Add BlitRGBtoRGBPixelAlphaLSX and SDL_FillRect4LSX opt

src/render/SDL_render.c

@@ -29,6 +29,12 @@
 #include "software/SDL_render_sw_c.h"
 #include "../video/SDL_pixels_c.h"
 
+#if defined(__loongarch__)
+#if SDL_VIDEO_OPENGL
+#include "SDL_opengl.h"
+#endif /* SDL_VIDEO_OPENGL */
+#endif
+
 #if defined(__ANDROID__)
 #include "../core/android/SDL_android.h"
 #endif
@@ -954,6 +960,34 @@ static void SDL_CalculateSimulatedVSyncInterval(SDL_Renderer *renderer, SDL_Wind
 }
 #endif /* !SDL_RENDER_DISABLED */
 
+#if defined(__loongarch__)
+static SDL_bool SDL_CheckIntegratedGraphics()
+{
+     SDL_Window *window = SDL_CreateWindow("OpenGL test", -32, -32, 32, 32,
+                                           SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
+     SDL_bool has_lg100_or_lg110 = SDL_FALSE;
+
+     if (window) {
+        SDL_GLContext context = SDL_GL_CreateContext(window);
+
+        if (context) {
+            const GLubyte *(APIENTRY *glGetStringFunc)(GLenum) =
+                SDL_GL_GetProcAddress("glGetString");
+            if (glGetStringFunc) {
+                const char *renderer = (const char*)glGetStringFunc(GL_RENDERER);
+                if (renderer && (SDL_strstr(renderer, "LG110") ||
+                    SDL_strstr(renderer, "LG100"))) {
+                    has_lg100_or_lg110 = SDL_TRUE;
+                }
+            }
+            SDL_GL_DeleteContext(context);
+        }
+        SDL_DestroyWindow(window);
+    }
+    return has_lg100_or_lg110;
+}
+#endif
+
 SDL_Renderer *SDL_CreateRenderer(SDL_Window *window, int index, Uint32 flags)
 {
 #ifndef SDL_RENDER_DISABLED
@@ -967,6 +1001,14 @@ SDL_Renderer *SDL_CreateRenderer(SDL_Window *window, int index, Uint32 flags)
     Android_ActivityMutex_Lock_Running();
 #endif
 
+/*CPU acceleration runs faster than integrated graphichs on Loongson at pressent.*/
+#if defined(__linux__) && defined(__loongarch__)
+    if (SDL_CheckIntegratedGraphics()) {
+        SDL_SetHint(SDL_HINT_FRAMEBUFFER_ACCELERATION, "0");
+        flags = SDL_RENDERER_SOFTWARE;
+    }
+#endif
+
     if (!window) {
         SDL_InvalidParamError("window");
         goto error;

src/video/SDL_blit_A.c

@@ -412,6 +412,66 @@ static void BlitRGBtoRGBPixelAlphaMMX(SDL_BlitInfo *info)
 
 #endif /* __MMX__ */
 
+#if defined(__loongarch_sx)
+
+static void
+BlitRGBtoRGBPixelAlphaLSX(SDL_BlitInfo * info)
+{
+    int width = info->dst_w;
+    int height = info->dst_h;
+    Uint32 *srcp = (Uint32 *) info->src;
+    int srcskip = info->src_skip >> 2;
+    Uint32 *dstp = (Uint32 *) info->dst;
+    int dstskip = info->dst_skip >> 2;
+    SDL_PixelFormat *sf = info->src_fmt;
+    Uint32 amask = sf->Amask;
+    Uint32 ashift = sf->Ashift;
+    Uint64 multmask, multmask2;
+
+    __m128i src1, src2, src3, dst1, alpha, alpha2;
+    multmask = 0x00FF;
+    multmask <<= (ashift * 2);
+    multmask2 = 0x00FF00FF00FF00FFULL;
+
+    while (height--) {
+        /* *INDENT-OFF* */
+        DUFFS_LOOP4({
+        Uint32 alpha1 = *srcp & amask;
+        if (alpha1 == 0) {
+            /* do nothing */
+        } else if (alpha1 == amask) {
+            *dstp = *srcp;
+        } else {
+            src1 = __lsx_vreplgr2vr_w(*srcp);
+            src1 = __lsx_vinsgr2vr_w(src1, *dstp, 1);
+            src2 = __lsx_vsllwil_hu_bu(src1, 0);
+
+            alpha = __lsx_vreplgr2vr_w(alpha1);
+            alpha = __lsx_vsrl_d(alpha, __lsx_vreplgr2vr_d(ashift));
+            alpha = __lsx_vilvl_h(alpha, alpha);
+            alpha2 = __lsx_vilvl_w(alpha, alpha);
+            alpha = __lsx_vor_v(alpha2, __lsx_vreplgr2vr_d(multmask));
+            alpha2 = __lsx_vxor_v(alpha2, __lsx_vreplgr2vr_d(multmask2));
+
+            src3 = __lsx_vilvl_d(alpha2, alpha);
+            src1 = __lsx_vmul_h(src2, src3);
+            src1 = __lsx_vsrli_h(src1, 8);
+            src2 = __lsx_vilvh_d(src1, src1);
+            src1 = __lsx_vadd_h(src1, src2);
+            dst1 = __lsx_vssrlni_bu_h(src1, src1, 0);
+            __lsx_vstelm_w(dst1, dstp, 0, 0);
+        }
+        ++srcp;
+        ++dstp;
+        }, width);
+        /* *INDENT-ON* */
+        srcp += srcskip;
+        dstp += dstskip;
+    }
+}
+
+#endif /* __loongarch_sx */
+
 #ifdef SDL_ARM_SIMD_BLITTERS
 void BlitARGBto565PixelAlphaARMSIMDAsm(int32_t w, int32_t h, uint16_t *dst, int32_t dst_stride, uint32_t *src, int32_t src_stride);
 
