Dreamcast: Squeeze a bit more performance out of the system

third_party/gldc/src/sh4.c

@@ -1,7 +1,6 @@
 #include "platform.h"
 #include "sh4.h"
 
-
 #define CLIP_DEBUG 0
 
 #define PVR_VERTEX_BUF_SIZE 2560 * 256
@@ -118,13 +117,15 @@ static volatile uint32_t* PVR_LMMODE0 = (uint32_t*) 0xA05F6884;
 static volatile uint32_t* PVR_LMMODE1 = (uint32_t*) 0xA05F6888;
 static volatile uint32_t* QACR = (uint32_t*) 0xFF000038;
 
+#define V0_VIS (1 << 0)
+#define V1_VIS (1 << 1)
+#define V2_VIS (1 << 2)
+
 void SceneListSubmit(Vertex* v2, int n) {
     TRACE();
 
     /* You need at least a header, and 3 vertices to render anything */
-    if(n < 4) {
+    if(n < 4) return;
-        return;
-    }
 
     const float h = vid_mode->height;
 
@@ -146,7 +147,6 @@ void SceneListSubmit(Vertex* v2, int n) {
     fprintf(stderr, "----\n");
 #endif
     uint8_t visible_mask = 0;
-    uint8_t counter = 0;
 
     sq = SQ_BASE_ADDRESS;
 
@@ -154,20 +154,11 @@ void SceneListSubmit(Vertex* v2, int n) {
         PREFETCH(v2 + 1);
         switch(v2->flags) {
         case GPU_CMD_VERTEX_EOL:
-            if(counter < 2) {
-                continue;
-            }
-            counter = 0;
             break;
         case GPU_CMD_VERTEX:
-            ++counter;
-            if(counter < 3) {
             continue;
-            }
-            break;
         default:
             _glPushHeaderOrVertex(v2);
-            counter = 0;
             continue;
         };
 
@@ -177,12 +168,12 @@ void SceneListSubmit(Vertex* v2, int n) {
         visible_mask = (
             (v0->xyz[2] > -v0->w) << 0 |
             (v1->xyz[2] > -v1->w) << 1 |
-            (v2->xyz[2] > -v2->w) << 2 |
+            (v2->xyz[2] > -v2->w) << 2
-            (counter == 0) << 3
         );
+	Vertex __attribute__((aligned(32))) scratch[4];
 
         switch(visible_mask) {
-        case 15: /* All visible, but final vertex in strip */
+        case V0_VIS | V1_VIS | V2_VIS: /* All vertices visible */
         {
             _glPerspectiveDivideVertex(v0, h);
             _glPushHeaderOrVertex(v0);
@@ -194,15 +185,8 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(v2);
         }
         break;
-        case 7:
+        case V0_VIS: /* First vertex was visible */
-            /* All visible, push the first vertex and move on */
-            _glPerspectiveDivideVertex(v0, h);
-            _glPushHeaderOrVertex(v0);
-            break;
-        case 9:
-            /* First vertex was visible, last in strip */
         {
-                Vertex __attribute__((aligned(32))) scratch[2];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
 
@@ -222,36 +206,10 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(b);
         }
         break;
-        case 1:
+        case V1_VIS: /* Second vertex was visible */
-            /* First vertex was visible, but not last in strip */
         {
-                Vertex __attribute__((aligned(32))) scratch[2];
-                Vertex* a = &scratch[0];
-                Vertex* b = &scratch[1];
-
-                _glClipEdge(v0, v1, a);
-                a->flags = GPU_CMD_VERTEX;
-
-                _glClipEdge(v2, v0, b);
-                b->flags = GPU_CMD_VERTEX;
-
-                _glPerspectiveDivideVertex(v0, h);
-                _glPushHeaderOrVertex(v0);
-
-                _glPerspectiveDivideVertex(a, h);
-                _glPushHeaderOrVertex(a);
-
-                _glPerspectiveDivideVertex(b, h);
-                _glPushHeaderOrVertex(b);
-                _glPushHeaderOrVertex(b);
-            }
-            break;
-        case 10:
-        case 2:
             /* Second vertex was visible. In self case we need to create a triangle and produce
                 two new vertices: 1-2, and 2-3. */
-            {
-                Vertex __attribute__((aligned(32))) scratch[3];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
             Vertex* c = &scratch[2];
@@ -262,7 +220,7 @@ void SceneListSubmit(Vertex* v2, int n) {
             a->flags = GPU_CMD_VERTEX;
 
             _glClipEdge(v1, v2, b);
-                b->flags = v2->flags;
+            b->flags = GPU_CMD_VERTEX_EOL;
 
