Revert "quantize: replace highbd versions"

This reverts commit 2200039d33c49a9f7a5c438656df143755b022c4.

This causes failures with VP9/EndToEndTestLarge.EndtoEndPSNRTest/*; it
seems the assembly does not match the C code.

Bug: webm:1586
Change-Id: I4c63beebf88d4c12789d681b0d38014510b147fe

4 files changed, 163 insertions, 16 deletions

diff --git a/vpx_dsp/quantize.h b/vpx_dsp/quantize.h
index 0fcd77941..8e138445e 100644
--- a/vpx_dsp/quantize.h
+++ b/vpx_dsp/quantize.h
@@ -37,22 +37,6 @@ void vpx_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
                                   tran_low_t *qcoeff_ptr,
                                   tran_low_t *dqcoeff_ptr,
                                   const int16_t dequant, uint16_t *eob_ptr);
-
-// Only used for reference. The optimized versions can handle HBD.
-void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
-                             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, uint16_t *eob_ptr,
-                             const int16_t *scan, const int16_t *iscan);
-
-void vpx_highbd_quantize_b_32x32_c(
-    const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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, uint16_t *eob_ptr,
-    const int16_t *scan, const int16_t *iscan);
 #endif
 
 #ifdef __cplusplus
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index bf348c112..a1e511cce 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -318,6 +318,9 @@ DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3.h
 DSP_SRCS-$(HAVE_AVX)    += x86/quantize_avx.c
 DSP_SRCS-$(HAVE_NEON)   += arm/quantize_neon.c
 DSP_SRCS-$(HAVE_VSX)    += ppc/quantize_vsx.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
+endif
 
 # avg
 DSP_SRCS-yes           += avg.c
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 73f28ff92..83dbcfdda 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -714,6 +714,14 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 
   add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
   specialize qw/vpx_quantize_b_32x32 neon ssse3 avx vsx/;
+
+  if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/vpx_highbd_quantize_b sse2/;
+
+    add_proto qw/void vpx_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/vpx_highbd_quantize_b_32x32 sse2/;
+  }  # CONFIG_VP9_HIGHBITDEPTH
 }  # CONFIG_VP9_ENCODER
 
 if (vpx_config("CONFIG_ENCODERS") eq "yes") {
diff --git a/vpx_dsp/x86/highbd_quantize_intrin_sse2.c b/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
new file mode 100644
index 000000000..4535a0f7a
--- /dev/null
+++ b/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -0,0 +1,152 @@
+/*
+ *  Copyright (c) 2015 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 <assert.h>
+#include <emmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
+                                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, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  // Pre-scan pass
+  for (i = ((int)count / 4) - 1; i >= 0; i--) {
+    __m128i coeffs, cmp1, cmp2;
+    int test;
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+    cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+    cmp1 = _mm_and_si128(cmp1, cmp2);
+    test = _mm_movemask_epi8(cmp1);
+    if (test == 0xffff)
+      non_zero_regs--;
+    else
+      break;
+  }
+
+  // Quantization pass:
+  for (i = 0; i < non_zero_regs; i++) {
+    __m128i coeffs, coeffs_sign, tmp1, tmp2;
+    int test;
+    int abs_coeff[4];
+    int coeff_sign[4];
+
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    coeffs_sign = _mm_srai_epi32(coeffs, 31);
+    coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+    tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+    tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+    tmp1 = _mm_or_si128(tmp1, tmp2);
+    test = _mm_movemask_epi8(tmp1);
+    _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
+    _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
+
+    for (j = 0; j < 4; j++) {
+      if (test & (1 << (4 * j))) {
+        int k = 4 * i + j;
+        const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
+        const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
+        const uint32_t abs_qcoeff =
+            (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
+        qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
+        dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+        if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+void vpx_highbd_quantize_b_32x32_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+
+  zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  // Pre-scan pass
+  for (i = 0; i < n_coeffs / 4; i++) {
+    __m128i coeffs, cmp1, cmp2;
+    int test;
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+    cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+    cmp1 = _mm_and_si128(cmp1, cmp2);
+    test = _mm_movemask_epi8(cmp1);
+    if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+    if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+    if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+    if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+  }
+
+  // Quantization pass: only process the coefficients selected in
+  // pre-scan pass. Note: idx can be zero.
+  for (i = 0; i < idx; i++) {
+    const int rc = idx_arr[i];
+    const int coeff = coeff_ptr[rc];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+    const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+    qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+    if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif