+add SSE4.1 optimizations of function SynetTiledScale2D32f.

ermig1979 · Dec 19, 2024 · d546bb9 · d546bb9
1 parent 12ad5a8
commit d546bb9
Show file tree

Hide file tree

Showing 5 changed files with 54 additions and 7 deletions.
diff --git a/docs/2025.html b/docs/2025.html
@@ -41,7 +41,7 @@ <h4>Algorithms</h4>
 <h5>New features</h5>
 <ul>
  <li>Parameter add in function SimdSynetMergedConvolution16bInit.</li>
- <li>Base implementation of function SynetTiledScale2D32f.</li>
+ <li>Base implementation, SSE4.1 optimizations of function SynetTiledScale2D32f.</li>
 </ul>
 <h5>Improving</h5>
 <ul>

diff --git a/src/Simd/SimdLib.cpp b/src/Simd/SimdLib.cpp
@@ -6054,7 +6054,7 @@ SIMD_API void SimdSynetTiledScale2D32f(const float* src, size_t channels, size_t
     SIMD_EMPTY();
 #if defined(SIMD_SYNET_ENABLE)
     typedef void(*SimdSynetTiledScale2D32fPtr) (const float* src, size_t channels, size_t height, size_t width, SimdTensorFormatType format, const float* ver, const float* hor, float* dst);
-    const static SimdSynetTiledScale2D32fPtr simdSynetTiledScale2D32f = SIMD_FUNC0(SynetTiledScale2D32f);// , SIMD_AVX512BW_FUNC, SIMD_AVX2_FUNC, SIMD_SSE41_FUNC);
+    const static SimdSynetTiledScale2D32fPtr simdSynetTiledScale2D32f = SIMD_FUNC1(SynetTiledScale2D32f, SIMD_SSE41_FUNC);// , SIMD_AVX512BW_FUNC, SIMD_AVX2_FUNC);
 
     simdSynetTiledScale2D32f(src, channels, height, width, format, ver, hor, dst);
 #else

diff --git a/src/Simd/SimdSse41.h b/src/Simd/SimdSse41.h
@@ -524,6 +524,8 @@ namespace Simd
 
         void SynetScaleLayerForward(const float* src, const float* scale, const float* bias, size_t channels, size_t height, size_t width, float* dst, SimdTensorFormatType format, SimdSynetCompatibilityType compatibility);
 
+        void SynetTiledScale2D32f(const float* src, size_t channels, size_t height, size_t width, SimdTensorFormatType format, const float* ver, const float* hor, float* dst);
+
         void TextureBoostedSaturatedGradient(const uint8_t* src, size_t srcStride, size_t width, size_t height,
             uint8_t saturation, uint8_t boost, uint8_t* dx, size_t dxStride, uint8_t* dy, size_t dyStride);
 

diff --git a/src/Simd/SimdSse41Synet.cpp b/src/Simd/SimdSse41Synet.cpp
@@ -455,6 +455,51 @@ namespace Simd
             else
                 assert(0);
         }
+
+        //-------------------------------------------------------------------------------------------------
+
+        void SynetTiledScale2D32f(const float* src, size_t channels, size_t height, size_t width, SimdTensorFormatType format, const float* ver, const float* hor, float* dst)
+        {
+            if (format == SimdTensorFormatNchw)
+            {
+                size_t widthF = AlignLo(width, F);
+                for (size_t c = 0; c < channels; ++c)
+                {
+                    for (size_t y = 0; y < height; ++y)
+                    {
+                        __m128 _hor = _mm_set1_ps(hor[y]);
+                        size_t x = 0;
+                        for (; x < widthF; x += F)
+                            _mm_storeu_ps(dst + x, _mm_mul_ps(_mm_loadu_ps(src + x), _mm_mul_ps(_mm_loadu_ps(ver + x), _hor)));
+                        for (; x < width; x += 1)
+                            _mm_store_ss(dst + x, _mm_mul_ss(_mm_load_ss(src + x), _mm_mul_ss(_mm_load_ss(ver + x), _hor)));
+                        src += width, dst += width;
+                    }
+                    hor += height;
+                    ver += width;
+                }
+            }
+            else if (format == SimdTensorFormatNhwc)
+            {
+                size_t channelsF = AlignLo(channels, F);
+                for (size_t y = 0; y < height; ++y)
+                {
+                    const float* pVer = ver;
+                    for (size_t x = 0; x < width; ++x)
+                    {
+                        size_t c = 0;
+                        for (; c < channelsF; c += F)
+                            _mm_storeu_ps(dst + c, _mm_mul_ps(_mm_loadu_ps(src + c), _mm_mul_ps(_mm_loadu_ps(pVer + c), _mm_loadu_ps(hor + c))));
+                        for (; c < channels; c += 1)
+                            _mm_store_ss(dst + c, _mm_mul_ss(_mm_load_ss(src + c), _mm_mul_ss(_mm_load_ss(pVer + c), _mm_load_ss(hor + c))));
+                        src += channels, dst += channels, pVer += channels;
+                    }
+                    hor += channels;
+                }
+            }
+            else
+                assert(0);
+        }
    }
 #endif
 }
diff --git a/src/Test/TestSynet.cpp b/src/Test/TestSynet.cpp
@@ -431,11 +431,11 @@ namespace Test
         if (TestBase())
             result = result && SynetTiledScale2D32fAutoTest(FUNC_TS2D32F(Simd::Base::SynetTiledScale2D32f), FUNC_TS2D32F(SimdSynetTiledScale2D32f));
 
-//#ifdef SIMD_SSE41_ENABLE
-//        if (Simd::Sse41::Enable && TestSse41())
-//            result = result && SynetTiledScale2D32fAutoTest(FUNC_TS2D32F(Simd::Sse41::SynetTiledScale2D32f), FUNC_TS2D32F(SimdSynetTiledScale2D32f));
-//#endif 
-//
+#ifdef SIMD_SSE41_ENABLE
+        if (Simd::Sse41::Enable && TestSse41())
+            result = result && SynetTiledScale2D32fAutoTest(FUNC_TS2D32F(Simd::Sse41::SynetTiledScale2D32f), FUNC_TS2D32F(SimdSynetTiledScale2D32f));
+#endif 
+
 //#ifdef SIMD_AVX2_ENABLE
 //        if (Simd::Avx2::Enable && TestAvx2())
 //            result = result && SynetTiledScale2D32fAutoTest(FUNC_TS2D32F(Simd::Avx2::SynetTiledScale2D32f), FUNC_TS2D32F(SimdSynetTiledScale2D32f));