Optimise vertex transform in non-fast path to avoid storing xyz to memory and then loading it again

GL/attributes.c

@@ -37,67 +37,89 @@ GLuint* _glGetEnabledAttributes() {
 
 
 static void _readPosition3f3f(const GLubyte* __restrict__ in, GLubyte* __restrict__ out) {
-    vec3cpy(out, in);
+    const float* input = (const float*) in;
+    Vertex* it = (Vertex*) out;
+
+    float x = input[0];
+    float y = input[1];
+    float z = input[2];
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition3ub3f(const GLubyte* input, GLubyte* out) {
-    float* output = (float*) out;
+    Vertex* it = (Vertex*)out;
 
-    output[0] = input[0] * ONE_OVER_TWO_FIVE_FIVE;
+    float x = input[0] * ONE_OVER_TWO_FIVE_FIVE;
-    output[1] = input[1] * ONE_OVER_TWO_FIVE_FIVE;
+    float y = input[1] * ONE_OVER_TWO_FIVE_FIVE;
-    output[2] = input[2] * ONE_OVER_TWO_FIVE_FIVE;
+    float z = input[2] * ONE_OVER_TWO_FIVE_FIVE;
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition3us3f(const GLubyte* in, GLubyte* out) {
     const GLushort* input = (const GLushort*) in;
-    float* output = (float*) out;
+    Vertex* it = (Vertex*) out;
 
-    output[0] = input[0];
+    float x = input[0];
-    output[1] = input[1];
+    float y = input[1];
-    output[2] = input[2];
+    float z = input[2];
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition3ui3f(const GLubyte* in, GLubyte* out) {
     const GLuint* input = (const GLuint*) in;
-    float* output = (float*) out;
+    Vertex* it = (Vertex*) out;
 
-    output[0] = input[0];
+    float x = input[0];
-    output[1] = input[1];
+    float y = input[1];
-    output[2] = input[2];
+    float z = input[2];
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition2f3f(const GLubyte* in, GLubyte* out) {
     const float* input = (const float*) in;
-    float* output = (float*) out;
+    Vertex* it = (Vertex*) out;
 
-    vec2cpy(output, input);
+    float x = input[0];
-    output[2] = 0.0f;
+    float y = input[1];
+    float z = 0.0f;
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition2ub3f(const GLubyte* input, GLubyte* out) {
-    float* output = (float*) out;
+    Vertex* it = (Vertex*) out;
 
-    output[0] = input[0] * ONE_OVER_TWO_FIVE_FIVE;
+    float x = input[0] * ONE_OVER_TWO_FIVE_FIVE;
-    output[1] = input[1] * ONE_OVER_TWO_FIVE_FIVE;
+    float y = input[1] * ONE_OVER_TWO_FIVE_FIVE;
-    output[2] = 0.0f;
+    float z = 0.0f;
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition2us3f(const GLubyte* in, GLubyte* out) {
     const GLushort* input = (const GLushort*) in;
-    float* output = (float*) out;
+    Vertex* it = (Vertex*) out;
 
-    output[0] = input[0];
+    float x = input[0];
-    output[1] = input[1];
+    float y = input[1];
-    output[2] = 0.0f;
+    float z = 0.0f;
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static void _readPosition2ui3f(const GLubyte* in, GLubyte* out) {
     const GLuint* input = (const GLuint*) in;
-    float* output = (float*) out;
+    Vertex* it = (Vertex*)out;
 
-    output[0] = input[0];
+    float x = input[0];
-    output[1] = input[1];
+    float y = input[1];
-    output[2] = 0.0f;
+    float z = 0.0f;
+    float w = 1.0f;
+    TransformVertex(x, y, z, w, it->xyz, &it->w);
 }
 
 static ReadAttributeFunc calcReadPositionFunc() {

GL/draw.c

@@ -289,13 +289,10 @@ static void _readPositionData(const GLuint first, const GLuint count, Vertex* it
     const GLsizei vstride = ATTRIB_LIST.vertex.stride;
     const GLubyte* vptr = ((GLubyte*) ATTRIB_LIST.vertex.ptr + (first * vstride));
 
-    float pos[3], w = 1.0f;
-
     ITERATE(count) {
         PREFETCH(vptr + vstride);
-        func(vptr, (GLubyte*) pos);
+        func(vptr, (GLubyte*) it);
         it->flags = GPU_CMD_VERTEX;
-        TransformVertex(pos, &w, it->xyz, &it->w);
 
         vptr += vstride;
         ++it;
@@ -411,8 +408,7 @@ static void generateElements(
         st = (GLubyte*) ATTRIB_LIST.st.ptr + (idx * ststride);
         nxyz = (GLubyte*) ATTRIB_LIST.normal.ptr + (idx * nstride);
 
