Neon version of vp9_subtract_block()

On a Nexus 7, vpxenc (in realtime mode, speed -12)
reported a performance improvement of ~3.2%

Change-Id: I8862497264142171b7efc32df1a67714a23539f4

test/vp9_subtract_test.cc

@@ -95,4 +95,9 @@ INSTANTIATE_TEST_CASE_P(C, VP9SubtractBlockTest,
 INSTANTIATE_TEST_CASE_P(SSE2, VP9SubtractBlockTest,
                         ::testing::Values(vp9_subtract_block_sse2));
 #endif
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, VP9SubtractBlockTest,
+                        ::testing::Values(vp9_subtract_block_neon));
+#endif
+
 }  // namespace vp9

vp9/common/vp9_rtcd_defs.pl

@@ -712,7 +712,7 @@ add_proto qw/int64_t vp9_block_error/, "const int16_t *coeff, const int16_t *dqc
 specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
 
 add_proto qw/void vp9_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/vp9_subtract_block/, "$sse2_x86inc";
+specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_quantize_fp/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 specialize qw/vp9_quantize_fp neon/, "$ssse3_x86_64";

vp9/encoder/arm/neon/vp9_subtract_neon.c

@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2014 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 <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+void vp9_subtract_block_neon(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, c;
+
+  if (cols > 16) {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; c += 32) {
+        const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
+        const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
+        const uint8x16_t v_pred_00 = vld1q_u8(&pred[c +  0]);
+        const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
+        const uint16x8_t v_diff_lo_00 = vsubl_u8(vget_low_u8(v_src_00),
+                                                 vget_low_u8(v_pred_00));
+        const uint16x8_t v_diff_hi_00 = vsubl_u8(vget_high_u8(v_src_00),
+                                                 vget_high_u8(v_pred_00));
+        const uint16x8_t v_diff_lo_16 = vsubl_u8(vget_low_u8(v_src_16),
+                                                 vget_low_u8(v_pred_16));
+        const uint16x8_t v_diff_hi_16 = vsubl_u8(vget_high_u8(v_src_16),
+                                                 vget_high_u8(v_pred_16));
+        vst1q_s16(&diff[c +  0], vreinterpretq_s16_u16(v_diff_lo_00));
+        vst1q_s16(&diff[c +  8], vreinterpretq_s16_u16(v_diff_hi_00));
+        vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
+        vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
+      }
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 8) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x16_t v_src = vld1q_u8(&src[0]);
+      const uint8x16_t v_pred = vld1q_u8(&pred[0]);
+      const uint16x8_t v_diff_lo = vsubl_u8(vget_low_u8(v_src),
+                                            vget_low_u8(v_pred));
+      const uint16x8_t v_diff_hi = vsubl_u8(vget_high_u8(v_src),
+                                            vget_high_u8(v_pred));
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
+      vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 4) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x8_t v_src = vld1_u8(&src[0]);
+      const uint8x8_t v_pred = vld1_u8(&pred[0]);
+      const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c)
+        diff[c] = src[c] - pred[c];
+
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  }
+}

vp9/vp9cx.mk

@@ -133,5 +133,6 @@ VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_sad_neon.c
 VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_dct_neon.c
 VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_variance_neon.c
 VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_quantize_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_subtract_neon.c
 
 VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))