diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h
index de60d95d..0bf10a1a 100644
--- a/src/dsp/lossless.h
+++ b/src/dsp/lossless.h
@@ -182,9 +182,9 @@ extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];
 // -----------------------------------------------------------------------------
 // Huffman-cost related functions.
 
-typedef float (*VP8LCostFunc)(const uint32_t* population, int length);
-typedef float (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
-                                      int length);
+typedef uint32_t (*VP8LCostFunc)(const uint32_t* population, int length);
+typedef uint32_t (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
+                                         int length);
 typedef float (*VP8LCombinedShannonEntropyFunc)(const int X[256],
                                                 const int Y[256]);
 
diff --git a/src/dsp/lossless_enc.c b/src/dsp/lossless_enc.c
index cde12806..997d56c2 100644
--- a/src/dsp/lossless_enc.c
+++ b/src/dsp/lossless_enc.c
@@ -636,20 +636,25 @@ void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits,
 
 //------------------------------------------------------------------------------
 
-static float ExtraCost_C(const uint32_t* population, int length) {
+static uint32_t ExtraCost_C(const uint32_t* population, int length) {
   int i;
-  float cost = 0.f;
-  for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2];
+  uint32_t cost = population[4] + population[5];
+  assert(length % 2 == 0);
+  for (i = 2; i < length / 2 - 1; ++i) {
+    cost += i * (population[2 * i + 2] + population[2 * i + 3]);
+  }
   return cost;
 }
 
-static float ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
-                                  int length) {
+static uint32_t ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
+                                    int length) {
   int i;
-  float cost = 0.f;
-  for (i = 2; i < length - 2; ++i) {
-    const int xy = X[i + 2] + Y[i + 2];
-    cost += (i >> 1) * xy;
+  uint32_t cost = X[4] + Y[4] + X[5] + Y[5];
+  assert(length % 2 == 0);
+  for (i = 2; i < length / 2 - 1; ++i) {
+    const int xy0 = X[2 * i + 2] + Y[2 * i + 2];
+    const int xy1 = X[2 * i + 3] + Y[2 * i + 3];
+    cost += i * (xy0 + xy1);
   }
   return cost;
 }
diff --git a/src/dsp/lossless_enc_mips32.c b/src/dsp/lossless_enc_mips32.c
index 639f7866..e10f12da 100644
--- a/src/dsp/lossless_enc_mips32.c
+++ b/src/dsp/lossless_enc_mips32.c
@@ -103,8 +103,8 @@ static float FastLog2Slow_MIPS32(uint32_t v) {
 //     cost += i * *(pop + 1);
 //     pop += 2;
 //   }
-//   return (float)cost;
-static float ExtraCost_MIPS32(const uint32_t* const population, int length) {
+//   return cost;
+static uint32_t ExtraCost_MIPS32(const uint32_t* const population, int length) {
   int i, temp0, temp1;
   const uint32_t* pop = &population[4];
   const uint32_t* const LoopEnd = &population[length];
@@ -130,7 +130,7 @@ static float ExtraCost_MIPS32(const uint32_t* const population, int length) {
     : "memory", "hi", "lo"
   );
 
-  return (float)((int64_t)temp0 << 32 | temp1);
+  return ((int64_t)temp0 << 32 | temp1);
 }
 
 // C version of this function:
@@ -148,9 +148,9 @@ static float ExtraCost_MIPS32(const uint32_t* const population, int length) {
 //     pX += 2;
 //     pY += 2;
 //   }
-//   return (float)cost;
-static float ExtraCostCombined_MIPS32(const uint32_t* const X,
-                                      const uint32_t* const Y, int length) {
+//   return cost;
+static uint32_t ExtraCostCombined_MIPS32(const uint32_t* const X,
+                                         const uint32_t* const Y, int length) {
   int i, temp0, temp1, temp2, temp3;
   const uint32_t* pX = &X[4];
   const uint32_t* pY = &Y[4];
@@ -183,7 +183,7 @@ static float ExtraCostCombined_MIPS32(const uint32_t* const X,
     : "memory", "hi", "lo"
   );
 
-  return (float)((int64_t)temp0 << 32 | temp1);
+  return ((int64_t)temp0 << 32 | temp1);
 }
 
 #define HUFFMAN_COST_PASS                                 \
diff --git a/src/dsp/lossless_enc_sse41.c b/src/dsp/lossless_enc_sse41.c
index ad358a6f..7ab83c26 100644
--- a/src/dsp/lossless_enc_sse41.c
+++ b/src/dsp/lossless_enc_sse41.c
@@ -18,8 +18,53 @@
 #include <smmintrin.h>
 #include "src/dsp/lossless.h"
 
