#368 asm optimizations

XProger · Jul 2, 2022 · ea021e0 · ea021e0
1 parent daa0e35
commit ea021e0
Show file tree

Hide file tree

Showing 11 changed files with 213 additions and 204 deletions.
diff --git a/src/fixed/common.h b/src/fixed/common.h
@@ -1,7 +1,7 @@
 #ifndef H_COMMON
 #define H_COMMON
 //#define STATIC_ITEMS
-//#define PROFILING
+#define PROFILING
 #ifdef PROFILING
     #define STATIC_ITEMS
     #define PROFILE_FRAMETIME
@@ -29,6 +29,8 @@
 #define SND_FMT_OGG     (1 << 14)
 #define SND_FMT_MP3     (1 << 15)
 
+#define FIXED_SHIFT     14
+
 #if defined(__WIN32__)
     #define USE_DIV_TABLE
     #define MODEHW
@@ -348,133 +350,22 @@ X_INLINE int32 abs(int32 x) {
     extern uint16 fb[VRAM_WIDTH * FRAME_HEIGHT];
 #endif
 
-#ifdef PROFILING
-    #define PROFILE_FRAME\
-        CNT_UPDATE,\
-        CNT_RENDER
-
-    #define PROFILE_STAGES\
-        CNT_TRANSFORM,\
-        CNT_ADD,\
-        CNT_FLUSH,\
-        CNT_VERT,\
-        CNT_POLY
-
-    #define PROFILE_SOUND\
-        CNT_SOUND
-
-    #if defined(PROFILE_FRAMETIME)
-        enum ProfileCounterId {
-            PROFILE_FRAME,
-            CNT_MAX,
-            PROFILE_STAGES,
-            PROFILE_SOUND
-        };
-    #elif defined(PROFILE_SOUNDTIME)
-        enum ProfileCounterId {
-            PROFILE_SOUND,
-            CNT_MAX,
-            PROFILE_FRAME,
-            PROFILE_STAGES
-        };
-    #else
-        enum ProfileCounterId {
-            PROFILE_STAGES,
-            CNT_MAX,
-            PROFILE_FRAME,
-            PROFILE_SOUND
-        };
-    #endif
-
-    extern uint32 gCounters[CNT_MAX];
-
-    #if defined(__3DO__) || defined(__32X__) // should be first, armcpp bug (#elif)
-        extern int32 g_timer;
-
-        #define PROFILE_START() {\
-            g_timer = osGetSystemTimeMS();\
-        }
-
-        #define PROFILE_STOP(value) {\
-            value += (osGetSystemTimeMS() - g_timer);\
-        }
-    #elif defined(__WIN32__) || defined(__GBA_WIN__)
-        extern LARGE_INTEGER g_timer;
-        extern LARGE_INTEGER g_current;
-
-        #define PROFILE_START() {\
-            QueryPerformanceCounter(&g_timer);\
-        }
-
-        #define PROFILE_STOP(value) {\
-            QueryPerformanceCounter(&g_current);\
-            value += uint32(g_current.QuadPart - g_timer.QuadPart);\
-        }
-    #elif defined(__GBA__)
-        #ifdef PROFILE_SOUNDTIME
-            #define TIMER_FREQ_DIV 1
-        #else
-            #define TIMER_FREQ_DIV 3
-        #endif
-
-        #define PROFILE_START() {\
-            REG_TM2CNT_L = 0;\
-            REG_TM2CNT_H = (1 << 7) | TIMER_FREQ_DIV;\
-        }
-
-        #define PROFILE_STOP(value) {\
-            value += REG_TM2CNT_L;\
-            REG_TM2CNT_H = 0;\
-        }
-    #else
-        #define PROFILE_START() aaa
-        #define PROFILE_STOP(value) bbb
-    #endif
-
-    struct ProfileCounter
-    {
-        ProfileCounterId cnt;
-
-        ProfileCounter(ProfileCounterId cnt) : cnt(cnt) {
-            if (cnt < CNT_MAX) {
-                PROFILE_START()
-            }
-        }
-
-        ~ProfileCounter() {
-            if (cnt < CNT_MAX) {
-                PROFILE_STOP(gCounters[cnt]);
-            }
-        }
-    };
-
-    #define PROFILE(cnt) ProfileCounter profileCounter(cnt)
-    #define PROFILE_CLEAR() memset(gCounters, 0, sizeof(gCounters));
-#else
-    #define PROFILE(cnt)
-    #define PROFILE_CLEAR()
-#endif
-
-#ifdef __TNS__
-    void osSetPalette(uint16* palette);
-#endif
-
 #define STATIC_MESH_FLAG_NO_COLLISION   1
 #define STATIC_MESH_FLAG_VISIBLE        2
 #define MAX_STATIC_MESH_RADIUS          (5 * 1024)
 
 extern int32 fps;
 
