GDALInterleave2Byte(): provide SSE2 implementation for ICC and MSVC

rouault · Jan 5, 2025 · 9881325 · 9881325
1 parent e751efd
commit 9881325
Show file tree

Hide file tree

Showing 3 changed files with 58 additions and 4 deletions.
diff --git a/autotest/cpp/test_gdal.cpp b/autotest/cpp/test_gdal.cpp
@@ -5011,7 +5011,21 @@ TEST_F(test_gdal, GDALTranspose2D_Byte_optims)
         }
     }
 
-    // Optim H = 2
+    // Optim H = 2 with W < 16
+    {
+        constexpr int W = 15;
+        constexpr int H = 2;
+        GDALTranspose2D(in.data(), GDT_Byte, out.data(), GDT_Byte, W, H);
+        for (int y = 0; y < H; ++y)
+        {
+            for (int x = 0; x < W; ++x)
+            {
+                EXPECT_EQ(out[x * H + y], in[y * W + x]);
+            }
+        }
+    }
+
+    // Optim H = 2 with W >= 16
     {
         constexpr int W = 19;
         constexpr int H = 2;

diff --git a/gcore/rasterio.cpp b/gcore/rasterio.cpp
@@ -5746,6 +5746,44 @@ static void GDALTranspose2D(const void *pSrc, GDALDataType eSrcType, DST *pDst,
 /*                      GDALInterleave2Byte()                           */
 /************************************************************************/
 
+#if defined(HAVE_SSE2) &&                                                      \
+    (!defined(__GNUC__) || defined(__INTEL_CLANG_COMPILER))
+
+// ICC autovectorizer doesn't do a good job at generating good SSE code,
+// at least with icx 2024.0.2.20231213, but it nicely unrolls the below loop.
+#if defined(__GNUC__)
+__attribute__((noinline))
+#endif
+static void
+GDALInterleave2Byte(const uint8_t *CPL_RESTRICT pSrc,
+                    uint8_t *CPL_RESTRICT pDst, size_t nIters)
+{
+    size_t i = 0;
+    constexpr size_t VALS_PER_ITER = 16;
+    for (i = 0; i + VALS_PER_ITER <= nIters; i += VALS_PER_ITER)
+    {
+        __m128i xmm0 =
+            _mm_loadu_si128(reinterpret_cast<__m128i const *>(pSrc + i));
+        __m128i xmm1 = _mm_loadu_si128(
+            reinterpret_cast<__m128i const *>(pSrc + i + nIters));
+        _mm_storeu_si128(reinterpret_cast<__m128i *>(pDst + 2 * i),
+                         _mm_unpacklo_epi8(xmm0, xmm1));
+        _mm_storeu_si128(
+            reinterpret_cast<__m128i *>(pDst + 2 * i + VALS_PER_ITER),
+            _mm_unpackhi_epi8(xmm0, xmm1));
+    }
+#if defined(__clang__)
+#pragma clang loop vectorize(disable)
+#endif
+    for (; i < nIters; ++i)
+    {
+        pDst[2 * i + 0] = pSrc[i + 0 * nIters];
+        pDst[2 * i + 1] = pSrc[i + 1 * nIters];
+    }
+}
+
+#else
+
 #if defined(__GNUC__) && !defined(__clang__)
 __attribute__((optimize("tree-vectorize")))
 #endif
@@ -5756,7 +5794,7 @@ static void
 GDALInterleave2Byte(const uint8_t *CPL_RESTRICT pSrc,
                     uint8_t *CPL_RESTRICT pDst, size_t nIters)
 {
-#if defined(__clang__)
+#if defined(__clang__) && !defined(__INTEL_CLANG_COMPILER)
 #pragma clang loop vectorize(enable)
 #endif
     for (size_t i = 0; i < nIters; ++i)
@@ -5766,6 +5804,8 @@ GDALInterleave2Byte(const uint8_t *CPL_RESTRICT pSrc,
     }
 }
 
+#endif
+
 /************************************************************************/
 /*                      GDALInterleave4Byte()                           */
 /************************************************************************/
@@ -5780,7 +5820,7 @@ static void
 GDALInterleave4Byte(const uint8_t *CPL_RESTRICT pSrc,
                     uint8_t *CPL_RESTRICT pDst, size_t nIters)
 {
-#if defined(__clang__)
+#if defined(__clang__) && !defined(__INTEL_CLANG_COMPILER)
 #pragma clang loop vectorize(enable)
 #endif
     for (size_t i = 0; i < nIters; ++i)

diff --git a/gcore/rasterio_ssse3.cpp b/gcore/rasterio_ssse3.cpp
@@ -286,7 +286,7 @@ static void
 GDALInterleave3Byte_SSSE3(const uint8_t *CPL_RESTRICT pSrc,
                           uint8_t *CPL_RESTRICT pDst, size_t nIters)
 {
-#if defined(__clang__)
+#if defined(__clang__) && !defined(__INTEL_CLANG_COMPILER)
 #pragma clang loop vectorize(enable)
 #endif
     for (size_t i = 0; i < nIters; ++i)