From 0b1eb4c84152188e48024cafc220af308d3a8694 Mon Sep 17 00:00:00 2001
From: yuanhecai <[EMAIL REDACTED]>
Date: Tue, 28 Oct 2025 17:59:19 +0800
Subject: [PATCH] loongarch: add Blit8888to8888PixelAlphaSwizzleLSX opt
---
CMakeLists.txt | 10 +++-
include/SDL3/SDL_intrin.h | 6 ++-
src/video/SDL_blit_A.c | 102 ++++++++++++++++++++++++++++++++++++++
3 files changed, 115 insertions(+), 3 deletions(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 71d5e3b72cdfc..39cd099d69959 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -932,7 +932,15 @@ if(SDL_ASSEMBLY)
cmake_pop_check_state()
if(COMPILER_SUPPORTS_LSX AND HAVE_LSXINTRIN_H)
- set_property(SOURCE "${SDL3_SOURCE_DIR}/src/video/yuv2rgb/yuv_rgb_lsx.c" APPEND PROPERTY COMPILE_OPTIONS "-mlsx")
+ set_property(SOURCE
+ "${SDL3_SOURCE_DIR}/src/video/yuv2rgb/yuv_rgb_lsx.c"
+ "${SDL3_SOURCE_DIR}/src/video/SDL_blit_A.c"
+ APPEND PROPERTY COMPILE_OPTIONS "-mlsx")
+
+ set_property(SOURCE
+ "${SDL3_SOURCE_DIR}/src/video/yuv2rgb/yuv_rgb_lsx.c"
+ "${SDL3_SOURCE_DIR}/src/video/SDL_blit_A.c"
+ PROPERTY SKIP_PRECOMPILE_HEADERS 1)
set(HAVE_LSX TRUE)
endif()
endif()
diff --git a/include/SDL3/SDL_intrin.h b/include/SDL3/SDL_intrin.h
index ab2dfeeee5846..802c1953b0713 100644
--- a/include/SDL3/SDL_intrin.h
+++ b/include/SDL3/SDL_intrin.h
@@ -281,12 +281,14 @@ _m_prefetch(void *__P)
* \sa SDL_TARGETING
*/
#define SDL_HAS_TARGET_ATTRIBS
-
+#elif defined(__loongarch64) && defined(__GNUC__) && (__GNUC__ >= 15)
+/* LoongArch requires GCC 15+ for target attribute support */
+# define SDL_HAS_TARGET_ATTRIBS
#elif defined(__clang__) && defined(__has_attribute)
# if __has_attribute(target)
# define SDL_HAS_TARGET_ATTRIBS
# endif
-#elif defined(__GNUC__) && (__GNUC__ + (__GNUC_MINOR__ >= 9) > 4) /* gcc >= 4.9 */
+#elif defined(__GNUC__) && !defined(__loongarch64) && (__GNUC__ + (__GNUC_MINOR__ >= 9) > 4) /* gcc >= 4.9 */
# define SDL_HAS_TARGET_ATTRIBS
#elif defined(__ICC) && __ICC >= 1600
# define SDL_HAS_TARGET_ATTRIBS
diff --git a/src/video/SDL_blit_A.c b/src/video/SDL_blit_A.c
index fc5ad1d737002..965b0b2d21f47 100644
--- a/src/video/SDL_blit_A.c
+++ b/src/video/SDL_blit_A.c
@@ -242,6 +242,103 @@ static void SDL_TARGETING("sse2") Blit888to888SurfaceAlphaSSE2(SDL_BlitInfo *inf
#endif
+#ifdef SDL_LSX_INTRINSICS
+
+static void SDL_TARGETING("lsx") Blit8888to8888PixelAlphaSwizzleLSX(SDL_BlitInfo *info)
+{
+ int width = info->dst_w;
+ int height = info->dst_h;
+ Uint8 *src = info->src;
+ int srcskip = info->src_skip;
+ Uint8 *dst = info->dst;
+ int dstskip = info->dst_skip;
+ const SDL_PixelFormatDetails *srcfmt = info->src_fmt;
+ const SDL_PixelFormatDetails *dstfmt = info->dst_fmt;
+ bool fill_alpha = !dstfmt->Amask;
+ Uint32 dstAmask, dstAshift;
+ const Uint8 offsets[] = {0, 0, 0, 0, 4, 4, 4, 4, 8, 8, 8, 8, 12, 12, 12, 12};
+
+ SDL_Get8888AlphaMaskAndShift(dstfmt, &dstAmask, &dstAshift);
