Add a 32-bit friendly sse2 quantizer.

This is based on the 64-bit ssse3 quantizer.

1.1x speedup for screen content at speed 7.

Change-Id: I57d15415ef97c49165954bbe3daaaf9318e37448

vp9/common/vp9_rtcd_defs.pl

@@ -1155,7 +1155,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64";
 
   add_proto qw/void vp9_quantize_b/, "const tran_low_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, tran_low_t *qcoeff_ptr, tran_low_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_b/, "$ssse3_x86_64";
+  specialize qw/vp9_quantize_b sse2/, "$ssse3_x86_64";
 
   add_proto qw/void vp9_quantize_b_32x32/, "const tran_low_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, tran_low_t *qcoeff_ptr, tran_low_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_b_32x32/, "$ssse3_x86_64";

vp9/encoder/x86/vp9_quantize_sse2.c

+/*
+ *  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 <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "vpx/vpx_integer.h"
+
+void vp9_quantize_b_sse2(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_ptr,
+                         const int16_t* iscan_ptr) {
+  __m128i zero;
+  (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();
+  if (!skip_block) {
+    __m128i eob;
+    __m128i zbin;
+    __m128i round, quant, dequant, shift;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        __m128i zbin_oq;
+        __m128i pw_1;
+        zbin_oq = _mm_set1_epi16(zbin_oq_value);
+        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_add_epi16(zbin, zbin_oq);
+        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 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // 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);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, 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);
+      }
+
+      {
+        // 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 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // 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);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        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);
+      }
+
+      {
+        // 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);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}

vp9/vp9cx.mk

@@ -101,6 +101,7 @@ VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_sad4d_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE3) += encoder/x86/vp9_sad_sse3.asm
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c
 
 ifeq ($(CONFIG_USE_X86INC),yes)
 VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_dct_mmx.asm