Optimize quantization simd implementation

This commit allows the quantizer to compare the AC coefficients to
the quantization step size to determine if further multiplication
operations are needed. It makes the quantization process 20% faster
without coding statistics change.

Change-Id: I735aaf6a9c0874c82175bb565b20e131464db64a

2 files changed, 61 insertions, 33 deletions

diff --git a/vp9/encoder/x86/vp9_dct_ssse3.c b/vp9/encoder/x86/vp9_dct_ssse3.c
index bdc75e993..1c1005aee 100644
--- a/vp9/encoder/x86/vp9_dct_ssse3.c
+++ b/vp9/encoder/x86/vp9_dct_ssse3.c
@@ -293,7 +293,8 @@ void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
 
   if (!skip_block) {
     __m128i eob;
-    __m128i round, quant, dequant;
+    __m128i round, quant, dequant, thr;
+    int16_t nzflag;
     {
       __m128i coeff0, coeff1;
 
@@ -368,6 +369,7 @@ void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
 
     // AC only loop
     index = 2;
+    thr = _mm_srai_epi16(dequant, 1);
     while (n_coeffs < 0) {
       __m128i coeff0, coeff1;
       {
@@ -387,28 +389,39 @@ void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 
-        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
-        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
-        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
-        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
 
-        // Reinsert signs
-        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
-        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
-        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
 
-        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
-        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 
-        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
 
-        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
-        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
       }
 
-      {
+      if (nzflag) {
         // Scan for eob
         __m128i zero_coeff0, zero_coeff1;
         __m128i nzero_coeff0, nzero_coeff1;
diff --git a/vp9/encoder/x86/vp9_quantize_sse2.c b/vp9/encoder/x86/vp9_quantize_sse2.c
index 679c66e30..00abd3c49 100644
--- a/vp9/encoder/x86/vp9_quantize_sse2.c
+++ b/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -230,6 +230,8 @@ void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
                           const int16_t* scan_ptr,
                           const int16_t* iscan_ptr) {
   __m128i zero;
+  __m128i thr;
+  int16_t nzflag;
   (void)scan_ptr;
   (void)zbin_ptr;
   (void)quant_shift_ptr;
@@ -316,6 +318,8 @@ void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
       n_coeffs += 8 * 2;
     }
 
+    thr = _mm_srai_epi16(dequant, 1);
+
     // AC only loop
     while (n_coeffs < 0) {
       __m128i coeff0, coeff1;
@@ -335,28 +339,39 @@ void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 
-        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
-        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
-        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
-        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
 
-        // Reinsert signs
-        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
-        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
-        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
-        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
 
-        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
-        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 
-        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
-        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
 
-        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
-        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
       }
 
-      {
+      if (nzflag) {
         // Scan for eob
         __m128i zero_coeff0, zero_coeff1;
         __m128i nzero_coeff0, nzero_coeff1;