-        pos_func(xyz, (GLubyte*) pos);
+        pos_func(xyz, (GLubyte*) output);
-        TransformVertex((const float*) pos, &w, output->xyz, &output->w);
         uv_func(uv, (GLubyte*) output->uv);
         diffuse_func(bgra, output->bgra);
         st_func(st, (GLubyte*) ve->st);
@@ -460,8 +456,6 @@ static void generateElementsFastPath(
     VertexExtra* ve = aligned_vector_at(target->extras, 0);
     Vertex* it = start;
 
-    const float w = 1.0f;
-
     if(!pos) {
         return;
     }
@@ -472,7 +466,7 @@ static void generateElementsFastPath(
         it->flags = GPU_CMD_VERTEX;
 
         pos = (GLubyte*) ATTRIB_LIST.vertex.ptr + (idx * vstride);
-        TransformVertex((const float*) pos, &w, it->xyz, &it->w);
+        TransformVertex(((float*) pos)[0], ((float*) pos)[1], ((float*) pos)[2], 1.0f, it->xyz, &it->w);
 
         if(uv) {
             uv = (GLubyte*) ATTRIB_LIST.uv.ptr + (idx * uvstride);

GL/draw_fastpath.inc

@@ -5,7 +5,6 @@
 
 MAKE_FUNC(POLYMODE)
 {
-    static const float w = 1.0f;
     if(!(ATTRIB_LIST.enabled & VERTEX_ENABLED_FLAG)) {
         /* If we don't have vertices, do nothing */
         return;
@@ -75,7 +74,7 @@ MAKE_FUNC(POLYMODE)
         PREFETCH(ptr);
         for(int_fast32_t i = 0; i < loop; ++i, ++it) {
             PREFETCH(ptr + stride);
-            TransformVertex((const float*) ptr, &w, it->xyz, &it->w);
+            TransformVertex(((float*) ptr)[0], ((float*) ptr)[1], ((float*) ptr)[2], 1.0f, it->xyz, &it->w);
             PROCESS_VERTEX_FLAGS(it, min + i);
             ptr += stride;
         }

GL/platforms/sh4.h

@@ -106,15 +106,14 @@ inline void TransformVec4(float* x) {
 
 }
 
-GL_FORCE_INLINE void TransformVertex(const float* xyz, const float* w, float* oxyz, float* ow) {
+GL_FORCE_INLINE void TransformVertex(float x, float y, float z, float w, float* oxyz, float* ow) {
-    register float __x __asm__("fr12") = (xyz[0]);
+    register float __x __asm__("fr4") = x;
-    register float __y __asm__("fr13") = (xyz[1]);
+    register float __y __asm__("fr5") = y;
-    register float __z __asm__("fr14") = (xyz[2]);
+    register float __z __asm__("fr6") = z;
-    register float __w __asm__("fr15") = (*w);
+    register float __w __asm__("fr7") = w;
 
     __asm__ __volatile__(
-        "fldi1 fr15\n"
+        "ftrv   xmtrx,fv4\n"
-        "ftrv   xmtrx,fv12\n"
         : "=f" (__x), "=f" (__y), "=f" (__z), "=f" (__w)
         : "0" (__x), "1" (__y), "2" (__z), "3" (__w)
     );

GL/platforms/software.c

@@ -636,12 +636,12 @@ void TransformVertices(Vertex* vertices, const int count) {
     }
 }
 
-void TransformVertex(const float* xyz, const float* w, float* oxyz, float* ow) {
+void TransformVertex(float x, float y, float z, float w, float* oxyz, float* ow) {
     float ret[4];
-    ret[0] = xyz[0];
+    ret[0] = x;
-    ret[1] = xyz[1];
+    ret[1] = y;
-    ret[2] = xyz[2];
+    ret[2] = z;
-    ret[3] = *w;
+    ret[3] = w;
 
     TransformVec4(ret);
 

GL/platforms/software.h

@@ -53,7 +53,7 @@ static inline void TransformNormalNoMod(const float* xIn, float* xOut) {
 }
 
 void TransformVertices(Vertex* vertices, const int count);
-void TransformVertex(const float* xyz, const float* w, float* oxyz, float* ow);
+void TransformVertex(float x, float y, float z, float w, float* oxyz, float* ow);
 
 void InitGPU(_Bool autosort, _Bool fsaa);