We are currently migrating Bugzilla to GitHub issues.
Any changes made to the bug tracker now will be lost, so please do not post new bugs or make changes to them.
When we're done, all bug URLs will redirect to their equivalent location on the new bug tracker.

[patch] Reduce drawcalls of SDL_RenderCopy and SDL_RenderCopyEx.

SDL_opengles_merge_render_copies.patch (text/plain), 11.51 KB, created by sowfelicity on 2021-01-02 12:25:02 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: sowfelicity

Created: 2021-01-02 12:25:02 UTC

Size: 11.51 KB

patch

obsolete

>diff -r 0c1b34c8cb23 src/render/opengles/SDL_glesfuncs.h
>--- a/src/render/opengles/SDL_glesfuncs.h	Sat Jan 02 09:28:16 2021 +0100
>+++ b/src/render/opengles/SDL_glesfuncs.h	Sat Jan 02 11:09:38 2021 +0000
>@@ -52,6 +52,7 @@
> SDL_PROC(void, glTexParameteriv, (GLenum, GLenum, const GLint *))
> SDL_PROC(void, glTexSubImage2D, (GLenum, GLint, GLint, GLint, GLsizei, GLsizei, GLenum, GLenum, const GLvoid *))
> SDL_PROC(void, glVertexPointer, (GLint, GLenum, GLsizei, const GLvoid *))
>+SDL_PROC(void, glColorPointer, (GLint, GLenum, GLsizei, const GLvoid *))
> SDL_PROC(void, glViewport, (GLint, GLint, GLsizei, GLsizei))
> SDL_PROC_OES(void, glBindFramebufferOES, (GLenum, GLuint))
> SDL_PROC_OES(void, glFramebufferTexture2DOES, (GLenum, GLenum, GLenum, GLuint, GLint))
>diff -r 0c1b34c8cb23 src/render/opengles/SDL_render_gles.c
>--- a/src/render/opengles/SDL_render_gles.c	Sat Jan 02 09:28:16 2021 +0100
>+++ b/src/render/opengles/SDL_render_gles.c	Sat Jan 02 11:09:38 2021 +0000
>@@ -861,6 +861,163 @@
>     }
> }
> 
>+static GLfloat *
>+MergeRenderCopies(SDL_Renderer *renderer, SDL_RenderCommand **pcmd, const void *vertices, size_t *pcount)
>+{
>+    if(pcmd && *pcmd){
>+        SDL_RenderCommand *cmd = (*pcmd)->next;
>+        GLfloat *verts = NULL, *pverts = NULL, *puvs = NULL, *pcolors = NULL;
>+        size_t count = 1, i = 0, j = 0;
>+        const SDL_Texture *texture = (*pcmd)->data.draw.texture;
>+        const SDL_BlendMode blend = (*pcmd)->data.draw.blend;
>+
>+        while(cmd){
>+            if(cmd->command == SDL_RENDERCMD_SETDRAWCOLOR){
>+                cmd = cmd->next;
>+                continue;
>+            }else if(cmd->command != SDL_RENDERCMD_COPY){
>+                break;
>+            }
>+            if(cmd->data.draw.texture != texture || cmd->data.draw.blend != blend){
>+                break;
>+            }
>+            count ++;
>+            cmd = cmd->next;
>+        }
>+
>+        verts = (GLfloat *) SDL_malloc(count * 6 * (2 + 2 + 4) *
>+                sizeof(GLfloat));
>+        if(!verts){
>+            SDL_OutOfMemory();
>+            return verts;
>+        }
>+        *pcount = count;
>+        pverts = verts;
>+        puvs = verts + count * 6 * 2;
>+        pcolors = puvs + count * 6 * 2;
>+        cmd = (*pcmd);
>+        for(i = 0; i < count; i++){
>+            const Uint8 r = cmd->data.draw.r;
>+            const Uint8 g = cmd->data.draw.g;
>+            const Uint8 b = cmd->data.draw.b;
>+            const Uint8 a = cmd->data.draw.a;
>+            const GLfloat color[4] = { ((GLfloat) r) * inv255f, ((GLfloat) g) * inv255f, ((GLfloat) b) * inv255f, ((GLfloat) a) * inv255f};
>+            const GLfloat *src =  (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>+
>+            SDL_memcpy(pverts, src, 2 * sizeof(GLfloat));
>+            SDL_memcpy(pverts + 2, src, 4 * 2 * sizeof(GLfloat));
>+            SDL_memcpy(pverts + 5 * 2, src + 3 * 2, 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs, src + 4 * 2, 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs + 2, src + 4 * 2, 4 * 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs + 5 * 2, src + 4 * 2 + 3 * 2, 2 * sizeof(GLfloat));
>+            for(j = 0; j < 6; j++){
>+                SDL_memcpy(pcolors + j * 4, color, sizeof(color));
>+            }
>+
>+            if(i == count - 1){
>+                *pcmd = cmd;
>+            }else{
>+                pverts += 6 * 2;
>+                puvs +=  6 * 2;
>+                pcolors += 6 * 4;
>+                cmd = cmd->next;
>+                while(cmd->command != SDL_RENDERCMD_COPY){
>+                    cmd = cmd->next;
>+                }
>+            }
>+        }
>+        return verts;
>+    }
>+    return NULL;
>+}
>+
>+static SDL_INLINE void
