encoding SSE4.1 stub for StoreHistogram + Quantize + SSE_16xN

Visible speed-up, thanks to pshufb and pabsw and psignw use.

had to tweak configure.ac to make "smmintri.h" presence correctly
detected (we need to set the CPPFLAGS instead of the CFLAGS!)

Change-Id: I2ab99e16a27a64fdf1f09b2b4e30a5e74ccca080

Android.mk

@@ -79,6 +79,7 @@ dsp_enc_srcs := \
     src/dsp/enc_mips_dsp_r2.c \
     src/dsp/enc_neon.$(NEON) \
     src/dsp/enc_sse2.c \
+    src/dsp/enc_sse41.c \
     src/dsp/lossless_enc.c \
     src/dsp/lossless_enc_mips32.c \
     src/dsp/lossless_enc_mips_dsp_r2.c \

Makefile.vc

@@ -229,6 +229,7 @@ DSP_ENC_OBJS = \
     $(DIROBJ)\dsp\enc_mips_dsp_r2.obj \
     $(DIROBJ)\dsp\enc_neon.obj \
     $(DIROBJ)\dsp\enc_sse2.obj \
+    $(DIROBJ)\dsp\enc_sse41.obj \
     $(DIROBJ)\dsp\lossless_enc.obj \
     $(DIROBJ)\dsp\lossless_enc_mips32.obj \
     $(DIROBJ)\dsp\lossless_enc_mips_dsp_r2.obj \

configure.ac

@@ -99,13 +99,13 @@ AC_SUBST([AVX2_FLAGS])
 
 TEST_AND_ADD_CFLAGS([SSE41_FLAGS], [-msse4.1])
 AS_IF([test -n "$SSE41_FLAGS"], [
-  SAVED_CFLAGS=$CFLAGS
+  SAVED_CPPFLAGS=$CPPFLAGS
-  CFLAGS="$CFLAGS $SSE41_FLAGS"
+  CPPFLAGS="$CPPFLAGS $SSE41_FLAGS"
   AC_CHECK_HEADER([smmintrin.h],
                   [AC_DEFINE(WEBP_HAVE_SSE41, [1],
                    [Set to 1 if SSE4.1 is supported])],
                   [SSE41_FLAGS=""])
-  CFLAGS=$SAVED_CFLAGS])
+  CPPFLAGS=$SAVED_CPPFLAGS])
 AC_SUBST([SSE41_FLAGS])
 
 TEST_AND_ADD_CFLAGS([SSE2_FLAGS], [-msse2])

makefile.unix

@@ -159,6 +159,7 @@ DSP_ENC_OBJS = \
     src/dsp/enc_mips_dsp_r2.o \
     src/dsp/enc_neon.o \
     src/dsp/enc_sse2.o \
+    src/dsp/enc_sse41.o \
     src/dsp/lossless_enc.o \
     src/dsp/lossless_enc_mips32.o \
     src/dsp/lossless_enc_mips_dsp_r2.o \

src/dsp/Makefile.am

@@ -1,6 +1,6 @@
 noinst_LTLIBRARIES = libwebpdsp.la libwebpdsp_avx2.la
 noinst_LTLIBRARIES += libwebpdsp_sse2.la libwebpdspdecode_sse2.la
-noinst_LTLIBRARIES += libwebpdspdecode_sse41.la
+noinst_LTLIBRARIES += libwebpdsp_sse41.la libwebpdspdecode_sse41.la
 
 if BUILD_LIBWEBPDECODER
   noinst_LTLIBRARIES += libwebpdspdecode.la
@@ -83,6 +83,12 @@ libwebpdsp_sse2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
 libwebpdsp_sse2_la_CFLAGS = $(AM_CFLAGS) $(SSE2_FLAGS)
 libwebpdsp_sse2_la_LIBADD = libwebpdspdecode_sse2.la
 
+libwebpdsp_sse41_la_SOURCES =
+libwebpdsp_sse41_la_SOURCES += enc_sse41.c
+libwebpdsp_sse41_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+libwebpdsp_sse41_la_CFLAGS = $(AM_CFLAGS) $(SSE41_FLAGS)
+libwebpdsp_sse41_la_LIBADD = libwebpdspdecode_sse41.la
+
 libwebpdsp_la_SOURCES = $(COMMON_SOURCES) $(ENC_SOURCES)
 
 noinst_HEADERS =
@@ -93,7 +99,7 @@ libwebpdsp_la_CPPFLAGS = $(USE_EXPERIMENTAL_CODE) $(USE_SWAP_16BIT_CSP)
 libwebpdsp_la_LDFLAGS = -lm
 libwebpdsp_la_LIBADD =
 libwebpdsp_la_LIBADD += libwebpdsp_avx2.la libwebpdsp_sse2.la
-libwebpdsp_la_LIBADD += libwebpdspdecode_sse41.la
+libwebpdsp_la_LIBADD += libwebpdsp_sse41.la
 