+
+ const __m128i const_0xff00 = __lsx_vreplgr2vr_h(0xff00);
+ const __m128i const_128 = __lsx_vreplgr2vr_b((Uint8)128);
+ const __m128i const_32641 = __lsx_vreplgr2vr_h(32641);
+ const __m128i const_257 = __lsx_vreplgr2vr_h(257);
+
+ // The byte offsets for the start of each pixel
+ const __m128i mask_offsets = __lsx_vld(offsets, 0);
+
+ const __m128i convert_mask = __lsx_vadd_w(
+ __lsx_vreplgr2vr_w(
+ ((srcfmt->Rshift >> 3) << dstfmt->Rshift) |
+ ((srcfmt->Gshift >> 3) << dstfmt->Gshift) |
+ ((srcfmt->Bshift >> 3) << dstfmt->Bshift)),
+ mask_offsets);
+
+ const __m128i alpha_splat_mask = __lsx_vadd_b(__lsx_vreplgr2vr_b(srcfmt->Ashift >> 3), mask_offsets);
+ const __m128i alpha_fill_mask = __lsx_vreplgr2vr_w((int)dstAmask);
+
+ while (height--) {
+ int i = 0;
+
+ for (; i + 4 <= width; i += 4) {
+ __m128i src128 = __lsx_vld(src, 0);
+ __m128i dst128 = __lsx_vld(dst, 0);
+
+ __m128i srcA = __lsx_vshuf_b(src128, src128, alpha_splat_mask);
+ src128 = __lsx_vshuf_b(src128, src128, convert_mask);
+
+ src128 = __lsx_vor_v(src128, alpha_fill_mask);
+
+ __m128i srca_lo = __lsx_vilvl_b(srcA, srcA);
+ __m128i srca_hi = __lsx_vilvh_b(srcA, srcA);
+
+ srca_lo = __lsx_vxor_v(srca_lo, const_0xff00);
+ srca_hi = __lsx_vxor_v(srca_hi, const_0xff00);
+
+ src128 = __lsx_vsub_b(src128, const_128);
+ dst128 = __lsx_vsub_b(dst128, const_128);
+
+ __m128i tmp = __lsx_vilvl_b(dst128, src128);
+ __m128i dst_lo = __lsx_vsadd_h(__lsx_vmulwev_h_bu_b(srca_lo, tmp), __lsx_vmulwod_h_bu_b(srca_lo, tmp));
+ tmp = __lsx_vilvh_b(dst128, src128);
+ __m128i dst_hi = __lsx_vsadd_h(__lsx_vmulwev_h_bu_b(srca_hi, tmp), __lsx_vmulwod_h_bu_b(srca_hi, tmp));
+
+ dst_lo = __lsx_vadd_h(dst_lo, const_32641);
+ dst_hi = __lsx_vadd_h(dst_hi, const_32641);
+
+ dst_lo = __lsx_vmuh_hu(dst_lo, const_257);
+ dst_hi = __lsx_vmuh_hu(dst_hi, const_257);
+
+ dst128 = __lsx_vssrarni_bu_h(dst_hi, dst_lo, 0);
+ if (fill_alpha) {
+ dst128 = __lsx_vor_v(dst128, alpha_fill_mask);
+ }
+ __lsx_vst(dst128, dst, 0);
+
+ src += 16;
+ dst += 16;
+ }
+
+ for (; i < width; ++i) {
+ Uint32 src32 = *(Uint32 *)src;
+ Uint32 dst32 = *(Uint32 *)dst;
+ ALPHA_BLEND_SWIZZLE_8888(src32, dst32, srcfmt, dstfmt);
+ if (fill_alpha) {
+ dst32 |= dstAmask;
+ }
+ *(Uint32 *)dst = dst32;
+ src += 4;
+ dst += 4;
+ }
+
+ src += srcskip;
+ dst += dstskip;
+ }
+}
+
+#endif
+
// fast RGB888->(A)RGB888 blending with surface alpha=128 special case
static void BlitRGBtoRGBSurfaceAlpha128(SDL_BlitInfo *info)
{
@@ -1402,6 +1499,11 @@ SDL_BlitFunc SDL_CalculateBlitA(SDL_Surface *surface)
return Blit8888to8888PixelAlphaSwizzleSSE41;
}
#endif
+#ifdef SDL_LSX_INTRINSICS
+ if (SDL_HasLSX()) {
+ return Blit8888to8888PixelAlphaSwizzleLSX;
+ }
+#endif
#if defined(SDL_NEON_INTRINSICS) && (__ARM_ARCH >= 8)
// To prevent "unused function" compiler warnings/errors
(void)Blit8888to8888PixelAlpha;