>+TranslateRotatef(const double cs, const double sn, const GLfloat translatex, const GLfloat translatey, const GLfloat x, const GLfloat y, GLfloat *newx, GLfloat *newy)
>+{
>+    // x no - translatex or y no - translatey.
>+    *newx = translatex + cs * x - sn * y;
>+    *newy = translatey + sn * x + cs * y;
>+}
>+
>+static GLfloat *
>+MergeRenderCopyExes(SDL_Renderer *renderer, SDL_RenderCommand **pcmd, const void *vertices, size_t *pcount)
>+{
>+    if(pcmd && *pcmd){
>+        SDL_RenderCommand *cmd = (*pcmd)->next;
>+        GLfloat *verts = NULL, *pverts = NULL, *puvs = NULL, *pcolors = NULL;
>+        size_t count = 1, i = 0, j = 0;
>+        const SDL_Texture *texture = (*pcmd)->data.draw.texture;
>+        const SDL_BlendMode blend = (*pcmd)->data.draw.blend;
>+        const double PI = acos(-1);
>+
>+        while(cmd){
>+            if(cmd->command == SDL_RENDERCMD_SETDRAWCOLOR){
>+                cmd = cmd->next;
>+                continue;
>+            }else if(cmd->command != SDL_RENDERCMD_COPY_EX){
>+                break;
>+            }
>+            if(cmd->data.draw.texture != texture || cmd->data.draw.blend != blend){
>+                break;
>+            }
>+            count ++;
>+            cmd = cmd->next;
>+        }
>+
>+        verts = (GLfloat *) SDL_malloc(count * 6 * (2 + 2 + 4) * sizeof (GLfloat));
>+        if(!verts){
>+            SDL_OutOfMemory();
>+            return verts;
>+        }
>+        *pcount = count;
>+        pverts = verts;
>+        puvs = verts + count * 6 * 2;
>+        pcolors = puvs + count * 6 * 2;
>+        cmd = (*pcmd);
>+        for(i = 0; i < count; i++){
>+            const Uint8 r = cmd->data.draw.r;
>+            const Uint8 g = cmd->data.draw.g;
>+            const Uint8 b = cmd->data.draw.b;
>+            const Uint8 a = cmd->data.draw.a;
>+            const GLfloat color[4] = { ((GLfloat) r) * inv255f, ((GLfloat) g) * inv255f, ((GLfloat) b) * inv255f, ((GLfloat) a) * inv255f};
>+            const GLfloat *src =  (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>+            const GLfloat translatex = src[16];
>+            const GLfloat translatey = src[17];
>+            const GLfloat angle = src[18];
>+            double cs = SDL_cos(angle * PI / 180);
>+            double sn = SDL_sin(angle * PI / 180);
>+            TranslateRotatef(cs, sn, translatex, translatey, src[0], src[1], pverts, pverts + 1);
>+            SDL_memcpy(pverts + 2, pverts, 2 * sizeof(GLfloat));
>+            for(j = 1; j < 4; j++){
>+                TranslateRotatef(cs, sn, translatex, translatey, src[j * 2],
>+                        src[j * 2 + 1], pverts + 2 + j * 2, pverts + 2 + j * 2 +
>+                        1);
>+            }
>+            SDL_memcpy(pverts + 5 * 2, pverts + 4 * 2, 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs, src + 4 * 2, 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs + 2, src + 4 * 2, 4 * 2 * sizeof(GLfloat));
>+            SDL_memcpy(puvs + 5 * 2, src + 4 * 2 + 3 * 2, 2 * sizeof(GLfloat));
>+            for(j = 0; j < 6; j++){
>+                SDL_memcpy(pcolors + j * 4, color, sizeof(color));
>+            }
>+
>+            if(i == count - 1){
>+                *pcmd = cmd;
>+            }else{
>+                pverts += 6 * 2;
>+                puvs +=  6 * 2;
>+                pcolors += 6 * 4;
>+                cmd = cmd->next;
>+                while(cmd->command != SDL_RENDERCMD_COPY_EX){
>+                    cmd = cmd->next;
>+                }
>+            }
>+        }
>+        return verts;
>+    }
>+    return NULL;
>+}
>+
> static int
> GLES_RunCommandQueue(SDL_Renderer * renderer, SDL_RenderCommand *cmd, void *vertices, size_t vertsize)
> {
>@@ -962,29 +1119,63 @@
>             }
> 
>             case SDL_RENDERCMD_COPY: {
>-                const GLfloat *verts = (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>-                SetCopyState(data, cmd);
>-                data->glVertexPointer(2, GL_FLOAT, 0, verts);
>-                data->glTexCoordPointer(2, GL_FLOAT, 0, verts + 8);
>-                data->glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
>+                size_t count = 0;
>+                GLfloat *verts = MergeRenderCopies(renderer, &cmd, vertices, &count);
>+                if(verts){
>+                    SetCopyState(data, cmd);
>+                    data->glEnableClientState(GL_VERTEX_ARRAY);
>+                    data->glVertexPointer(2, GL_FLOAT, 0, verts);
