void variance_neon_w16(const uint8_t *a, int a_stride, const uint8_t *b,
int b_stride, int w, int h, uint32_t *sse,
@@ -127,7 +169,9 @@
const uint8_t *b, int b_stride, \
unsigned int *sse) { \int sum; \
- if (n == 8) \
+ if (n == 4) \
+ variance_neon_w4x4(a, a_stride, b, b_stride, m, sse, &sum); \
+ else if (n == 8) \
variance_neon_w8x2(a, a_stride, b, b_stride, m, sse, &sum); \
else \
variance_neon_w16(a, a_stride, b, b_stride, n, m, sse, &sum); \
@@ -137,6 +181,8 @@
return *sse - (uint32_t)(((int64_t)sum * sum) >> shift); \
}
+varianceNxM(4, 4, 4);
+varianceNxM(4, 8, 5);
varianceNxM(8, 4, 5);
varianceNxM(8, 8, 6);
varianceNxM(8, 16, 7);
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1142,12 +1142,10 @@
specialize qw/vpx_variance8x4 sse2 neon msa/;
add_proto qw/unsigned int vpx_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-# TODO(johannkoenig): neon
- specialize qw/vpx_variance4x8 sse2 msa/;
+ specialize qw/vpx_variance4x8 sse2 neon msa/;
add_proto qw/unsigned int vpx_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-# TODO(johannkoenig): neon
- specialize qw/vpx_variance4x4 sse2 msa/;
+ specialize qw/vpx_variance4x4 sse2 neon msa/;
#
# Specialty Variance
--
⑨