*improve SSE4.1 optimizations of class ResizerFloatBilinear (part 7: …

…case of large scale, channels = 3).
ermig1979 · Jan 15, 2025 · 51b4815 · 51b4815
1 parent 218fa22
commit 51b4815
Show file tree

Hide file tree

Showing 3 changed files with 28 additions and 4 deletions.
diff --git a/src/Simd/SimdBaseResizerBilinear.cpp b/src/Simd/SimdBaseResizerBilinear.cpp
@@ -334,11 +334,10 @@ namespace Simd
         ResizerFloatBilinear::ResizerFloatBilinear(const ResParam & param)
             : Resizer(param)
         {
-            _rowBuf = _param.align < 16 || ((_param.channels < 4 && _param.align > 16) || _param.channels == 3) || _param.dstH >= _param.srcH;
+            _rowBuf = _param.align < 16 || (_param.channels < 4 && _param.align > 16) || _param.dstH >= _param.srcH;
 #if defined(SIMD_ARM_ENABLE) || defined(SIMD_ARM64_ENABLE)
             _rowBuf = true;
 #endif
-
             _ay.Resize(_param.dstH, false, _param.align);
             _iy.Resize(_param.dstH, false, _param.align);
             EstimateIndexAlpha(_param, _param.srcH, _param.dstH, 1, 1, _iy.data, _ay.data);

diff --git a/src/Simd/SimdSse41ResizerBilinear.cpp b/src/Simd/SimdSse41ResizerBilinear.cpp
@@ -598,7 +598,7 @@ namespace Simd
         void ResizerFloatBilinear::Run(const float* src, size_t srcStride, float* dst, size_t dstStride)
         {
             size_t cn = _param.channels, cnF = AlignLo(cn, F), cnT = cn - cnF, cnL = cnT - F;
-            size_t dw = _param.dstW, dw2= AlignLo(dw, 2), dw4 = AlignLo(dw, 4);
+            size_t dw = _param.dstW, dw2 = AlignLo(dw, 2), dw4 = AlignLo(dw, 4), dw1 = dw - 1;
             __m128 _1 = _mm_set1_ps(1.0f);
             if (_rowBuf)
             {
@@ -770,6 +770,31 @@ namespace Simd
                             StoreHalf<0>(dst + od, _mm_add_ps(_mm_mul_ps(r0, fy0), _mm_mul_ps(r1, fy1)));
                         }
                     }
+                    else if (cn == 3)
+                    {
+                        size_t dx = 0, od = 0;
+                        for (; dx < dw1; dx += 1, od += 3)
+                        {
+                            size_t os = _ix[dx];
+                            __m128 fx1 = _mm_set1_ps(_ax[dx]);
+                            __m128 fx0 = _mm_sub_ps(_1, fx1);
+                            __m128 r0 = _mm_add_ps(_mm_mul_ps(_mm_loadu_ps(src0 + os), fx0), _mm_mul_ps(_mm_loadu_ps(src0 + os + 3), fx1));
+                            __m128 r1 = _mm_add_ps(_mm_mul_ps(_mm_loadu_ps(src1 + os), fx0), _mm_mul_ps(_mm_loadu_ps(src1 + os + 3), fx1));
+                            _mm_storeu_ps(dst + od, _mm_add_ps(_mm_mul_ps(r0, fy0), _mm_mul_ps(r1, fy1)));
+                        }
+                        if (dx < dw)
+                        {
+                            size_t os = _ix[dx];
+                            __m128 fx1 = _mm_set1_ps(_ax[dx]);
+                            __m128 fx0 = _mm_sub_ps(_1, fx1);
+                            for (size_t ed = od + 3; od < ed; od++, os++)
+                            {
+                                __m128 r0 = _mm_add_ps(_mm_mul_ps(_mm_load_ss(src0 + os), fx0), _mm_mul_ps(_mm_load_ss(src0 + os + 3), fx1));
+                                __m128 r1 = _mm_add_ps(_mm_mul_ps(_mm_load_ss(src1 + os), fx0), _mm_mul_ps(_mm_load_ss(src1 + os + 3), fx1));
+                                _mm_store_ss(dst + od, _mm_add_ps(_mm_mul_ps(r0, fy0), _mm_mul_ps(r1, fy1)));
+                            }
+                        }
+                    }
                     else
                     {
                         for (size_t dx = 0; dx < dw; dx++)

diff --git a/src/Test/TestResize.cpp b/src/Test/TestResize.cpp
@@ -245,7 +245,7 @@ namespace Test
     bool ResizerAutoTest(const FuncRS & f1, const FuncRS & f2)
     {
         bool result = true;
-        result = result && ResizerAutoTest(SimdResizeMethodBilinear, SimdResizeChannelFloat, 2, f1, f2);
+
 #if 1
 #if defined(SIMD_X64_ENABLE) 
         result = result && ResizerAutoTest(SimdResizeMethodBilinear, SimdResizeChannelFloat, 64, f1, f2);