Split HistogramAdd to only have the high level logic in C.

Change-Id: Ic9eaebf7128ca0215b49d2a13bde1f5b94a28061
(cherry picked from commit dea3e89983f299b3325898fa5b9474be258553b2)

4 files changed, 72 insertions, 122 deletions

diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h
index b2bbdfc9..f709cc86 100644
--- a/src/dsp/lossless.h
+++ b/src/dsp/lossless.h
@@ -163,7 +163,7 @@ extern VP8LCostCombinedFunc VP8LExtraCostCombined;
 extern VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy;
 
 typedef struct {        // small struct to hold counters
-  int counts[2];        // index: 0=zero steak, 1=non-zero streak
+  int counts[2];        // index: 0=zero streak, 1=non-zero streak
   int streaks[2][2];    // [zero/non-zero][streak<3 / streak>=3]
 } VP8LStreaks;
 
@@ -194,10 +194,14 @@ extern VP8LGetEntropyUnrefinedFunc VP8LGetEntropyUnrefined;
 void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
                               VP8LBitEntropy* const entropy);
 
-typedef void (*VP8LHistogramAddFunc)(const VP8LHistogram* const a,
-                                     const VP8LHistogram* const b,
-                                     VP8LHistogram* const out);
-extern VP8LHistogramAddFunc VP8LHistogramAdd;
+typedef void (*VP8LAddVectorFunc)(const uint32_t* a, const uint32_t* b,
+                                  uint32_t* out, int size);
+extern VP8LAddVectorFunc VP8LAddVector;
+typedef void (*VP8LAddVectorEqFunc)(const uint32_t* a, uint32_t* out, int size);
+extern VP8LAddVectorEqFunc VP8LAddVectorEq;
+void VP8LHistogramAdd(const VP8LHistogram* const a,
+                      const VP8LHistogram* const b,
+                      VP8LHistogram* const out);
 
 // -----------------------------------------------------------------------------
 // PrefixEncode()
diff --git a/src/dsp/lossless_enc.c b/src/dsp/lossless_enc.c
index d608326f..6ec9e46a 100644
--- a/src/dsp/lossless_enc.c
+++ b/src/dsp/lossless_enc.c
@@ -632,36 +632,34 @@ static double ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
 
 //------------------------------------------------------------------------------
 
-static void HistogramAdd_C(const VP8LHistogram* const a,
-                           const VP8LHistogram* const b,
-                           VP8LHistogram* const out) {
+static void AddVector_C(const uint32_t* a, const uint32_t* b, uint32_t* out,
+                        int size) {
   int i;
+  for (i = 0; i < size; ++i) out[i] = a[i] + b[i];
+}
+
+static void AddVectorEq_C(const uint32_t* a, uint32_t* out, int size) {
+  int i;
+  for (i = 0; i < size; ++i) out[i] += a[i];
+}
+
+void VP8LHistogramAdd(const VP8LHistogram* const a,
+                      const VP8LHistogram* const b, VP8LHistogram* const out) {
   const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
   assert(a->palette_code_bits_ == b->palette_code_bits_);
   if (b != out) {
-    for (i = 0; i < literal_size; ++i) {
-      out->literal_[i] = a->literal_[i] + b->literal_[i];
-    }
-    for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-      out->distance_[i] = a->distance_[i] + b->distance_[i];
-    }
-    for (i = 0; i < NUM_LITERAL_CODES; ++i) {
-      out->red_[i] = a->red_[i] + b->red_[i];
-      out->blue_[i] = a->blue_[i] + b->blue_[i];
-      out->alpha_[i] = a->alpha_[i] + b->alpha_[i];
-    }
+    VP8LAddVector(a->literal_, b->literal_, out->literal_, literal_size);
+    VP8LAddVector(a->distance_, b->distance_, out->distance_,
+                  NUM_DISTANCE_CODES);
+    VP8LAddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
+    VP8LAddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
+    VP8LAddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
   } else {
-    for (i = 0; i < literal_size; ++i) {
-      out->literal_[i] += a->literal_[i];
-    }
-    for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-      out->distance_[i] += a->distance_[i];
-    }
-    for (i = 0; i < NUM_LITERAL_CODES; ++i) {
-      out->red_[i] += a->red_[i];
-      out->blue_[i] += a->blue_[i];
-      out->alpha_[i] += a->alpha_[i];
-    }
+    VP8LAddVectorEq(a->literal_, out->literal_, literal_size);
+    VP8LAddVectorEq(a->distance_, out->distance_, NUM_DISTANCE_CODES);
+    VP8LAddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES);
+    VP8LAddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES);
+    VP8LAddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
   }
 }
 
@@ -848,7 +846,8 @@ VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy;
 VP8LGetEntropyUnrefinedFunc VP8LGetEntropyUnrefined;
 VP8LGetCombinedEntropyUnrefinedFunc VP8LGetCombinedEntropyUnrefined;
 
