*improve AVX2 optimizations of class ResizerFloatBilinear (part 10: c…

…ase of large scale, channels >= 8).
ermig1979 · Jan 17, 2025 · 4edcec4 · 4edcec4
1 parent 87bb494
commit 4edcec4
Show file tree

Hide file tree

Showing 2 changed files with 29 additions and 7 deletions.
diff --git a/src/Simd/SimdAvx2ResizerBilinear.cpp b/src/Simd/SimdAvx2ResizerBilinear.cpp
@@ -1009,11 +1009,6 @@ namespace Simd
             }
             else
             {
-                if (cn > 7)
-                {
-                    Sse41::ResizerFloatBilinear::Run(src, srcStride, dst, dstStride);
-                    return;
-                }
                 for (size_t dy = 0; dy < _param.dstH; dy++, dst += dstStride)
                 {
                     __m256 fy1 = _mm256_set1_ps(_ay[dy]);
@@ -1212,6 +1207,29 @@ namespace Simd
                             }
                         }
                     }
+                    else
+                    {
+                        for (size_t dx = 0; dx < dw; dx++)
+                        {
+                            size_t os = _ix[dx], eF = os + cnF, od = dx * cn;
+                            __m256 fx1 = _mm256_set1_ps(_ax[dx]);
+                            __m256 fx0 = _mm256_sub_ps(_1, fx1);
+                            for (; os < eF; os += F, od += F)
+                            {
+                                __m256 r0 = _mm256_fmadd_ps(_mm256_loadu_ps(src0 + os), fx0, _mm256_mul_ps(_mm256_loadu_ps(src0 + os + cn), fx1));
+                                __m256 r1 = _mm256_fmadd_ps(_mm256_loadu_ps(src1 + os), fx0, _mm256_mul_ps(_mm256_loadu_ps(src1 + os + cn), fx1));
+                                _mm256_storeu_ps(dst + od, _mm256_fmadd_ps(r0, fy0, _mm256_mul_ps(r1, fy1)));
+                            }
+                            if (cnTF)
+                            {
+                                os += cnLF;
+                                od += cnLF;
+                                __m256 r0 = _mm256_fmadd_ps(_mm256_loadu_ps(src0 + os), fx0, _mm256_mul_ps(_mm256_loadu_ps(src0 + os + cn), fx1));
+                                __m256 r1 = _mm256_fmadd_ps(_mm256_loadu_ps(src1 + os), fx0, _mm256_mul_ps(_mm256_loadu_ps(src1 + os + cn), fx1));
+                                _mm256_storeu_ps(dst + od, _mm256_fmadd_ps(r0, fy0, _mm256_mul_ps(r1, fy1)));
+                            }
+                        }
+                    }
                 }
             }
         }

diff --git a/src/Simd/SimdAvx512bwResizerBilinear.cpp b/src/Simd/SimdAvx512bwResizerBilinear.cpp
@@ -732,9 +732,9 @@ namespace Simd
 
         void ResizerFloatBilinear::Run(const float* src, size_t srcStride, float* dst, size_t dstStride)
         {
+            size_t cn = _param.channels;
             if (_rowBuf)
             {
-                size_t cn = _param.channels;
                 size_t rs = _param.dstW * cn;
                 float* pbx[2] = { _bx[0].data, _bx[1].data };
                 int32_t prev = -2;
@@ -861,7 +861,11 @@ namespace Simd
             }
             else
             {
-                Avx2::ResizerFloatBilinear::Run(src, srcStride, dst, dstStride);
+                if (cn > 0)
+                {
+                    Avx2::ResizerFloatBilinear::Run(src, srcStride, dst, dstStride);
+                    return;
+                }
             }
         }