quantize sse2: copy opts from ssse3

Simplify eob calculations based on ssse3 implementation.

General clean up and re-scoping.

Change-Id: I48f282bf9bd28ee9bc2c7a6779be9d45b5a3a3ee

diff --git a/vpx_dsp/x86/quantize_sse2.c b/vpx_dsp/x86/quantize_sse2.c
index 257377b..2635b27 100644 (file)
--- a/vpx_dsp/x86/quantize_sse2.c
+++ b/vpx_dsp/x86/quantize_sse2.c
@@ -23,185 +23,151 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
                          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;
-  __m128i eob;
-  __m128i zbin;
-  __m128i round, quant, dequant, shift;
+  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 qtmp0, qtmp1;
+  __m128i zero_coeff0, zero_coeff1, iscan0, iscan1;
+  __m128i eob, eob0, eob1;
+
   (void)scan_ptr;
   (void)skip_block;
   assert(!skip_block);
 
-  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 coeff0, coeff1;
-
-    // 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);
-    }
-
-    {
-      __m128i coeff0_sign, coeff1_sign;
-      __m128i qcoeff0, qcoeff1;
-      __m128i qtmp0, qtmp1;
-      __m128i cmp_mask0, cmp_mask1;
-      // Do DC and first 15 AC
-      coeff0 = load_tran_low(coeff_ptr + n_coeffs);
-      coeff1 = load_tran_low(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);
-      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);
-
-      // 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_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
-      store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
-
-      coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-      dequant = _mm_unpackhi_epi64(dequant, dequant);
-      coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
-      store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
-      store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 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;
-  }
-
-  // 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_tran_low(coeff_ptr + n_coeffs);
-      coeff1 = load_tran_low(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_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
-      store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
-
-      coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-      coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
-      store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
-      store_tran_low(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;
+  // Setup global values.
+  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 *)quant_shift_ptr);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_tran_low(coeff_ptr);
+  coeff1 = load_tran_low(coeff_ptr + 8);
+
+  // Poor man's abs().
+  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);
+
+  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);
+
+  // 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_tran_low(qcoeff0, qcoeff_ptr);
+  store_tran_low(qcoeff1, qcoeff_ptr + 8);
+
+  coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+  dequant = _mm_unpackhi_epi64(dequant, dequant);
+  coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+  store_tran_low(coeff0, dqcoeff_ptr);
+  store_tran_low(coeff1, dqcoeff_ptr + 8);
+
+  // Scan for eob.
+  zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+  zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+  iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
+  iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8));
+  // Add one to convert from indices to counts
+  iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
+  iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
+  eob = _mm_andnot_si128(zero_coeff0, iscan0);
+  eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
+  eob = _mm_max_epi16(eob, eob1);
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_tran_low(coeff_ptr + index);
+    coeff1 = load_tran_low(coeff_ptr + index + 8);
+
+    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);
+
+    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);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_tran_low(qcoeff0, qcoeff_ptr + index);
+    store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
+
+    coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+    coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+    store_tran_low(coeff0, dqcoeff_ptr + index);
+    store_tran_low(coeff1, dqcoeff_ptr + index + 8);
+
+    zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+    zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+    iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index));
+    iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
+    iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
+    iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
+    eob0 = _mm_andnot_si128(zero_coeff0, iscan0);
+    eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
+    eob0 = _mm_max_epi16(eob0, eob1);
+    eob = _mm_max_epi16(eob, eob0);
+
+    index += 16;
   }
 
-  // Accumulate EOB
+  // Accumulate eob.
   {
     __m128i eob_shuffled;
     eob_shuffled = _mm_shuffle_epi32(eob, 0xe);