-VP8LHistogramAddFunc VP8LHistogramAdd;
+VP8LAddVectorFunc VP8LAddVector;
+VP8LAddVectorEqFunc VP8LAddVectorEq;
 
 VP8LVectorMismatchFunc VP8LVectorMismatch;
 VP8LBundleColorMapFunc VP8LBundleColorMap;
@@ -885,7 +884,8 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) {
   VP8LGetEntropyUnrefined = GetEntropyUnrefined_C;
   VP8LGetCombinedEntropyUnrefined = GetCombinedEntropyUnrefined_C;
 
-  VP8LHistogramAdd = HistogramAdd_C;
+  VP8LAddVector = AddVector_C;
+  VP8LAddVectorEq = AddVectorEq_C;
 
   VP8LVectorMismatch = VectorMismatch_C;
   VP8LBundleColorMap = VP8LBundleColorMap_C;
@@ -971,7 +971,8 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) {
   assert(VP8LCombinedShannonEntropy != NULL);
   assert(VP8LGetEntropyUnrefined != NULL);
   assert(VP8LGetCombinedEntropyUnrefined != NULL);
-  assert(VP8LHistogramAdd != NULL);
+  assert(VP8LAddVector != NULL);
+  assert(VP8LAddVectorEq != NULL);
   assert(VP8LVectorMismatch != NULL);
   assert(VP8LBundleColorMap != NULL);
   assert(VP8LPredictorsSub[0] != NULL);
diff --git a/src/dsp/lossless_enc_mips32.c b/src/dsp/lossless_enc_mips32.c
index e7b58f4e..0412a093 100644
--- a/src/dsp/lossless_enc_mips32.c
+++ b/src/dsp/lossless_enc_mips32.c
@@ -344,65 +344,29 @@ static void GetCombinedEntropyUnrefined_MIPS32(const uint32_t X[],
     ASM_END_COMMON_0                                    \
     ASM_END_COMMON_1
 
-#define ADD_VECTOR(A, B, OUT, SIZE, EXTRA_SIZE)  do {   \
-  const uint32_t* pa = (const uint32_t*)(A);            \
-  const uint32_t* pb = (const uint32_t*)(B);            \
-  uint32_t* pout = (uint32_t*)(OUT);                    \
-  const uint32_t* const LoopEnd = pa + (SIZE);          \
-  assert((SIZE) % 4 == 0);                              \
-  ASM_START                                             \
-  ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout)              \
-  ASM_END_0                                             \
-  if ((EXTRA_SIZE) > 0) {                               \
-    const int last = (EXTRA_SIZE);                      \
-    int i;                                              \
-    for (i = 0; i < last; ++i) pout[i] = pa[i] + pb[i]; \
-  }                                                     \
-} while (0)
-
-#define ADD_VECTOR_EQ(A, OUT, SIZE, EXTRA_SIZE)  do {   \
-  const uint32_t* pa = (const uint32_t*)(A);            \
-  uint32_t* pout = (uint32_t*)(OUT);                    \
-  const uint32_t* const LoopEnd = pa + (SIZE);          \
-  assert((SIZE) % 4 == 0);                              \
-  ASM_START                                             \
-  ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout)            \
-  ASM_END_1                                             \
-  if ((EXTRA_SIZE) > 0) {                               \
-    const int last = (EXTRA_SIZE);                      \
-    int i;                                              \
-    for (i = 0; i < last; ++i) pout[i] += pa[i];        \
-  }                                                     \
-} while (0)
-
-static void HistogramAdd_MIPS32(const VP8LHistogram* const a,
-                                const VP8LHistogram* const b,
-                                VP8LHistogram* const out) {
+static void AddVector_MIPS32(const uint32_t* pa, const uint32_t* pb,
+                             uint32_t* pout, int size) {
   uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
-  const int extra_cache_size = VP8LHistogramNumCodes(a->palette_code_bits_)
-                             - (NUM_LITERAL_CODES + NUM_LENGTH_CODES);
-  assert(a->palette_code_bits_ == b->palette_code_bits_);
-
-  if (b != out) {
-    ADD_VECTOR(a->literal_, b->literal_, out->literal_,
-               NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size);
-    ADD_VECTOR(a->distance_, b->distance_, out->distance_,
-               NUM_DISTANCE_CODES, 0);
-    ADD_VECTOR(a->red_, b->red_, out->red_, NUM_LITERAL_CODES, 0);
-    ADD_VECTOR(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES, 0);
-    ADD_VECTOR(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES, 0);
-  } else {
-    ADD_VECTOR_EQ(a->literal_, out->literal_,
-                  NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size);
-    ADD_VECTOR_EQ(a->distance_, out->distance_, NUM_DISTANCE_CODES, 0);
-    ADD_VECTOR_EQ(a->red_, out->red_, NUM_LITERAL_CODES, 0);
-    ADD_VECTOR_EQ(a->blue_, out->blue_, NUM_LITERAL_CODES, 0);
-    ADD_VECTOR_EQ(a->alpha_, out->alpha_, NUM_LITERAL_CODES, 0);
-  }
+  const uint32_t end = ((size) / 4) * 4;
+  const uint32_t* const LoopEnd = pa + end;
+  int i;
+  ASM_START
+  ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout)
+  ASM_END_0
+  for (i = end; i < size; ++i) pout[i] = pa[i] + pb[i];
+}
+
+static void AddVectorEq_MIPS32(const uint32_t* pa, uint32_t* pout, int size) {
+  uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+  const uint32_t end = ((size) / 4) * 4;
+  const uint32_t* const LoopEnd = pa + end;
+  int i;
+  ASM_START
+  ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout)
+  ASM_END_1
+  for (i = end; i < size; ++i) pout[i] += pa[i];
 }
 
