Neon version of vp9_quantize_fp()

On a Nexus 7, vpxenc (in realtime mode, speed -12)
reported a performance improvement of ~12.4%

Change-Id: Id29d215acf58bb108489e218a259adf74b4768d7

3 files changed, 104 insertions, 1 deletions

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index d3d874dfc..3b1ca161d 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -715,7 +715,7 @@ add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, p
 specialize qw/vp9_subtract_block/, "$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/, "$ssse3_x86_64";
+specialize qw/vp9_quantize_fp neon/, "$ssse3_x86_64";
 
 add_proto qw/void vp9_quantize_fp_32x32/, "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_32x32/, "$ssse3_x86_64";
diff --git a/vp9/encoder/arm/neon/vp9_quantize_neon.c b/vp9/encoder/arm/neon/vp9_quantize_neon.c
new file mode 100644
index 000000000..2d5ec79b3
--- /dev/null
+++ b/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -0,0 +1,102 @@
+/*
+ *  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 <math.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          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) {
+  int i;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)zbin_oq_value;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+
+    for (i = 0; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_abs_coeff = vabsq_s16(v_coeff);
+      const int16x8_t v_tmp = vqaddq_s16(v_abs_coeff, v_round);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan =
+          vandq_s16(vmvnq_s16(vreinterpretq_s16_u16(v_nz_mask)), v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    {
+      const int16x4_t v_eobmax_3210 =
+          vmax_s16(vget_low_s16(v_eobmax_76543210),
+                   vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index 77b968bda..c8e7c5e6d 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -132,5 +132,6 @@ VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_avx2.c
 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-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))