-#define FIXED_SHIFT     14
-
 #ifndef F16_SHIFT
     #define F16_SHIFT 0
 #endif
 
 #ifdef USE_MATRIX_INT16
-    #define MATRIX_FIXED_SHIFT  FIXED_SHIFT
-#else
-    #define MATRIX_FIXED_SHIFT  0
+    #define MATRIX_FIXED_SHIFT FIXED_SHIFT
+#endif
+
+#ifndef MATRIX_FIXED_SHIFT
+   #define MATRIX_FIXED_SHIFT 0
 #endif
 
 #define SND_MAX_DIST    (8 * 1024)
@@ -2966,4 +2857,112 @@ void osSetPalette(const uint16* palette);
 const void* osLoadScreen(LevelID id);
 const void* osLoadLevel(LevelID id);
 
+#ifdef PROFILING
+    #define PROFILE_FRAME\
+        CNT_UPDATE,\
+        CNT_RENDER
+
+    #define PROFILE_STAGES\
+        CNT_TRANSFORM,\
+        CNT_ADD,\
+        CNT_FLUSH,\
+        CNT_VERT,\
+        CNT_POLY
+
+    #define PROFILE_SOUND\
+        CNT_SOUND
+
+    #if defined(PROFILE_FRAMETIME)
+        enum ProfileCounterId {
+            PROFILE_FRAME,
+            CNT_MAX,
+            PROFILE_STAGES,
+            PROFILE_SOUND
+        };
+    #elif defined(PROFILE_SOUNDTIME)
+        enum ProfileCounterId {
+            PROFILE_SOUND,
+            CNT_MAX,
+            PROFILE_FRAME,
+            PROFILE_STAGES
+        };
+    #else
+        enum ProfileCounterId {
+            PROFILE_STAGES,
+            CNT_MAX,
+            PROFILE_FRAME,
+            PROFILE_SOUND
+        };
+    #endif
+
+    extern uint32 gCounters[CNT_MAX];
+
+    #if defined(__3DO__) || defined(__32X__) // should be first, armcpp bug (#elif)
+        extern int32 g_timer;
+
+        #define PROFILE_START() {\
+            g_timer = osGetSystemTimeMS();\
+        }
+
+        #define PROFILE_STOP(value) {\
+            value += (osGetSystemTimeMS() - g_timer);\
+        }
+    #elif defined(__WIN32__) || defined(__GBA_WIN__)
+        extern LARGE_INTEGER g_timer;
+        extern LARGE_INTEGER g_current;
+
+        #define PROFILE_START() {\
+            QueryPerformanceCounter(&g_timer);\
+        }
+
+        #define PROFILE_STOP(value) {\
+            QueryPerformanceCounter(&g_current);\
+            value += uint32(g_current.QuadPart - g_timer.QuadPart);\
+        }
+    #elif defined(__GBA__)
+        #ifdef PROFILE_SOUNDTIME
+            #define TIMER_FREQ_DIV 1
+        #else
+            #define TIMER_FREQ_DIV 3
+        #endif
+
+        #define PROFILE_START() {\
+            REG_TM2CNT_L = 0;\
+            REG_TM2CNT_H = (1 << 7) | TIMER_FREQ_DIV;\
+        }
+
+        #define PROFILE_STOP(value) {\
+            value += REG_TM2CNT_L;\
+            REG_TM2CNT_H = 0;\
+        }
+    #else
+        #define PROFILE_START() aaa
+        #define PROFILE_STOP(value) bbb
+    #endif
+
+    struct ProfileCounter
+    {
+        ProfileCounterId cnt;
+
+        ProfileCounter(ProfileCounterId cnt) : cnt(cnt) {
+            if (cnt < CNT_MAX) {
+                PROFILE_START()
+            }
+        }
+
+        ~ProfileCounter() {
+            if (cnt < CNT_MAX) {
+                PROFILE_STOP(gCounters[cnt]);
+            }
+        }
+    };
+
+    #define PROFILE(cnt) ProfileCounter profileCounter(cnt)
+    #define PROFILE_CLEAR() memset(gCounters, 0, sizeof(gCounters));
+#else
+    #define PROFILE(cnt)
+    #define PROFILE_CLEAR()
+#endif
+
+
 #endif
diff --git a/src/platform/gba/asm/common_asm.inc b/src/platform/gba/asm/common_asm.inc
@@ -94,6 +94,7 @@
 .equ MIN_INT32,     0x80000000
 .equ MAX_INT32,     0x7FFFFFFF
 
