port libvpx sse2 quantize

libvpx got a lot of quantize improvements after libaom was forked.

For libaom we just need to remove skip_block and always assume the
large tran_low_t values.

Change-Id: Iaba4e5ae44146b77a5ee06f9f8af852457b71cf3
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index d4de33d..11ff737 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -183,6 +183,7 @@
               "${AOM_ROOT}/aom_dsp/x86/highbd_subtract_sse2.c"
               "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse2.c"
               "${AOM_ROOT}/aom_dsp/x86/quantize_sse2.c"
+              "${AOM_ROOT}/aom_dsp/x86/quantize_x86.h"
               "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.c"
               "${AOM_ROOT}/aom_dsp/x86/variance_sse2.c")
 
diff --git a/aom_dsp/x86/quantize_sse2.c b/aom_dsp/x86/quantize_sse2.c
index 6d86353..d3de6e2 100644
--- a/aom_dsp/x86/quantize_sse2.c
+++ b/aom_dsp/x86/quantize_sse2.c
@@ -9,37 +9,35 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 
+#include <assert.h>
 #include <emmintrin.h>
 #include <xmmintrin.h>
 
 #include "config/aom_dsp_rtcd.h"
 
 #include "aom/aom_integer.h"
+#include "aom_dsp/x86/quantize_x86.h"
 
 static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) {
-  if (sizeof(tran_low_t) == 4) {
-    return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1],
-                          (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3],
-                          (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5],
-                          (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]);
-  } else {
-    return _mm_load_si128((const __m128i *)coeff_ptr);
-  }
+  assert(sizeof(tran_low_t) == 4);
+
+  return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1],
+                        (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3],
+                        (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5],
+                        (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]);
 }
 
 static INLINE void store_coefficients(__m128i coeff_vals,
                                       tran_low_t *coeff_ptr) {
-  if (sizeof(tran_low_t) == 4) {
-    __m128i one = _mm_set1_epi16(1);
-    __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
-    __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
-    __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
-    __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
-    _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1);
-    _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2);
-  } else {
-    _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals);
-  }
+  assert(sizeof(tran_low_t) == 4);
+
+  __m128i one = _mm_set1_epi16(1);
+  __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
+  __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
+  __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
+  __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
+  _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1);
+  _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2);
 }
 
 void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
@@ -49,191 +47,101 @@
                          tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
                          uint16_t *eob_ptr, const int16_t *scan_ptr,
                          const int16_t *iscan_ptr) {
-  __m128i zero;
+  const __m128i zero = _mm_setzero_si128();
+  int index = 16;
+
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i eob, eob0;
+
   (void)scan_ptr;
 
-  coeff_ptr += n_coeffs;
-  iscan_ptr += n_coeffs;
-  qcoeff_ptr += n_coeffs;
-  dqcoeff_ptr += n_coeffs;
-  n_coeffs = -n_coeffs;
-  zero = _mm_setzero_si128();
-  __m128i eob;
-  __m128i zbin;
-  __m128i round, quant, dequant, shift;
-  {
-    __m128i coeff0, coeff1;
+  // Setup global values.
+  load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
+                dequant_ptr, &dequant, quant_shift_ptr, &shift);
 
-    // Setup global values
-    {
-      __m128i pw_1;
-      zbin = _mm_load_si128((const __m128i *)zbin_ptr);
-      round = _mm_load_si128((const __m128i *)round_ptr);
-      quant = _mm_load_si128((const __m128i *)quant_ptr);
-      pw_1 = _mm_set1_epi16(1);
-      zbin = _mm_sub_epi16(zbin, pw_1);
-      dequant = _mm_load_si128((const __m128i *)dequant_ptr);
-      shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
-    }
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
 
-    {
-      __m128i coeff0_sign, coeff1_sign;
-      __m128i qcoeff0, qcoeff1;
-      __m128i qtmp0, qtmp1;
-      __m128i cmp_mask0, cmp_mask1;
-      // Do DC and first 15 AC
-      coeff0 = load_coefficients(coeff_ptr + n_coeffs);
-      coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
+  // Poor man's abs().
+  coeff0_sign = _mm_srai_epi16(coeff0, 15);
+  coeff1_sign = _mm_srai_epi16(coeff1, 15);
+  qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
 