 if BUILD_LIBWEBPDECODER
   libwebpdspdecode_la_SOURCES = $(COMMON_SOURCES)

src/dsp/enc.c

@@ -722,6 +722,7 @@ VP8BlockCopy VP8Copy4x4;
 VP8BlockCopy VP8Copy16x8;
 
 extern void VP8EncDspInitSSE2(void);
+extern void VP8EncDspInitSSE41(void);
 extern void VP8EncDspInitAVX2(void);
 extern void VP8EncDspInitNEON(void);
 extern void VP8EncDspInitMIPS32(void);
@@ -761,6 +762,11 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       VP8EncDspInitSSE2();
+#if defined(WEBP_USE_SSE41)
+      if (VP8GetCPUInfo(kSSE4_1)) {
+        VP8EncDspInitSSE41();
+      }
+#endif
     }
 #endif
 #if defined(WEBP_USE_AVX2)

src/dsp/enc_sse41.c

@@ -0,0 +1,263 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING 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.
+// -----------------------------------------------------------------------------
+//
+// SSE4 version of some encoding functions.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE41)
+#include <smmintrin.h>
+
+#include "../enc/vp8enci.h"
+
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms.
+
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
+  int j;
+  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
+  for (j = start_block; j < end_block; ++j) {
+    int16_t out[16];
+    int k;
+
+    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+
+    // Convert coefficients to bin (within out[]).
+    {
+      // Load.
+      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
+      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
+      // v = abs(out) >> 3
+      const __m128i abs0 = _mm_abs_epi16(out0);
+      const __m128i abs1 = _mm_abs_epi16(out1);
+      const __m128i v0 = _mm_srai_epi16(abs0, 3);
+      const __m128i v1 = _mm_srai_epi16(abs1, 3);
+      // bin = min(v, MAX_COEFF_THRESH)
+      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
+      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
+      // Store.
+      _mm_storeu_si128((__m128i*)&out[0], bin0);
+      _mm_storeu_si128((__m128i*)&out[8], bin1);
+    }
+
+    // Convert coefficients to bin.
+    for (k = 0; k < 16; ++k) {
+      ++distribution[out[k]];
+    }
+  }
+  VP8SetHistogramData(distribution, histo);
+}
+
+//------------------------------------------------------------------------------
+// Metric
+
+static WEBP_INLINE __m128i SubtractAndAccumulate(const __m128i a,
+                                                 const __m128i b) {
+  // take abs(a-b) in 8b
+  const __m128i a_b = _mm_subs_epu8(a, b);
+  const __m128i b_a = _mm_subs_epu8(b, a);
+  const __m128i abs_a_b = _mm_or_si128(a_b, b_a);
+  // zero-extend to 16b
+  const __m128i C0 = _mm_cvtepu8_epi16(abs_a_b);
+  const __m128i C1 = _mm_cvtepu8_epi16(_mm_srli_si128(abs_a_b, 8));
+  // multiply with self
+  const __m128i D0 = _mm_madd_epi16(C0, C0);
+  const __m128i D1 = _mm_madd_epi16(C1, C1);
+  // accumulate
+  const __m128i sum = _mm_add_epi32(D0, D1);
+  return sum;
+}
+
+static int SSE_16xN(const uint8_t* a, const uint8_t* b, int num_pairs) {
+  __m128i sum = _mm_setzero_si128();
+  int32_t tmp[4];
+
+  while (num_pairs-- > 0) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[BPS * 0]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[BPS * 1]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[BPS * 0]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[BPS * 1]);
+    const __m128i sum1 = SubtractAndAccumulate(a0, b0);
+    const __m128i sum2 = SubtractAndAccumulate(a1, b1);
+    const __m128i sum12 = _mm_add_epi32(sum1, sum2);
+    sum = _mm_add_epi32(sum, sum12);
+    a += 2 * BPS;
+    b += 2 * BPS;
+  }
+  _mm_storeu_si128((__m128i*)tmp, sum);
+  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+}
+
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  return SSE_16xN(a, b, 8);
+}
+
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  return SSE_16xN(a, b, 4);
+}
+
+//------------------------------------------------------------------------------
+// Quantization
+//
+
+// Generates a pshufb constant for shuffling 16b words.
+#define PSHUFB_CST(A,B,C,D,E,F,G,H) \
+  _mm_set_epi8(2 * (H) + 1, 2 * (H) + 0, 2 * (G) + 1, 2 * (G) + 0, \
+               2 * (F) + 1, 2 * (F) + 0, 2 * (E) + 1, 2 * (E) + 0, \
+               2 * (D) + 1, 2 * (D) + 0, 2 * (C) + 1, 2 * (C) + 0, \
+               2 * (B) + 1, 2 * (B) + 0, 2 * (A) + 1, 2 * (A) + 0)
+
+static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
+                                       const uint16_t* const sharpen,
+                                       const VP8Matrix* const mtx) {
+  const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i out0, out8;
+  __m128i packed_out;
+
+  // Load all inputs.
+  // TODO(cduvivier): Make variable declarations and allocations aligned so that
+  //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
+  __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
+  __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