             _glPerspectiveDivideVertex(a, h);
             _glPushHeaderOrVertex(a);
@@ -274,10 +232,8 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(b);
         }
         break;
-        case 11:
+        case V0_VIS | V1_VIS: /* First and second vertex were visible */
-        case 3:  /* First and second vertex were visible */
         {
-            Vertex __attribute__((aligned(32))) scratch[3];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
             Vertex* c = &scratch[2];
@@ -291,7 +247,7 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(v0);
 
             _glClipEdge(v1, v2, a);
-            a->flags = v2->flags;
+            a->flags = GPU_CMD_VERTEX_EOL;
 
             _glPerspectiveDivideVertex(c, h);
             _glPushHeaderOrVertex(c);
@@ -304,11 +260,8 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(a);
         }
         break;
-        case 12:
+        case V2_VIS: /* Third vertex was visible */
-        case 4:
-            /* Third vertex was visible. */
         {
-                Vertex __attribute__((aligned(32))) scratch[3];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
             Vertex* c = &scratch[2];
@@ -324,10 +277,6 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPerspectiveDivideVertex(a, h);
             _glPushHeaderOrVertex(a);
 
-                if(counter % 2 == 1) {
-                    _glPushHeaderOrVertex(a);
-                }
-
             _glPerspectiveDivideVertex(b, h);
             _glPushHeaderOrVertex(b);
 
@@ -335,9 +284,8 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(c);
         }
         break;
-        case 13:
+        case V0_VIS | V2_VIS: /* First and third vertex were visible */
         {
-            Vertex __attribute__((aligned(32))) scratch[3];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
             Vertex* c = &scratch[2];
@@ -366,39 +314,8 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(c);
         }
         break;
-        case 5:  /* First and third vertex were visible */
+        case V1_VIS | V2_VIS: /* Second and third vertex were visible */
         {
-            Vertex __attribute__((aligned(32))) scratch[3];
-            Vertex* a = &scratch[0];
-            Vertex* b = &scratch[1];
-            Vertex* c = &scratch[2];
-
-            memcpy_vertex(c, v2);
-            c->flags = GPU_CMD_VERTEX;
-
-            _glClipEdge(v0, v1, a);
-            a->flags = GPU_CMD_VERTEX;
-
-            _glClipEdge(v1, v2, b);
-            b->flags = GPU_CMD_VERTEX;
-
-            _glPerspectiveDivideVertex(v0, h);
-            _glPushHeaderOrVertex(v0);
-
-            _glPerspectiveDivideVertex(a, h);
-            _glPushHeaderOrVertex(a);
-
-            _glPerspectiveDivideVertex(c, h);
-            _glPushHeaderOrVertex(c);
-            _glPerspectiveDivideVertex(b, h);
-            _glPushHeaderOrVertex(b);
-            _glPushHeaderOrVertex(c);
-        }
-        break;
-        case 14:
-        case 6:  /* Second and third vertex were visible */
-        {
-            Vertex __attribute__((aligned(32))) scratch[4];
             Vertex* a = &scratch[0];
             Vertex* b = &scratch[1];
             Vertex* c = &scratch[2];
@@ -427,9 +344,6 @@ void SceneListSubmit(Vertex* v2, int n) {
             _glPushHeaderOrVertex(d);
         }
         break;
-        case 8:
-        default:
-            break;
         }
     }
 

third_party/gldc/src/sh4.h

@@ -62,8 +62,6 @@ GL_FORCE_INLINE void* memcpy_fast(void *dest, const void *src, size_t len) {
 
 #define MEMCPY4(dst, src, bytes) memcpy_fast(dst, src, bytes)
 
-#define MEMSET4(dst, v, size) memset4((dst), (v), (size))
-
 GL_FORCE_INLINE void TransformVertex(const float* xyz, const float* w, float* oxyz, float* ow) {
     register float __x __asm__("fr12") = (xyz[0]);
     register float __y __asm__("fr13") = (xyz[1]);
@@ -85,10 +83,6 @@ GL_FORCE_INLINE void TransformVertex(const float* xyz, const float* w, float* ox
 
 void InitGPU(_Bool autosort, _Bool fsaa);
 
-static inline size_t GPUMemoryAvailable() {
-    return pvr_mem_available();
-}
-
 static inline void* GPUMemoryAlloc(size_t size) {
     return pvr_mem_malloc(size);
 }

third_party/gldc/src/texture.c

@@ -58,7 +58,7 @@ GLubyte _glInitTextures() {
     _glInitializeTextureObject(default_tex, 0);
     TEXTURE_ACTIVE = default_tex;
 
-    size_t vram_free = GPUMemoryAvailable();
+    size_t vram_free = pvr_mem_available();
     YALLOC_SIZE = vram_free - PVR_MEM_BUFFER_SIZE; /* Take all but 64kb VRAM */
     YALLOC_BASE = GPUMemoryAlloc(YALLOC_SIZE);