-      // Poor man's sign extract
-      coeff0_sign = _mm_srai_epi16(coeff0, 15);
-      coeff1_sign = _mm_srai_epi16(coeff1, 15);
-      qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
-      qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
-      qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
 
-      cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
-      zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
-      cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
-      qcoeff0 = _mm_adds_epi16(qcoeff0, round);
-      round = _mm_unpackhi_epi64(round, round);
-      qcoeff1 = _mm_adds_epi16(qcoeff1, round);
-      qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
-      quant = _mm_unpackhi_epi64(quant, quant);
-      qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
-      qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
-      qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
-      qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
-      shift = _mm_unpackhi_epi64(shift, shift);
-      qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+  calculate_qcoeff(&qcoeff0, round, quant, shift);
 
-      // Reinsert signs
-      qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
-      qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+  round = _mm_unpackhi_epi64(round, round);
+  quant = _mm_unpackhi_epi64(quant, quant);
+  shift = _mm_unpackhi_epi64(shift, shift);
 
-      // Mask out zbin threshold coeffs
-      qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
-      qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+  calculate_qcoeff(&qcoeff1, round, quant, shift);
 
-      store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
-      store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+  // Reinsert signs
+  qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
 
-      coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-      dequant = _mm_unpackhi_epi64(dequant, dequant);
-      coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+  // Mask out zbin threshold coeffs
+  qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+  qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
 
-      store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
-      store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
-    }
+  store_coefficients(qcoeff0, qcoeff_ptr);
+  store_coefficients(qcoeff1, qcoeff_ptr + 8);
 
-    {
-      // Scan for eob
-      __m128i zero_coeff0, zero_coeff1;
-      __m128i nzero_coeff0, nzero_coeff1;
-      __m128i iscan0, iscan1;
-      __m128i eob1;
-      zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
-      zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
-      nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
-      nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
-      iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
-      iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
-      // Add one to convert from indices to counts
-      iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
-      iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
-      eob = _mm_and_si128(iscan0, nzero_coeff0);
-      eob1 = _mm_and_si128(iscan1, nzero_coeff1);
-      eob = _mm_max_epi16(eob, eob1);
-    }
-    n_coeffs += 8 * 2;
+  coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+  dequant = _mm_unpackhi_epi64(dequant, dequant);
+  coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+  store_coefficients(coeff0, dqcoeff_ptr);
+  store_coefficients(coeff1, dqcoeff_ptr + 8);
+
+  eob =
+      scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, zero);
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    coeff0_sign = _mm_srai_epi16(coeff0, 15);
+    coeff1_sign = _mm_srai_epi16(coeff1, 15);
+    qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+    coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+    store_coefficients(coeff0, dqcoeff_ptr + index);
+    store_coefficients(coeff1, dqcoeff_ptr + index + 8);
+
+    eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr,
+                        index, zero);
+    eob = _mm_max_epi16(eob, eob0);
+
+    index += 16;
   }
 