-// For sign-extended multiplying constants, pre-shifted by 5:
-#define CST_5b(X)  (((int16_t)((uint16_t)(X) << 8)) >> 5)
+//------------------------------------------------------------------------------
+// Cost operations.
+
+static WEBP_INLINE uint32_t HorizontalSum_SSE41(__m128i cost) {
+  cost = _mm_add_epi32(cost, _mm_srli_si128(cost, 8));
+  cost = _mm_add_epi32(cost, _mm_srli_si128(cost, 4));
+  return _mm_cvtsi128_si32(cost);
+}
+
+static uint32_t ExtraCost_SSE41(const uint32_t* const a, int length) {
+  int i;
+  __m128i cost = _mm_set_epi32(2 * a[7], 2 * a[6], a[5], a[4]);
+  assert(length % 8 == 0);
+
+  for (i = 8; i + 8 <= length; i += 8) {
+    const int j = (i - 2) >> 1;
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i w = _mm_set_epi32(j + 3, j + 2, j + 1, j);
+    const __m128i a2 = _mm_hadd_epi32(a0, a1);
+    const __m128i mul = _mm_mullo_epi32(a2, w);
+    cost = _mm_add_epi32(mul, cost);
+  }
+  return HorizontalSum_SSE41(cost);
+}
+
+static uint32_t ExtraCostCombined_SSE41(const uint32_t* const a,
+                                        const uint32_t* const b, int length) {
+  int i;
+  __m128i cost = _mm_add_epi32(_mm_set_epi32(2 * a[7], 2 * a[6], a[5], a[4]),
+                               _mm_set_epi32(2 * b[7], 2 * b[6], b[5], b[4]));
+  assert(length % 8 == 0);
+
+  for (i = 8; i + 8 <= length; i += 8) {
+    const int j = (i - 2) >> 1;
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
+    const __m128i w = _mm_set_epi32(j + 3, j + 2, j + 1, j);
+    const __m128i a2 = _mm_hadd_epi32(a0, a1);
+    const __m128i b2 = _mm_hadd_epi32(b0, b1);
+    const __m128i mul = _mm_mullo_epi32(_mm_add_epi32(a2, b2), w);
+    cost = _mm_add_epi32(mul, cost);
+  }
+  return HorizontalSum_SSE41(cost);
+}
 
 //------------------------------------------------------------------------------
 // Subtract-Green Transform
@@ -44,6 +89,9 @@ static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data,
 //------------------------------------------------------------------------------
 // Color Transform
 
+// For sign-extended multiplying constants, pre-shifted by 5:
+#define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5)
+
 #define MK_CST_16(HI, LO) \
   _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
 
@@ -143,6 +191,8 @@ static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride,
 extern void VP8LEncDspInitSSE41(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) {
+  VP8LExtraCost = ExtraCost_SSE41;
+  VP8LExtraCostCombined = ExtraCostCombined_SSE41;
   VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE41;
   VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE41;
   VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE41;
diff --git a/src/enc/histogram_enc.c b/src/enc/histogram_enc.c
index 8418def2..9e1d18f1 100644
--- a/src/enc/histogram_enc.c
+++ b/src/enc/histogram_enc.c
@@ -358,15 +358,17 @@ static WEBP_INLINE float GetCombinedEntropy(const uint32_t* const X,
 
 // Estimates the Entropy + Huffman + other block overhead size cost.
 float VP8LHistogramEstimateBits(VP8LHistogram* const p) {
-  return
-      PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_),
-                     NULL, &p->is_used_[0])
-      + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1])
-      + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2])
-      + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3])
-      + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4])
-      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
-      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
+  return PopulationCost(p->literal_,
+                        VP8LHistogramNumCodes(p->palette_code_bits_), NULL,
+                        &p->is_used_[0]) +
+         PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1]) +
+         PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2]) +
+         PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3]) +
+         PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL,
+                        &p->is_used_[4]) +
+         (float)VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES,
+                              NUM_LENGTH_CODES) +
+         (float)VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
 }
 
 // -----------------------------------------------------------------------------
@@ -381,9 +383,9 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
   *cost += GetCombinedEntropy(a->literal_, b->literal_,
                               VP8LHistogramNumCodes(palette_code_bits),
                               a->is_used_[0], b->is_used_[0], 0);
-  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
-                                 b->literal_ + NUM_LITERAL_CODES,
-                                 NUM_LENGTH_CODES);
+  *cost += (float)VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
+                                        b->literal_ + NUM_LITERAL_CODES,
+                                        NUM_LENGTH_CODES);
   if (*cost > cost_threshold) return 0;
 
   if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM &&
@@ -417,8 +419,8 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
   *cost +=
       GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES,
                          a->is_used_[4], b->is_used_[4], 0);
-  *cost +=
-      VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES);
+  *cost += (float)VP8LExtraCostCombined(a->distance_, b->distance_,
+                                        NUM_DISTANCE_CODES);
   if (*cost > cost_threshold) return 0;
 
   return 1;
@@ -506,11 +508,11 @@ static void UpdateHistogramCost(VP8LHistogram* const h) {
       PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym, &h->is_used_[3]);
   const float distance_cost =
       PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) +
-      VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
+      (float)VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
   const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
   h->literal_cost_ =
       PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) +
-          VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES);
+      (float)VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES);
   h->red_cost_ =
       PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]);
   h->blue_cost_ =