From bccd1aa788de9709ae09aeae797e422ad405695a Mon Sep 17 00:00:00 2001
From: Luke Benstead <kazade@gmail.com>
Date: Mon, 20 Aug 2018 21:19:12 +0100
Subject: [PATCH] Start refactoring the submission process

---
 GL/clip.h    |  4 ++++
 GL/draw.c    | 36 +++++++++---------------------------
 GL/flush.c   |  7 ++++---
 GL/private.h | 10 ++++++++--
 4 files changed, 25 insertions(+), 32 deletions(-)

diff --git a/GL/clip.h b/GL/clip.h
index 58d379f..4035162 100644
--- a/GL/clip.h
+++ b/GL/clip.h
@@ -25,11 +25,15 @@ typedef enum {
 
 
 typedef struct {
+    /* Same 32 byte layout as pvr_vertex_t */
     uint32_t flags;
     float xyz[3];
     float uv[2];
     uint8_t bgra[4];
+    uint32_t oargb;
 
+    /* Important, we have 24 bytes here. That means when submitting to the SQs we need to
+     * increment the pointer by 6 */
     float nxyz[3]; /* Normal */
     float w;
     float st[2];
diff --git a/GL/draw.c b/GL/draw.c
index 25783ea..8b72b0b 100644
--- a/GL/draw.c
+++ b/GL/draw.c
@@ -741,22 +741,14 @@ static void divide(AlignedVector* vertices) {
     }
 }
 
-typedef struct {
-    unsigned int list_type;
-    pvr_poly_hdr_t hdr;
-} ListToHeader;
-
-
-#define MAX_LISTS 5
-
 static void push(const AlignedVector* vertices, PolyList* activePolyList, GLshort textureUnit) {
     /* Copy the vertices to the active poly list */
 
     // Make room for the element + the header
-    PVRCommand* dst = (PVRCommand*) aligned_vector_extend(&activePolyList->vector, vertices->size);
+    ClipVertex* dst = (ClipVertex*) aligned_vector_extend(&activePolyList->vector, vertices->size + 1);
 
     // Store a pointer to the header
-    pvr_poly_hdr_t* hdr = (pvr_poly_hdr_t*) dst;
+    PVRHeader* header = (PVRHeader*) dst;
 
     // Compile
     pvr_poly_cxt_t cxt = *getPVRContext();
@@ -764,32 +756,22 @@ static void push(const AlignedVector* vertices, PolyList* activePolyList, GLshor
 
     _glUpdatePVRTextureContext(&cxt, textureUnit);
 
-    pvr_poly_compile(hdr, &cxt);
+    pvr_poly_compile(&header->hdr, &cxt);
 
     // Point dest at the first new vertex to populate, if we're not sending a header
     // we won't increment and instead overwrite the header we just created with the
     // first vertex
     dst++;
 
-    // Add one more to the list
-    aligned_vector_extend(&activePolyList->vector, 1);
-
-    GLsizei i = vertices->size;
     ClipVertex* vin = aligned_vector_at(vertices, 0);
-    pvr_vertex_t* vout = (pvr_vertex_t*) dst;
+    ClipVertex* vout = dst;
 
+    GLuint i = vertices->size;
     while(i--) {
-        vout->flags = vin->flags;
-        vout->x = vin->xyz[0];
-        vout->y = vin->xyz[1];
-        vout->z = vin->xyz[2];
-        vout->u = vin->uv[0];
-        vout->v = vin->uv[1];
-        vout->argb = *((uint32_t*) vin->bgra);
-        vout->oargb = 0;
-
-        vin++;
-        vout++;
+        vin->oargb = 0;
+        *vout = *vin;
+        ++vout;
+        ++vin;
     }
 }
 
diff --git a/GL/flush.c b/GL/flush.c
index 116a317..f38f592 100644
--- a/GL/flush.c
+++ b/GL/flush.c
@@ -30,6 +30,7 @@ static void pvr_list_submit(void *src, int n) {
         d[7] = *(s++);
         __asm__("pref @%0" : : "r"(d));
         d += 8;
+        s += CLIP_VERTEX_INT_PADDING;
     }
 
     /* Wait for both store queues to complete */
@@ -81,9 +82,9 @@ void APIENTRY glKosInit() {
     PT_LIST.list_type = PVR_LIST_PT_POLY;
     TR_LIST.list_type = PVR_LIST_TR_POLY;
 
-    aligned_vector_init(&OP_LIST.vector, sizeof(PVRCommand));
-    aligned_vector_init(&PT_LIST.vector, sizeof(PVRCommand));
-    aligned_vector_init(&TR_LIST.vector, sizeof(PVRCommand));
+    aligned_vector_init(&OP_LIST.vector, sizeof(ClipVertex));
+    aligned_vector_init(&PT_LIST.vector, sizeof(ClipVertex));
+    aligned_vector_init(&TR_LIST.vector, sizeof(ClipVertex));
 }
 
 void APIENTRY glKosSwapBuffers() {
diff --git a/GL/private.h b/GL/private.h
index 87d7674..c50f767 100644
--- a/GL/private.h
+++ b/GL/private.h
@@ -11,9 +11,12 @@
 
 #define MAX_TEXTURE_SIZE 1024
 
+#define CLIP_VERTEX_INT_PADDING 6
+
 typedef struct {
-    unsigned int cmd[8];
-} PVRCommand;
+    pvr_poly_hdr_t hdr;
+    unsigned int padding[CLIP_VERTEX_INT_PADDING];
+} PVRHeader;
 
 typedef struct {
     unsigned int flags;      /* Constant PVR_CMD_USERCLIP */
@@ -22,6 +25,9 @@ typedef struct {
              sy,         /* Start y */
              ex,         /* End x */
              ey;         /* End y */
+
+    /* Padding to match clip vertex */
+    unsigned int padding[CLIP_VERTEX_INT_PADDING];
 } PVRTileClipCommand; /* Tile Clip command for the pvr */
 
 typedef struct {