From e1e3eaf51b36bbd8ecaeef4beff1d55613a64a4f Mon Sep 17 00:00:00 2001
From: Luke Benstead <kazade@gmail.com>
Date: Tue, 20 Apr 2021 19:27:16 +0100
Subject: [PATCH] Testing with batched arrays

---
 GL/draw.c | 50 +++++++++++++++++++++++++++++++++-----------------
 1 file changed, 33 insertions(+), 17 deletions(-)

diff --git a/GL/draw.c b/GL/draw.c
index 1f0914e..873c7f3 100644
--- a/GL/draw.c
+++ b/GL/draw.c
@@ -616,11 +616,10 @@ ReadNormalFunc calcReadNormalFunc() {
     }
 }
 
-void _readPositionData(const GLuint first, const GLuint count, const Vertex* output) {
+static void _readPositionData(ReadDiffuseFunc func, const GLuint first, const GLuint count, const Vertex* output) {
     const GLsizei vstride = (VERTEX_POINTER.stride) ? VERTEX_POINTER.stride : VERTEX_POINTER.size * byte_size(VERTEX_POINTER.type);
     const void* vptr = ((GLubyte*) VERTEX_POINTER.ptr + (first * vstride));
 
-    ReadDiffuseFunc func = calcReadPositionFunc();
     GLubyte* out = (GLubyte*) output[0].xyz;
     uint32_t* flags;
 
@@ -637,11 +636,10 @@ void _readPositionData(const GLuint first, const GLuint count, const Vertex* out
     }
 }
 
-void _readUVData(const GLuint first, const GLuint count, const Vertex* output) {
+static void _readUVData(ReadUVFunc func, const GLuint first, const GLuint count, const Vertex* output) {
     const GLsizei uvstride = (UV_POINTER.stride) ? UV_POINTER.stride : UV_POINTER.size * byte_size(UV_POINTER.type);
     const void* uvptr = ((GLubyte*) UV_POINTER.ptr + (first * uvstride));
 
-    ReadUVFunc func = calcReadUVFunc();
     GLubyte* out = (GLubyte*) output[0].uv;
 
     ITERATE(count) {
@@ -651,11 +649,10 @@ void _readUVData(const GLuint first, const GLuint count, const Vertex* output) {
     }
 }
 
-void _readSTData(const GLuint first, const GLuint count, const VertexExtra* extra) {
+static void _readSTData(ReadUVFunc func, const GLuint first, const GLuint count, const VertexExtra* extra) {
     const GLsizei ststride = (ST_POINTER.stride) ? ST_POINTER.stride : ST_POINTER.size * byte_size(ST_POINTER.type);
     const void* stptr = ((GLubyte*) ST_POINTER.ptr + (first * ststride));
 
-    ReadUVFunc func = calcReadSTFunc();
     GLubyte* out = (GLubyte*) extra[0].st;
 
     ITERATE(count) {
@@ -665,11 +662,10 @@ void _readSTData(const GLuint first, const GLuint count, const VertexExtra* extr
     }
 }
 
-void _readNormalData(const GLuint first, const GLuint count, const VertexExtra* extra) {
+static void _readNormalData(ReadNormalFunc func, const GLuint first, const GLuint count, const VertexExtra* extra) {
     const GLsizei nstride = (NORMAL_POINTER.stride) ? NORMAL_POINTER.stride : NORMAL_POINTER.size * byte_size(NORMAL_POINTER.type);
     const void* nptr = ((GLubyte*) NORMAL_POINTER.ptr + (first * nstride));
 
-    ReadNormalFunc func = calcReadNormalFunc();
     GLubyte* out = (GLubyte*) extra[0].nxyz;
 
     ITERATE(count) {
@@ -696,12 +692,11 @@ void _readNormalData(const GLuint first, const GLuint count, const VertexExtra*
     }
 }
 
-void _readDiffuseData(const GLuint first, const GLuint count, const Vertex* output) {
+static void _readDiffuseData(ReadDiffuseFunc func, const GLuint first, const GLuint count, const Vertex* output) {
     const GLuint size = (DIFFUSE_POINTER.size == GL_BGRA) ? 4 : DIFFUSE_POINTER.size;
     const GLuint cstride = (DIFFUSE_POINTER.stride) ? DIFFUSE_POINTER.stride : size * byte_size(DIFFUSE_POINTER.type);
     const GLubyte* cptr = ((GLubyte*) DIFFUSE_POINTER.ptr) + (first * cstride);
 
-    ReadDiffuseFunc func = calcReadDiffuseFunc();
     GLubyte* out = (GLubyte*) output[0].bgra;
 
     ITERATE(count) {
@@ -788,19 +783,40 @@ static void generateArraysFastPath(SubmissionTarget* target, const GLsizei first
 
     VertexExtra* ve = aligned_vector_at(target->extras, 0);
 
-    _readNormalData(first, count, ve);
-    _readSTData(first, count, ve);
+    ReadNormalFunc nfunc = calcReadNormalFunc();
+    ReadUVFunc stfunc = calcReadSTFunc();
+
+    _readNormalData(nfunc, first, count, ve);
+    _readSTData(stfunc, first, count, ve);
 }
 
+#define BATCH_SIZE 32
+
 static void generateArrays(SubmissionTarget* target, const GLsizei first, const GLuint count) {
     Vertex* start = _glSubmissionTargetStart(target);
     VertexExtra* ve = aligned_vector_at(target->extras, 0);
 
-    _readPositionData(first, count, start);
-    _readDiffuseData(first, count, start);
-    _readUVData(first, count, start);
-    _readNormalData(first, count, ve);
-    _readSTData(first, count, ve);
+    GLsizei s = first;
+    GLuint e = s + BATCH_SIZE;
+
+    ReadPositionFunc pfunc = calcReadPositionFunc();
+    ReadDiffuseFunc dfunc = calcReadDiffuseFunc();
+    ReadUVFunc uvfunc = calcReadUVFunc();
+    ReadNormalFunc nfunc = calcReadNormalFunc();
+    ReadUVFunc stfunc = calcReadSTFunc();
+
+    do {
+        _readPositionData(pfunc, s, BATCH_SIZE, start);
+        _readDiffuseData(dfunc, s, BATCH_SIZE, start);
+        _readUVData(uvfunc, s, BATCH_SIZE, start);
+        _readNormalData(nfunc, s, BATCH_SIZE, ve);
+        _readSTData(stfunc, s, BATCH_SIZE, ve);
+
+        s = e;
+        e += BATCH_SIZE;
+        start += BATCH_SIZE;
+        ve += BATCH_SIZE;
+    } while (s < count);
 }
 
 static void generate(SubmissionTarget* target, const GLenum mode, const GLsizei first, const GLuint count,