+// res = divTable[x] (uint16)
 .macro divLUT res, x
     add \res, \x, #DIVLUT_ADDR
     ldrh \res, [\res, \x]

diff --git a/src/platform/gba/asm/faceAddMeshQuads.s b/src/platform/gba/asm/faceAddMeshQuads.s
@@ -24,10 +24,6 @@ vy1         .req vg3
 vx2         .req vg2
 vy2         .req vg2
 
-vz0         .req vg0
-vz1         .req vg1
-vz2         .req vg2
-vz3         .req vg3
 depth       .req vg0
 
 tmp         .req flags
@@ -52,10 +48,12 @@ faceAddMeshQuads_asm:
     add polys, #2   // skip flags
 
 .loop:
-    ldrb vp0, [polys], #1
-    ldrb vp1, [polys], #1
-    ldrb vp2, [polys], #1
-    ldrb vp3, [polys], #3   // + flags
+    ldrh vp0, [polys], #2
+    lsr vp1, vp0, #8
+    and vp0, #0xFF
+    ldrh vp2, [polys], #4   // + flags
+    lsr vp3, vp2, #8
+    and vp2, #0xFF
 
     add vp0, vp, vp0, lsl #3
     add vp1, vp, vp1, lsl #3
@@ -64,36 +62,33 @@ faceAddMeshQuads_asm:
 
     CCW .skip
 
-    // fetch clip flags
-    ldrb vg0, [vp0, #VERTEX_CLIP]
-    ldrb vg1, [vp1, #VERTEX_CLIP]
-    ldrb vg2, [vp2, #VERTEX_CLIP]
-    ldrb vg3, [vp3, #VERTEX_CLIP]
+    // fetch [c, g, zz]
+    ldr vg0, [vp0, #VERTEX_Z]
+    ldr vg1, [vp1, #VERTEX_Z]
+    ldr vg2, [vp2, #VERTEX_Z]
+    ldr vg3, [vp3, #VERTEX_Z]
 
     // check clipping
     and tmp, vg0, vg1
     and tmp, vg2
     and tmp, vg3
-    tst tmp, #(CLIP_DISCARD >> 8)
+    tst tmp, #(CLIP_DISCARD << 16)
     bne .skip
 
     // mark if should be clipped by viewport
     orr tmp, vg0, vg1
     orr tmp, vg2
     orr tmp, vg3
-    tst tmp, #(CLIP_FRAME >> 8)
+    tst tmp, #(CLIP_FRAME << 16)
     ldrh flags, [polys, #-8]
     orrne flags, #FACE_CLIPPED
 
-    // vz0 = AVG_Z4 (depth)
-    ldrh vz0, [vp0, #VERTEX_Z]
-    ldrh vz1, [vp1, #VERTEX_Z]
-    ldrh vz2, [vp2, #VERTEX_Z]
-    ldrh vz3, [vp3, #VERTEX_Z]
-    add depth, vz0, vz1
-    add depth, vz2
-    add depth, vz3
-    lsr depth, #(2 + OT_SHIFT)
+    // depth = AVG_Z4
+    lsl vg0, #16
+    add depth, vg0, vg1, lsl #16
+    add depth, vg2, lsl #16
+    add depth, vg3, lsl #16
+    lsr depth, #(16 + 2 + OT_SHIFT)
 
     // faceAdd
     rsb vp0, vertices, vp0, lsr #3