Merge pull request #183 from aous72/unnecessarily_coded_blocks_bug

This fixes a bug that results in coding all-zero codeblocks, producing files that are larger than they need to be.
The increase in size is not fixed and can change from one run to the other.
The bug is on the SIMD accelerated paths only.

Author: aous72

src/core/codestream/ojph_codestream_avx2.cpp

@@ -88,7 +88,7 @@ namespace ojph {
       __m256i m0 = _mm256_set1_epi32(INT_MIN);
       __m256i tmax = _mm256_loadu_si256((__m256i*)max_val);
       __m256i *p = (__m256i*)sp;
-      for (ui32 i = 0; i < count; i += 8, p += 1, dp += 8)
+      for ( ; count >= 8; count -= 8, p += 1, dp += 8)
       {
         __m256i v = _mm256_loadu_si256(p);
         __m256i sign = _mm256_and_si256(v, m0);
@@ -98,6 +98,22 @@ namespace ojph {
         val = _mm256_or_si256(val, sign);
         _mm256_storeu_si256((__m256i*)dp, val);
       }
+      if (count)
+      {
+        __m256i v = _mm256_loadu_si256(p);
+        __m256i sign = _mm256_and_si256(v, m0);
+        __m256i val = _mm256_abs_epi32(v);
+        val = _mm256_slli_epi32(val, (int)shift);
+
+        __m256i c = _mm256_set1_epi32((si32)count);
+        __m256i idx = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+        __m256i mask = _mm256_cmpgt_epi32(c, idx);
+        c = _mm256_and_si256(val, mask);
+        tmax = _mm256_or_si256(tmax, c);
+
+        val = _mm256_or_si256(val, sign);
+        _mm256_storeu_si256((__m256i*)dp, val);
+      }
       _mm256_storeu_si256((__m256i*)max_val, tmax);
     }
 
@@ -113,7 +129,7 @@ namespace ojph {
       __m256i tmax = _mm256_loadu_si256((__m256i*)max_val);
       float *p = (float*)sp;
 
-      for (ui32 i = 0; i < count; i += 8, p += 8, dp += 8)
+      for ( ; count >= 8; count -= 8, p += 8, dp += 8)
       {
         __m256 vf = _mm256_loadu_ps(p);
         vf = _mm256_mul_ps(vf, d);                // multiply
@@ -124,6 +140,23 @@ namespace ojph {
         val = _mm256_or_si256(val, sign);
         _mm256_storeu_si256((__m256i*)dp, val);
       }
+      if (count)
+      {
+        __m256 vf = _mm256_loadu_ps(p);
+        vf = _mm256_mul_ps(vf, d);                // multiply
+        __m256i val = _mm256_cvtps_epi32(vf);     // convert to int
+        __m256i sign = _mm256_and_si256(val, m0); // get sign
+        val = _mm256_abs_epi32(val);
+
+        __m256i c = _mm256_set1_epi32((si32)count);
+        __m256i idx = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+        __m256i mask = _mm256_cmpgt_epi32(c, idx);
+        c = _mm256_and_si256(val, mask);
+        tmax = _mm256_or_si256(tmax, c);
+
+        val = _mm256_or_si256(val, sign);
+        _mm256_storeu_si256((__m256i*)dp, val);
+      }
       _mm256_storeu_si256((__m256i*)max_val, tmax);
     }
 
