Rework matrix-related code.

GLSL matrices are stored in column-major order while GL_ARB_vertex/fragment_program
use row-major.  So, need to use STATE_MATRIX_TRANSPOSE for built-in matrices.
Unfortunately, this means that the expression M * V isn't very efficient since we
need to extract the rows out of M.  And that's the typical expression for vertex
transformation:  gl_ModelViewProjectionMatrix * gl_Position.
Solve this inefficiency by looking for M*V expressions and replacing them
with V*Transpose(M).
Also, add support for GLSL 1.20's MatrixTranspose, Inverse and InverseTranspose
matrices.

1 files changed, 43 insertions, 0 deletions

diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c
index fc000dbf7be..1c037d4304b 100644
--- a/src/mesa/shader/slang/slang_codegen.c
+++ b/src/mesa/shader/slang/slang_codegen.c
@@ -2178,6 +2178,48 @@ _slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper)
 }
 
 
+/**
+ * Look for expressions such as:  gl_ModelviewMatrix * gl_Vertex
+ * and replace with this:  gl_Vertex * gl_ModelviewMatrixTranpose
+ * Since matrices are stored in column-major order, the second form of
+ * multiplication is much more efficient (just 4 dot products).
+ */
+static void
+_slang_check_matmul_optimization(slang_assemble_ctx *A, slang_operation *oper)
+{
+   static const struct {
+      const char *orig;
+      const char *tranpose;
+   } matrices[] = {
+      {"gl_ModelViewMatrix", "gl_ModelViewMatrixTranspose"},
+      {"gl_ProjectionMatrix", "gl_ProjectionMatrixTranspose"},
+      {"gl_ModelViewProjectionMatrix", "gl_ModelViewProjectionMatrixTranspose"},
+      {"gl_TextureMatrix", "gl_TextureMatrixTranspose"},
+      {"gl_NormalMatrix", "__NormalMatrixTranspose"},
+      { NULL, NULL }
+   };
+
+   assert(oper->type == SLANG_OPER_MULTIPLY);
+   if (oper->children[0].type == SLANG_OPER_IDENTIFIER) {
+      GLuint i;
+      for (i = 0; matrices[i].orig; i++) {
+         if (oper->children[0].a_id
+             == slang_atom_pool_atom(A->atoms, matrices[i].orig)) {
+            /*
+            _mesa_printf("Replace %s with %s\n",
+                         matrices[i].orig, matrices[i].tranpose);
+            */
+            assert(oper->children[0].type == SLANG_OPER_IDENTIFIER);
+            oper->children[0].a_id
+               = slang_atom_pool_atom(A->atoms, matrices[i].tranpose);
+            /* finally, swap the operands */
+            _slang_operation_swap(&oper->children[0], &oper->children[1]);
+            return;
+         }
+      }
+   }
+}
+
 
 /**
  * Generate IR tree for a slang_operation (AST node)
@@ -2309,6 +2351,7 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper)
       {
 	 slang_ir_node *n;
          assert(oper->num_children == 2);
+         _slang_check_matmul_optimization(A, oper);
          n = _slang_gen_function_call_name(A, "*", oper, NULL);
 	 return n;
       }