Merge changes I6dfedfd5,I2376e2dc into main

* changes: rework AddVectorEq_SSE2 rework AddVector_SSE2
{ios,xcframework}build.sh: fix compilation w/Xcode 16
2024-12-26 13:48:21 +01:00 · 2024-11-15 02:58:10 +00:00 · 2024-11-14 20:26:57 +00:00 · 2024-11-14 12:21:39 -08:00 · 2024-11-14 12:21:33 -08:00
3 changed files with 64 additions and 24 deletions
--- a/iosbuild.sh
+++ b/iosbuild.sh
@ -53,7 +53,7 @@ DEMUXLIBLIST=''
 if [[ -z "${SDK}" ]]; then
  echo "iOS SDK not available"
  exit 1
-elif [[ ${SDK%%.*} -gt 8 ]]; then
+elif [[ ${SDK%%.*} -gt 8 && "${XCODE%%.*}" -lt 16 ]]; then
  EXTRA_CFLAGS="-fembed-bitcode"
 elif [[ ${SDK%%.*} -le 6 ]]; then
  echo "You need iOS SDK version 6.0 or above"
--- a/src/dsp/lossless_enc_sse2.c
+++ b/src/dsp/lossless_enc_sse2.c
@ -175,64 +175,102 @@ static void CollectColorRedTransforms_SSE2(const uint32_t* WEBP_RESTRICT argb,

 // Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
 // that's ok since the histogram values are less than 1<<28 (max picture size).
-#define LINE_SIZE 16    // 8 or 16
 static void AddVector_SSE2(const uint32_t* WEBP_RESTRICT a,
                           const uint32_t* WEBP_RESTRICT b,
                           uint32_t* WEBP_RESTRICT out, int size) {
-  int i;
-  for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
+  int i = 0;
+  int aligned_size = size & ~15;
+  // Size is, at minimum, NUM_DISTANCE_CODES (40) and may be as large as
+  // NUM_LITERAL_CODES (256) + NUM_LENGTH_CODES (24) + (0 or a non-zero power of
+  // 2). See the usage in VP8LHistogramAdd().
+  assert(size >= 16);
+  assert(size % 2 == 0);
+
+  do {
    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
-#if (LINE_SIZE == 16)
    const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i +  8]);
    const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
-#endif
    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i +  0]);
    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i +  4]);
-#if (LINE_SIZE == 16)
    const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i +  8]);
    const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
-#endif
    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
-#if (LINE_SIZE == 16)
    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
-#endif
+    i += 16;
+  } while (i != aligned_size);
+
+  if ((size & 8) != 0) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
+    _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
+    _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
+    i += 8;
  }
-  for (; i < size; ++i) {
-    out[i] = a[i] + b[i];
+
+  size &= 7;
+  if (size == 4) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i]);
+    _mm_storeu_si128((__m128i*)&out[i], _mm_add_epi32(a0, b0));
+  } else if (size == 2) {
+    const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[i]);
+    const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[i]);
+    _mm_storel_epi64((__m128i*)&out[i], _mm_add_epi32(a0, b0));
  }
 }

 static void AddVectorEq_SSE2(const uint32_t* WEBP_RESTRICT a,
                             uint32_t* WEBP_RESTRICT out, int size) {
-  int i;
-  for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
+  int i = 0;
+  int aligned_size = size & ~15;
+  // Size is, at minimum, NUM_DISTANCE_CODES (40) and may be as large as
+  // NUM_LITERAL_CODES (256) + NUM_LENGTH_CODES (24) + (0 or a non-zero power of
+  // 2). See the usage in VP8LHistogramAdd().
+  assert(size >= 16);
+  assert(size % 2 == 0);
+
+  do {
    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
-#if (LINE_SIZE == 16)
    const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i +  8]);
    const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
-#endif
    const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i +  0]);
    const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i +  4]);
-#if (LINE_SIZE == 16)
    const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i +  8]);
    const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
-#endif
    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
-#if (LINE_SIZE == 16)
    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
-#endif
+    i += 16;
+  } while (i != aligned_size);
+
+  if ((size & 8) != 0) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]);
+    _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
+    _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
+    i += 8;
  }
-  for (; i < size; ++i) {
-    out[i] += a[i];
+
+  size &= 7;
+  if (size == 4) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i]);
+    _mm_storeu_si128((__m128i*)&out[i], _mm_add_epi32(a0, b0));
+  } else if (size == 2) {
+    const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[i]);
+    const __m128i b0 = _mm_loadl_epi64((const __m128i*)&out[i]);
+    _mm_storel_epi64((__m128i*)&out[i], _mm_add_epi32(a0, b0));
  }
 }
-#undef LINE_SIZE

 //------------------------------------------------------------------------------
 // Entropy
--- a/xcframeworkbuild.sh
+++ b/xcframeworkbuild.sh
@ -172,7 +172,9 @@ for (( i = 0; i < $NUM_PLATFORMS; ++i )); do
    CFLAGS="-pipe -isysroot ${SDKROOT} -O3 -DNDEBUG"
    case "${PLATFORM}" in
      iPhone*)
-        CFLAGS+=" -fembed-bitcode"
+        if [[ "${XCODE%%.*}" -lt 16 ]]; then
+          CFLAGS+=" -fembed-bitcode"
+        fi
        CFLAGS+=" -target ${ARCH}-apple-ios${IOS_MIN_VERSION}"
        [[ "${PLATFORM}" == *Simulator* ]] && CFLAGS+="-simulator"
        ;;
Author	SHA1	Message	Date
James Zern	0ab789e067	Merge changes I6dfedfd5,I2376e2dc into main * changes: rework AddVectorEq_SSE2 rework AddVector_SSE2	2024-11-15 02:58:10 +00:00
James Zern	0323645066	{ios,xcframework}build.sh: fix compilation w/Xcode 16 Don't use `-fembed-bitcode`, fixes: ld: warning: -bitcode_bundle is no longer supported and will be ignored ld: -mllvm and -bitcode_bundle (Xcode setting ENABLE_BITCODE=YES) cannot be used together Change-Id: I4ead0fc71da39bb5ec92c1f5ba467b95ad8b7461	2024-11-14 20:26:57 +00:00
James Zern	61e2cfdadd	rework AddVectorEq_SSE2 Take advantage of the known sizes used by VP8LHistogramAdd() and remove loop for the remainder. The loop was being auto-vectorized making the code larger and slower than the vectorized C code. For larger sizes the new code is ~3-4.5% faster than the old code with about the same improvement against the vectorized C code. For the minimal size (40), the new code is ~30% faster than the C and old SSE2 code. The LINE_SIZE==8 option is removed with this change. It had been set to 16 for its entire life and clang-16 was unrolling the LINE_SIZE==8 case by 2 in any case; they both profile similarly. Change-Id: I6dfedfd57474f44d15e2ce510a48e5252221077a	2024-11-14 12:21:39 -08:00
James Zern	7bda3deb89	rework AddVector_SSE2 Take advantage of the known sizes used by VP8LHistogramAdd() and remove loop for the remainder. The loop was being auto-vectorized making the code larger and slower than the vectorized C code. For larger sizes the new code is ~4-7% faster than the old code with about the same improvement against the vectorized C code. For the minimal size (40), the new code is ~30% faster than the C and old SSE2 code. The LINE_SIZE==8 option is removed with this change. It had been set to 16 for its entire life and clang-16 was unrolling the LINE_SIZE==8 case by 2 in any case; they both profile similarly. Change-Id: I2376e2dca3bffa38477b4a432f4c533419e3be0e	2024-11-14 12:21:33 -08:00