@@ -178,7 +211,7 @@ namespace ojph {
       __m256i one = _mm256_set1_epi64x(1);
       __m256i tmax = _mm256_loadu_si256((__m256i*)max_val);
       __m256i *p = (__m256i*)sp;
-      for (ui32 i = 0; i < count; i += 4, p += 1, dp += 4)
+      for ( ; count >= 4; count -= 4, p += 1, dp += 4)
       {
         __m256i v = _mm256_loadu_si256(p);
         __m256i sign = _mm256_cmpgt_epi64(zero, v);
@@ -191,6 +224,25 @@ namespace ojph {
         val = _mm256_or_si256(val, sign);
         _mm256_storeu_si256((__m256i*)dp, val);
       }
+      if (count)
+      {
+        __m256i v = _mm256_loadu_si256(p);
+        __m256i sign = _mm256_cmpgt_epi64(zero, v);
+        __m256i val = _mm256_xor_si256(v, sign);  // negate 1's complement
+        __m256i ones = _mm256_and_si256(sign, one);
+        val = _mm256_add_epi64(val, ones);        // 2's complement
+        sign = _mm256_and_si256(sign, m0);
+        val = _mm256_slli_epi64(val, (int)shift);
+
+        __m256i c = _mm256_set1_epi64x(count);
+        __m256i idx = _mm256_set_epi64x(3, 2, 1, 0);
+        __m256i mask = _mm256_cmpgt_epi64(c, idx);
+        c = _mm256_and_si256(val, mask);
+        tmax = _mm256_or_si256(tmax, c);
+
+        val = _mm256_or_si256(val, sign);
+        _mm256_storeu_si256((__m256i*)dp, val);
+      }
       _mm256_storeu_si256((__m256i*)max_val, tmax);
     }
 

src/core/codestream/ojph_codestream_sse2.cpp

@@ -87,7 +87,7 @@ namespace ojph {
       __m128i one = _mm_set1_epi32(1);
       __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
       __m128i *p = (__m128i*)sp;
-      for (ui32 i = 0; i < count; i += 4, p += 1, dp += 4)
+      for ( ; count >= 4; count -= 4, p += 1, dp += 4)
       {
         __m128i v = _mm_loadu_si128(p);
         __m128i sign = _mm_cmplt_epi32(v, zero);
@@ -100,6 +100,25 @@ namespace ojph {
         val = _mm_or_si128(val, sign);
         _mm_storeu_si128((__m128i*)dp, val);
       }
+      if (count)
+      {
+        __m128i v = _mm_loadu_si128(p);
+        __m128i sign = _mm_cmplt_epi32(v, zero);
+        __m128i val = _mm_xor_si128(v, sign); // negate 1's complement
+        __m128i ones = _mm_and_si128(sign, one);
+        val = _mm_add_epi32(val, ones);        // 2's complement
+        sign = _mm_and_si128(sign, m0);
+        val = _mm_slli_epi32(val, (int)shift);
+
+        __m128i c = _mm_set1_epi32((si32)count);
+        __m128i idx = _mm_set_epi32(3, 2, 1, 0);
+        __m128i mask = _mm_cmpgt_epi32(c, idx);
+        c = _mm_and_si128(val, mask);
+        tmax = _mm_or_si128(tmax, c);
+
+        val = _mm_or_si128(val, sign);
+        _mm_storeu_si128((__m128i*)dp, val);
+      }
       _mm_storeu_si128((__m128i*)max_val, tmax);
     }
 
@@ -116,7 +135,7 @@ namespace ojph {
       __m128i one = _mm_set1_epi32(1);
       __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
       float *p = (float*)sp;
-      for (ui32 i = 0; i < count; i += 4, p += 4, dp += 4)
+      for ( ; count >= 4; count -= 4, p += 4, dp += 4)
       {
         __m128 vf = _mm_loadu_ps(p);
         vf = _mm_mul_ps(vf, d);                    // multiply
@@ -130,6 +149,26 @@ namespace ojph {
         val = _mm_or_si128(val, sign);
         _mm_storeu_si128((__m128i*)dp, val);
       }
+      if (count)
+      {
+        __m128 vf = _mm_loadu_ps(p);
+        vf = _mm_mul_ps(vf, d);                    // multiply
+        __m128i val = _mm_cvtps_epi32(vf);         // convert to int
+        __m128i sign = _mm_cmplt_epi32(val, zero); // get sign
+        val = _mm_xor_si128(val, sign);            // negate 1's complement
+        __m128i ones = _mm_and_si128(sign, one);
+        val = _mm_add_epi32(val, ones);            // 2's complement
+
+        __m128i c = _mm_set1_epi32((si32)count);
+        __m128i idx = _mm_set_epi32(3, 2, 1, 0);
+        __m128i mask = _mm_cmpgt_epi32(c, idx);
+        c = _mm_and_si128(val, mask);
+        tmax = _mm_or_si128(tmax, c);
+
+        sign = _mm_slli_epi32(sign, 31);
+        val = _mm_or_si128(val, sign);
+        _mm_storeu_si128((__m128i*)dp, val);
+      }
       _mm_storeu_si128((__m128i*)max_val, tmax);
     }
 