+  const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
+  const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
+  const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
+  const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
+
+  // coeff = abs(in)
+  __m128i coeff0 = _mm_abs_epi16(in0);
+  __m128i coeff8 = _mm_abs_epi16(in8);
+
+  // coeff = abs(in) + sharpen
+  if (sharpen != NULL) {
+    const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
+    const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
+    coeff0 = _mm_add_epi16(coeff0, sharpen0);
+    coeff8 = _mm_add_epi16(coeff8, sharpen8);
+  }
+
+  // out = (coeff * iQ + B) >> QFIX
+  {
+    // doing calculations with 32b precision (QFIX=17)
+    // out = (coeff * iQ)
+    const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
+    const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
+    const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
+    const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
+    __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
+    __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
+    __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
+    __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
+    // out = (coeff * iQ + B)
+    const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
+    const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
+    const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
+    const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
+    out_00 = _mm_add_epi32(out_00, bias_00);
+    out_04 = _mm_add_epi32(out_04, bias_04);
+    out_08 = _mm_add_epi32(out_08, bias_08);
+    out_12 = _mm_add_epi32(out_12, bias_12);
+    // out = QUANTDIV(coeff, iQ, B, QFIX)
+    out_00 = _mm_srai_epi32(out_00, QFIX);
+    out_04 = _mm_srai_epi32(out_04, QFIX);
+    out_08 = _mm_srai_epi32(out_08, QFIX);
+    out_12 = _mm_srai_epi32(out_12, QFIX);
+
+    // pack result as 16b
+    out0 = _mm_packs_epi32(out_00, out_04);
+    out8 = _mm_packs_epi32(out_08, out_12);
+
+    // if (coeff > 2047) coeff = 2047
+    out0 = _mm_min_epi16(out0, max_coeff_2047);
+    out8 = _mm_min_epi16(out8, max_coeff_2047);
+  }
+
+  // put sign back
+  out0 = _mm_sign_epi16(out0, in0);
+  out8 = _mm_sign_epi16(out8, in8);
+
+  // in = out * Q
+  in0 = _mm_mullo_epi16(out0, q0);
+  in8 = _mm_mullo_epi16(out8, q8);
+
+  _mm_storeu_si128((__m128i*)&in[0], in0);
+  _mm_storeu_si128((__m128i*)&in[8], in8);
+
+  // zigzag the output before storing it. The re-ordering is:
+  //    0 1 2 3 4 5 6 7 | 8  9 10 11 12 13 14 15
+  // -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15
+  // There's only two misplaced entries ([8] and [7]) that are crossing the
+  // reg's boundaries.
+  // We use pshufb instead of pshuflo/pshufhi.
+  {
+    const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6);
+    const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1);
+    const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo);
+    const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7);  // extract #7
+    const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7);
+    const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1);
+    const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi);
+    const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8);  // extract #8
+    const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8);
+    const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7);
+    _mm_storeu_si128((__m128i*)&out[0], out_z0);
+    _mm_storeu_si128((__m128i*)&out[8], out_z8);
+    packed_out = _mm_packs_epi16(out_z0, out_z8);
+  }
+
+  // detect if all 'out' values are zeroes or not
+  return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
+}
+
+#undef PSHUFB_CST
+
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx);
+}
+
+static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
+                            const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, NULL, mtx);
+}
+
+static int Quantize2Blocks(int16_t in[32], int16_t out[32],
+                           const VP8Matrix* const mtx) {
+  int nz;
+  const uint16_t* const sharpen = &mtx->sharpen_[0];
+  nz  = DoQuantizeBlock(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
+  nz |= DoQuantizeBlock(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1;
+  return nz;
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8EncDspInitSSE41(void);
+WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) {
+  VP8CollectHistogram = CollectHistogram;
+  VP8EncQuantizeBlock = QuantizeBlock;
+  VP8EncQuantize2Blocks = Quantize2Blocks;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE16x8 = SSE16x8;
+}
+
+#else  // !WEBP_USE_SSE41
+
+extern void VP8EncDspInitSSE41(void);
+WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) {}
+
+#endif  // WEBP_USE_SSE41