@@ -1456,7 +1516,7 @@ SDL_BlitFunc SDL_CalculateBlitA(SDL_Surface *surface)
 
         case 4:
             if (sf->Rmask == df->Rmask && sf->Gmask == df->Gmask && sf->Bmask == df->Bmask && sf->BytesPerPixel == 4) {
-#if defined(__MMX__) || defined(__3dNOW__)
+#if defined(__MMX__) || defined(__3dNOW__) || defined(__loongarch_sx)
                 if (sf->Rshift % 8 == 0 && sf->Gshift % 8 == 0 && sf->Bshift % 8 == 0 && sf->Ashift % 8 == 0 && sf->Aloss == 0) {
 #ifdef __3dNOW__
                     if (SDL_Has3DNow()) {
@@ -1467,9 +1527,14 @@ SDL_BlitFunc SDL_CalculateBlitA(SDL_Surface *surface)
                     if (SDL_HasMMX()) {
                         return BlitRGBtoRGBPixelAlphaMMX;
                     }
+#endif
+#ifdef __loongarch_sx
+                    if (SDL_HasLSX()) {
+                        return BlitRGBtoRGBPixelAlphaLSX;
+                    }
 #endif
                 }
-#endif /* __MMX__ || __3dNOW__ */
+#endif /* __MMX__ || __3dNOW__ || __loongarch_sx*/
                 if (sf->Amask == 0xff000000) {
 #ifdef SDL_ARM_NEON_BLITTERS
                     if (SDL_HasNEON()) {

src/video/SDL_fillrect.c

@@ -135,6 +135,61 @@ DEFINE_SSE_FILLRECT(4, Uint32)
 /* *INDENT-ON* */ /* clang-format on */
 #endif            /* __SSE__ */
 
+#if defined(__loongarch_sx)
+
+#define LSX_BEGIN __m128i c128 = __lsx_vreplgr2vr_w(color);
+
+#define LSX_WORK \
+    for (i = n / 64; i--;) { \
+        __lsx_vst(c128, p, 0); \
+        __lsx_vst(c128, p, 16); \
+        __lsx_vst(c128, p, 32); \
+        __lsx_vst(c128, p, 48); \
+        p += 64; \
+    }
+
+#define DEFINE_LSX_FILLRECT(bpp, type) \
+static void \
+SDL_FillRect##bpp##LSX(Uint8 *pixels, int pitch, Uint32 color, int w, int h) \
+{ \
+    int i, n; \
+    Uint8 *p = NULL; \
+ \
+    LSX_BEGIN; \
+ \
+    while (h--) { \
+        n = w * bpp; \
+        p = pixels; \
+ \
+        if (n > 63) { \
+            int adjust = 16 - ((uintptr_t)p & 15); \
+            if (adjust < 16) { \
+                n -= adjust; \
+                adjust /= bpp; \
+                while (adjust--) { \
+                    *((type *)p) = (type)color; \
+                    p += bpp; \
+                } \
+            } \
+            LSX_WORK; \
+        } \
+        if (n & 63) { \
+            int remainder = (n & 63); \
+            remainder /= bpp; \
+            while (remainder--) { \
+                *((type *)p) = (type)color; \
+                p += bpp; \
+            } \
+        } \
+        pixels += pitch; \
+    } \
+ \
+}
+
+DEFINE_LSX_FILLRECT(4, Uint32)
+
+#endif
+
 static void SDL_FillRect1(Uint8 *pixels, int pitch, Uint32 color, int w, int h)
 {
     int n;
@@ -423,6 +478,13 @@ int SDL_FillRects(SDL_Surface *dst, const SDL_Rect *rects, int count,
                 break;
             }
 #endif
+
+#ifdef __loongarch_sx
+                if (SDL_HasLSX()) {
+                    fill_function = SDL_FillRect4LSX;
+                    break;
+                }
+#endif
             fill_function = SDL_FillRect4;
             break;
         }