@@ -189,7 +228,7 @@ namespace ojph {
       __m128i one = _mm_set1_epi64x(1);
       __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
       __m128i *p = (__m128i*)sp;
-      for (ui32 i = 0; i < count; i += 2, p += 1, dp += 2)
+      for ( ; count >= 2; count -= 2, p += 1, dp += 2)
       {
         __m128i v = _mm_loadu_si128(p);
         __m128i sign = _mm_cmplt_epi32(v, zero);
@@ -203,6 +242,24 @@ namespace ojph {
         val = _mm_or_si128(val, sign);
         _mm_storeu_si128((__m128i*)dp, val);
       }
+      if (count)
+      {
+        __m128i v = _mm_loadu_si128(p);
+        __m128i sign = _mm_cmplt_epi32(v, zero);
+        sign = _mm_shuffle_epi32(sign, 0xF5);  // sign = sign[1,1,3,3];
+        __m128i val = _mm_xor_si128(v, sign);  // negate 1's complement
+        __m128i ones = _mm_and_si128(sign, one);
+        val = _mm_add_epi64(val, ones);        // 2's complement
+        sign = _mm_and_si128(sign, m0);
+        val = _mm_slli_epi64(val, (int)shift);
+
+        __m128i c = _mm_set_epi32(0, 0, (si32)0xFFFFFFFF, (si32)0xFFFFFFFF);
+        c = _mm_and_si128(val, c);
+        tmax = _mm_or_si128(tmax, c);
+
+        val = _mm_or_si128(val, sign);
+        _mm_storeu_si128((__m128i*)dp, val);
+      }
       _mm_storeu_si128((__m128i*)max_val, tmax);
     }
 
@@ -222,10 +279,10 @@ namespace ojph {
         __m128i val = _mm_and_si128(v, m1);
         val = _mm_srli_epi64(val, (int)shift);
         __m128i sign = _mm_cmplt_epi32(v, zero);
-        sign = _mm_shuffle_epi32(sign, 0xF5);  // sign = sign[1,1,3,3];
-        val = _mm_xor_si128(val, sign); // negate 1's complement
+        sign = _mm_shuffle_epi32(sign, 0xF5);      // sign = sign[1,1,3,3];
+        val = _mm_xor_si128(val, sign);            // negate 1's complement
         __m128i ones = _mm_and_si128(sign, one);
-        val = _mm_add_epi64(val, ones); // 2's complement
+        val = _mm_add_epi64(val, ones);            // 2's complement
         _mm_storeu_si128((__m128i*)p, val);
       }
     }

src/core/codestream/ojph_codestream_wasm.cpp

@@ -88,7 +88,7 @@ namespace ojph {
       v128_t one = wasm_i32x4_splat(1);
       v128_t tmax = wasm_v128_load(max_val);
       si32 *p = (si32*)sp;
-      for (ui32 i = 0; i < count; i += 4, p += 4, dp += 4)
+      for ( ; count >= 4; count -= 4, p += 4, dp += 4)
       {
         v128_t v = wasm_v128_load(p);
         v128_t sign = wasm_i32x4_lt(v, zero);
@@ -101,6 +101,25 @@ namespace ojph {
         val = wasm_v128_or(val, sign);
         wasm_v128_store(dp, val);
       }
+      if (count)
+      {
+        v128_t v = wasm_v128_load(p);
+        v128_t sign = wasm_i32x4_lt(v, zero);
+        v128_t val = wasm_v128_xor(v, sign); // negate 1's complement
+        v128_t ones = wasm_v128_and(sign, one);
+        val = wasm_i32x4_add(val, ones);     // 2's complement
+        sign = wasm_v128_and(sign, m0);
+        val = wasm_i32x4_shl(val, shift);
+
+        v128_t c = wasm_i32x4_splat((si32)count);
+        v128_t idx = wasm_i32x4_make(0, 1, 2, 3);
+        v128_t mask = wasm_i32x4_gt(c, idx);
+        c = wasm_v128_and(val, mask);
+        tmax = wasm_v128_or(tmax, c);
+
+        val = wasm_v128_or(val, sign);
+        wasm_v128_store(dp, val);
+      }
       wasm_v128_store(max_val, tmax);
     }
 