-  // AC only loop
-  while (n_coeffs < 0) {
-    __m128i coeff0, coeff1;
-    {
-      __m128i coeff0_sign, coeff1_sign;
-      __m128i qcoeff0, qcoeff1;
-      __m128i qtmp0, qtmp1;
-      __m128i cmp_mask0, cmp_mask1;
-
-      coeff0 = load_coefficients(coeff_ptr + n_coeffs);
-      coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
-
-      // Poor man's sign extract
-      coeff0_sign = _mm_srai_epi16(coeff0, 15);
-      coeff1_sign = _mm_srai_epi16(coeff1, 15);
-      qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
-      qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
-      qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
-
-      cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
-      cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
-      qcoeff0 = _mm_adds_epi16(qcoeff0, round);
-      qcoeff1 = _mm_adds_epi16(qcoeff1, round);
-      qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
-      qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
-      qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
-      qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
-      qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
-      qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
-
-      // Reinsert signs
-      qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
-      qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-      qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
-
-      // Mask out zbin threshold coeffs
-      qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
-      qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
-
-      store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
-      store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
-
-      coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-      coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
-      store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
-      store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
-    }
-
-    {
-      // Scan for eob
-      __m128i zero_coeff0, zero_coeff1;
-      __m128i nzero_coeff0, nzero_coeff1;
-      __m128i iscan0, iscan1;
-      __m128i eob0, eob1;
-      zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
-      zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
-      nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
-      nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
-      iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
-      iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
-      // Add one to convert from indices to counts
-      iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
-      iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
-      eob0 = _mm_and_si128(iscan0, nzero_coeff0);
-      eob1 = _mm_and_si128(iscan1, nzero_coeff1);
-      eob0 = _mm_max_epi16(eob0, eob1);
-      eob = _mm_max_epi16(eob, eob0);
-    }
-    n_coeffs += 8 * 2;
-  }
-
-  // Accumulate EOB
-  {
-    __m128i eob_shuffled;
-    eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
-    eob = _mm_max_epi16(eob, eob_shuffled);
-    eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
-    eob = _mm_max_epi16(eob, eob_shuffled);
-    eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
-    eob = _mm_max_epi16(eob, eob_shuffled);
-    *eob_ptr = _mm_extract_epi16(eob, 1);
-  }
+  *eob_ptr = accumulate_eob(eob);
 }
diff --git a/aom_dsp/x86/quantize_x86.h b/aom_dsp/x86/quantize_x86.h
new file mode 100644
index 0000000..4eed7dd
--- /dev/null
+++ b/aom_dsp/x86/quantize_x86.h
@@ -0,0 +1,77 @@
+/*
+ *  Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+
+#include "aom/aom_integer.h"
+
+static INLINE void load_b_values(const int16_t *zbin_ptr, __m128i *zbin,
+                                 const int16_t *round_ptr, __m128i *round,
+                                 const int16_t *quant_ptr, __m128i *quant,
+                                 const int16_t *dequant_ptr, __m128i *dequant,
+                                 const int16_t *shift_ptr, __m128i *shift) {
+  *zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  *round = _mm_load_si128((const __m128i *)round_ptr);
+  *quant = _mm_load_si128((const __m128i *)quant_ptr);
+  *zbin = _mm_sub_epi16(*zbin, _mm_set1_epi16(1));
+  *dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  *shift = _mm_load_si128((const __m128i *)shift_ptr);
+}
+
+// With ssse3 and later abs() and sign() are preferred.
+static INLINE __m128i invert_sign_sse2(__m128i a, __m128i sign) {
+  a = _mm_xor_si128(a, sign);
+  return _mm_sub_epi16(a, sign);
+}
+
+static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round,
+                                    const __m128i quant, const __m128i shift) {
+  __m128i tmp, qcoeff;
+  qcoeff = _mm_adds_epi16(*coeff, round);
+  tmp = _mm_mulhi_epi16(qcoeff, quant);
+  qcoeff = _mm_add_epi16(tmp, qcoeff);
+  *coeff = _mm_mulhi_epi16(qcoeff, shift);
+}
+
+static INLINE __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) {
+  return _mm_mullo_epi16(qcoeff, dequant);
+}
+
+// Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing
+// to zbin to add 1 to the index in 'scan'.
+static INLINE __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1,
+                                   const __m128i zbin_mask0,
+                                   const __m128i zbin_mask1,
+                                   const int16_t *scan_ptr, const int index,
+                                   const __m128i zero) {
+  const __m128i zero_coeff0 = _mm_cmpeq_epi16(*coeff0, zero);
+  const __m128i zero_coeff1 = _mm_cmpeq_epi16(*coeff1, zero);
+  __m128i scan0 = _mm_load_si128((const __m128i *)(scan_ptr + index));
+  __m128i scan1 = _mm_load_si128((const __m128i *)(scan_ptr + index + 8));
+  __m128i eob0, eob1;
+  // Add one to convert from indices to counts
+  scan0 = _mm_sub_epi16(scan0, zbin_mask0);
+  scan1 = _mm_sub_epi16(scan1, zbin_mask1);
+  eob0 = _mm_andnot_si128(zero_coeff0, scan0);
+  eob1 = _mm_andnot_si128(zero_coeff1, scan1);
+  return _mm_max_epi16(eob0, eob1);
+}
+
+static INLINE int16_t accumulate_eob(__m128i eob) {
+  __m128i eob_shuffled;
+  eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  return _mm_extract_epi16(eob, 1);
+}