-#undef ADD_VECTOR_EQ
-#undef ADD_VECTOR
 #undef ASM_END_1
 #undef ASM_END_0
 #undef ASM_END_COMMON_1
@@ -422,7 +386,8 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) {
   VP8LExtraCostCombined = ExtraCostCombined_MIPS32;
   VP8LGetEntropyUnrefined = GetEntropyUnrefined_MIPS32;
   VP8LGetCombinedEntropyUnrefined = GetCombinedEntropyUnrefined_MIPS32;
-  VP8LHistogramAdd = HistogramAdd_MIPS32;
+  VP8LAddVector = AddVector_MIPS32;
+  VP8LAddVectorEq = AddVectorEq_MIPS32;
 }
 
 #else  // !WEBP_USE_MIPS32
diff --git a/src/dsp/lossless_enc_sse2.c b/src/dsp/lossless_enc_sse2.c
index f84a9909..36478c49 100644
--- a/src/dsp/lossless_enc_sse2.c
+++ b/src/dsp/lossless_enc_sse2.c
@@ -170,12 +170,13 @@ static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride,
 
 //------------------------------------------------------------------------------
 
+// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
+// that's ok since the histogram values are less than 1<<28 (max picture size).
 #define LINE_SIZE 16    // 8 or 16
 static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
                            int size) {
   int i;
-  assert(size % LINE_SIZE == 0);
-  for (i = 0; i < size; i += LINE_SIZE) {
+  for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
     const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
     const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
 #if (LINE_SIZE == 16)
@@ -195,12 +196,14 @@ static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
     _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
 #endif
   }
+  for (; i < size; ++i) {
+    out[i] = a[i] + b[i];
+  }
 }
 
 static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
   int i;
-  assert(size % LINE_SIZE == 0);
-  for (i = 0; i < size; i += LINE_SIZE) {
+  for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
     const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
     const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
 #if (LINE_SIZE == 16)
@@ -220,35 +223,11 @@ static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
     _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
 #endif
   }
-}
-#undef LINE_SIZE
-
-// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
-// that's ok since the histogram values are less than 1<<28 (max picture size).
-static void HistogramAdd_SSE2(const VP8LHistogram* const a,
-                              const VP8LHistogram* const b,
-                              VP8LHistogram* const out) {
-  int i;
-  const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
-  assert(a->palette_code_bits_ == b->palette_code_bits_);
-  if (b != out) {
-    AddVector_SSE2(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
-    AddVector_SSE2(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
-    AddVector_SSE2(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
-    AddVector_SSE2(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
-  } else {
-    AddVectorEq_SSE2(a->literal_, out->literal_, NUM_LITERAL_CODES);
-    AddVectorEq_SSE2(a->red_, out->red_, NUM_LITERAL_CODES);
-    AddVectorEq_SSE2(a->blue_, out->blue_, NUM_LITERAL_CODES);
-    AddVectorEq_SSE2(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
-  }
-  for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
-    out->literal_[i] = a->literal_[i] + b->literal_[i];
-  }
-  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-    out->distance_[i] = a->distance_[i] + b->distance_[i];
+  for (; i < size; ++i) {
+    out[i] += a[i];
   }
 }
+#undef LINE_SIZE
 
 //------------------------------------------------------------------------------
 // Entropy
@@ -675,7 +654,8 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
   VP8LTransformColor = TransformColor_SSE2;
   VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE2;
   VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2;
-  VP8LHistogramAdd = HistogramAdd_SSE2;
+  VP8LAddVector = AddVector_SSE2;
+  VP8LAddVectorEq = AddVectorEq_SSE2;
   VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2;
   VP8LVectorMismatch = VectorMismatch_SSE2;
   VP8LBundleColorMap = BundleColorMap_SSE2;