@@ -117,7 +136,7 @@ namespace ojph {
       v128_t one = wasm_i32x4_splat(1);
       v128_t tmax = wasm_v128_load(max_val);
       float *p = (float*)sp;
-      for (ui32 i = 0; i < count; i += 4, p += 4, dp += 4)
+      for ( ; count >= 4; count -= 4, p += 4, dp += 4)
       {
         v128_t vf = wasm_v128_load(p);
         vf = wasm_f32x4_mul(vf, d);                   // multiply
@@ -131,6 +150,26 @@ namespace ojph {
         val = wasm_v128_or(val, sign);
         wasm_v128_store(dp, val);
       }
+      if (count)
+      {
+        v128_t vf = wasm_v128_load(p);
+        vf = wasm_f32x4_mul(vf, d);                   // multiply
+        v128_t val = wasm_i32x4_trunc_sat_f32x4(vf);  // convert to signed int
+        v128_t sign = wasm_i32x4_lt(val, zero);       // get sign
+        val = wasm_v128_xor(val, sign);               // negate 1's complement
+        v128_t ones = wasm_v128_and(sign, one);
+        val = wasm_i32x4_add(val, ones);              // 2's complement
+
+        v128_t c = wasm_i32x4_splat((si32)count);
+        v128_t idx = wasm_i32x4_make(0, 1, 2, 3);
+        v128_t mask = wasm_i32x4_gt(c, idx);
+        c = wasm_v128_and(val, mask);
+        tmax = wasm_v128_or(tmax, c);
+
+        sign = wasm_i32x4_shl(sign, 31);
+        val = wasm_v128_or(val, sign);
+        wasm_v128_store(dp, val);
+      }
       wasm_v128_store(max_val, tmax);
     }
 
@@ -190,7 +229,7 @@ namespace ojph {
       v128_t one = wasm_i64x2_splat(1);
       v128_t tmax = wasm_v128_load(max_val);
       si64 *p = (si64*)sp;
-      for (ui32 i = 0; i < count; i += 2, p += 2, dp += 2)
+      for ( ; count >= 2; count -= 2, p += 2, dp += 2)
       {
         v128_t v = wasm_v128_load(p);
         v128_t sign = wasm_i64x2_lt(v, zero);
@@ -203,6 +242,24 @@ namespace ojph {
         val = wasm_v128_or(val, sign);
         wasm_v128_store(dp, val);
       }
+      if (count)
+      {
+        v128_t v = wasm_v128_load(p);
+        v128_t sign = wasm_i64x2_lt(v, zero);
+        v128_t val = wasm_v128_xor(v, sign); // negate 1's complement
+        v128_t ones = wasm_v128_and(sign, one);
+        val = wasm_i64x2_add(val, ones);     // 2's complement
+        sign = wasm_v128_and(sign, m0);
+        val = wasm_i64x2_shl(val, shift);
+
+        v128_t c = wasm_i32x4_make((si32)0xFFFFFFFF, (si32)0xFFFFFFFF, 0, 0);
+        c = wasm_v128_and(val, c);
+        tmax = wasm_v128_or(tmax, c);
+
+        val = wasm_v128_or(val, sign);
+        wasm_v128_store(dp, val);
+      }
+
       wasm_v128_store(max_val, tmax);
     }   
 

src/core/common/ojph_version.h

@@ -35,4 +35,4 @@
 
 #define OPENJPH_VERSION_MAJOR 0
 #define OPENJPH_VERSION_MINOR 21
-#define OPENJPH_VERSION_PATCH 1
+#define OPENJPH_VERSION_PATCH 2