*fix bug: Error in Base implementation, SSE4.1, AVX2, AVX-512BW optim…

…izations of function Float32ToBFloat16.
ermig1979 · Jul 16, 2024 · 4d745b6 · 4d745b6
1 parent 89783e2
commit 4d745b6
Show file tree

Hide file tree

Showing 15 changed files with 53 additions and 267 deletions.
diff --git a/docs/2024.html b/docs/2024.html
@@ -42,21 +42,20 @@ <h5>New features</h5>
  <li>Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetRelu16b.</li>
  <li>API of SynetAdd16b framework.</li>
  <li>Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of class SynetAdd16bUniform.</li>
- <li>Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Float32ToBFloat16NearestEven.</li>
 </ul>
 <h5>Bug fixing</h5>
 <ul>
  <li>Error in Base implementation of class SynetMergedConvolution16bCdc.</li>
  <li>Error in Base implementation of class SynetMergedConvolution16bDc.</li>
  <li>Error in Base implementation of class SynetInnerProduct16bGemmNN.</li>
+ <li>Error in Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Float32ToBFloat16.</li>
 </ul>
 
 <h4>Test framework</h4>
 <h5>New features</h5>
 <ul>
  <li>Tests for verifying functionality of function SynetRelu16b.</li>
  <li>Tests for verifying functionality of SynetAdd16b framework.</li>
- <li>Tests for verifying functionality of function Float32ToBFloat16NearestEven.</li>
 </ul>
 
 <a href="#HOME">Home</a>

diff --git a/docs/help/group__bfloat16.html b/docs/help/group__bfloat16.html
diff --git a/src/Simd/SimdAmxBf16.h b/src/Simd/SimdAmxBf16.h
@@ -31,7 +31,7 @@ namespace Simd
 #ifdef SIMD_AMXBF16_ENABLE    
     namespace AmxBf16
     {
-        void Float32ToBFloat16NearestEven(const float* src, size_t size, uint16_t* dst);
+        void Float32ToBFloat16(const float* src, size_t size, uint16_t* dst);
 
         void ChangeColors(const uint8_t* src, size_t srcStride, size_t width, size_t height, const uint8_t* colors, uint8_t* dst, size_t dstStride);
 

diff --git a/src/Simd/SimdAmxBf16BFloat16.cpp b/src/Simd/SimdAmxBf16BFloat16.cpp
@@ -30,20 +30,20 @@ namespace Simd
 #ifdef SIMD_AMXBF16_ENABLE    
     namespace AmxBf16
     {
-        void Float32ToBFloat16NearestEven(const float* src, size_t size, uint16_t* dst)
+        void Float32ToBFloat16(const float* src, size_t size, uint16_t* dst)
         {
             size_t size32 = AlignLo(size, 32);
             __mmask16 srcMask[2];
             __mmask32 dstMask[1];
             size_t i = 0;
             for (; i < size32; i += 32)
-                Float32ToBFloat16NearestEven<false, false>(src + i, dst + i, srcMask, dstMask);
+                Float32ToBFloat16<false, false>(src + i, dst + i, srcMask, dstMask);
             if (size32 < size)
             {
                 srcMask[0] = TailMask16(size - size32 - F * 0);
                 srcMask[1] = TailMask16(size - size32 - F * 1);
                 dstMask[0] = TailMask32(size - size32);
-                Float32ToBFloat16NearestEven<false, true>(src + i, dst + i, srcMask, dstMask);
+                Float32ToBFloat16<false, true>(src + i, dst + i, srcMask, dstMask);
             }
         }
     }

diff --git a/src/Simd/SimdAvx2BFloat16.cpp b/src/Simd/SimdAvx2BFloat16.cpp
@@ -61,35 +61,6 @@ namespace Simd
 
         //---------------------------------------------------------------------------------------------
 
-        void Float32ToBFloat16NearestEven(const float* src, size_t size, uint16_t* dst)
-        {
-            size_t size16 = Simd::AlignLo(size, 16);
-            size_t size8 = Simd::AlignLo(size, 8);
-            size_t size4 = Simd::AlignLo(size, 4);
-            size_t i = 0;
-            for (; i < size16; i += 16)
-            {
-                __m256i d0 = Float32ToBFloat16NearestEven(_mm256_loadu_ps(src + i + 0));
-                __m256i d1 = Float32ToBFloat16NearestEven(_mm256_loadu_ps(src + i + 8));
-                _mm256_storeu_si256((__m256i*)(dst + i), _mm256_permute4x64_epi64(_mm256_packus_epi32(d0, d1), 0xD8));
-            }
-            for (; i < size8; i += 8)
-            {
-                __m128i d0 = Sse41::Float32ToBFloat16NearestEven(_mm_loadu_ps(src + i + 0));
-                __m128i d1 = Sse41::Float32ToBFloat16NearestEven(_mm_loadu_ps(src + i + 4));
-                _mm_storeu_si128((__m128i*)(dst + i), _mm_packus_epi32(d0, d1));
-            }
-            for (; i < size4; i += 4)
-            {
-                __m128i d0 = Sse41::Float32ToBFloat16NearestEven(_mm_loadu_ps(src + i + 0));
-                _mm_storel_epi64((__m128i*)(dst + i), _mm_packus_epi32(d0, Sse41::K_ZERO));
-            }
-            for (; i < size; ++i)
-                dst[i] = Base::Float32ToBFloat16NearestEven(src[i]);
-        }
-
-        //---------------------------------------------------------------------------------------------
-
         void BFloat16ToFloat32(const uint16_t* src, size_t size, float* dst)
         {
             size_t size16 = Simd::AlignLo(size, 16);

diff --git a/src/Simd/SimdAvx512bwBFloat16.cpp b/src/Simd/SimdAvx512bwBFloat16.cpp
@@ -49,25 +49,6 @@ namespace Simd
 
         //---------------------------------------------------------------------------------------------
 
-        void Float32ToBFloat16NearestEven(const float* src, size_t size, uint16_t* dst)
-        {
-            size_t size32 = AlignLo(size, 32);
-            __mmask16 srcMask[2];
-            __mmask32 dstMask[1];
-            size_t i = 0;
-            for (; i < size32; i += 32)
-                Float32ToBFloat16NearestEven<false, false>(src + i, dst + i, srcMask, dstMask);
-            if (size32 < size)
-            {
-                srcMask[0] = TailMask16(size - size32 - F * 0);
-                srcMask[1] = TailMask16(size - size32 - F * 1);
-                dstMask[0] = TailMask32(size - size32);
-                Float32ToBFloat16NearestEven<false, true>(src + i, dst + i, srcMask, dstMask);
-            }
-        }
-
-        //---------------------------------------------------------------------------------------------
-
         template<bool align, bool mask> SIMD_INLINE void BFloat16ToFloat32(const uint16_t* src, float* dst, __mmask32 srcMask[1], __mmask16 dstMask[2])
         {
             __m512i _src = Load<align, mask>(src, srcMask[0]);