>+                    data->glEnableClientState(GL_TEXTURE_COORD_ARRAY);
>+                    data->glTexCoordPointer(2, GL_FLOAT, 0, verts + count * 6 * 2);
>+                    data->glEnableClientState(GL_COLOR_ARRAY);
>+                    data->glColorPointer(4, GL_FLOAT, 0, verts + count * 6 * (2 + 2));
>+                    data->glDrawArrays(GL_TRIANGLE_STRIP, 0, count * 6);
>+                    data->glDisableClientState(GL_COLOR_ARRAY);
>+                    data->glDisableClientState(GL_TEXTURE_COORD_ARRAY);
>+                    data->glDisableClientState(GL_VERTEX_ARRAY);
>+                    SDL_free(verts);
>+                }else{
>+                    const GLfloat *verts = (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>+                    SetCopyState(data, cmd);
>+                    data->glVertexPointer(2, GL_FLOAT, 0, verts);
>+                    data->glTexCoordPointer(2, GL_FLOAT, 0, verts + 8);
>+                    data->glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
>+                }
>                 break;
>             }
> 
>             case SDL_RENDERCMD_COPY_EX: {
>-                const GLfloat *verts = (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>-                const GLfloat translatex = verts[16];
>-                const GLfloat translatey = verts[17];
>-                const GLfloat angle = verts[18];
>-                SetCopyState(data, cmd);
>-                data->glVertexPointer(2, GL_FLOAT, 0, verts);
>-                data->glTexCoordPointer(2, GL_FLOAT, 0, verts + 8);
>+                size_t count = 0;
>+                GLfloat *verts = MergeRenderCopyExes(renderer, &cmd, vertices, &count);
>+                if(verts){
>+                    SetCopyState(data, cmd);
>+                    data->glEnableClientState(GL_VERTEX_ARRAY);
>+                    data->glVertexPointer(2, GL_FLOAT, 0, verts);
>+                    data->glEnableClientState(GL_TEXTURE_COORD_ARRAY);
>+                    data->glTexCoordPointer(2, GL_FLOAT, 0, verts + count * 6 * 2);
>+                    data->glEnableClientState(GL_COLOR_ARRAY);
>+                    data->glColorPointer(4, GL_FLOAT, 0, verts + count * 6 * (2 + 2));
>+                    data->glDrawArrays(GL_TRIANGLE_STRIP, 0, count * 6);
>+                    data->glDisableClientState(GL_COLOR_ARRAY);
>+                    data->glDisableClientState(GL_TEXTURE_COORD_ARRAY);
>+                    data->glDisableClientState(GL_VERTEX_ARRAY);
>+                    SDL_free(verts);
>+                }else{
>+                    const GLfloat *verts = (GLfloat *) (((Uint8 *) vertices) + cmd->data.draw.first);
>+                    const GLfloat translatex = verts[16];
>+                    const GLfloat translatey = verts[17];
>+                    const GLfloat angle = verts[18];
>+                    SetCopyState(data, cmd);
>+                    data->glVertexPointer(2, GL_FLOAT, 0, verts);
>+                    data->glTexCoordPointer(2, GL_FLOAT, 0, verts + 8);
> 
>-                /* Translate to flip, rotate, translate to position */
>-                data->glPushMatrix();
>-                data->glTranslatef(translatex, translatey, 0.0f);
>-                data->glRotatef(angle, 0.0, 0.0, 1.0);
>-                data->glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
>-                data->glPopMatrix();
>+                    /* Translate to flip, rotate, translate to position */
>+                    data->glPushMatrix();
>+                    data->glTranslatef(translatex, translatey, 0.0f);
>+                    data->glRotatef(angle, 0.0, 0.0, 1.0);
>+                    data->glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
>+                    data->glPopMatrix();
>+                }
>                 break;
>             }
>

Actions: View | Diff

Attachments on bug 5434: 4627

We are currently migrating Bugzilla to GitHub issues.Any changes made to the bug tracker now will be lost, so please do not post new bugs or make changes to them.When we're done, all bug URLs will redirect to their equivalent location on the new bug tracker.

We are currently migrating Bugzilla to GitHub issues.
Any changes made to the bug tracker now will be lost, so please do not post new bugs or make changes to them.
When we're done, all bug URLs will redirect to their equivalent location on the new bug tracker.