Redoing the way we handle vertex shaders for the draw module.

11 files changed, 151 insertions, 57 deletions

diff --git a/src/mesa/pipe/draw/draw_context.c b/src/mesa/pipe/draw/draw_context.c
index 66c66ff698..5efb173228 100644
--- a/src/mesa/pipe/draw/draw_context.c
+++ b/src/mesa/pipe/draw/draw_context.c
@@ -155,14 +155,6 @@ void draw_set_viewport_state( struct draw_context *draw,
 }
 
 
-void
-draw_set_vertex_shader(struct draw_context *draw,
-                       const struct pipe_shader_state *shader)
-{
-   draw_flush( draw );
-   draw->vertex_shader = *shader;
-}
-
 
 void
 draw_set_vertex_buffer(struct draw_context *draw,
diff --git a/src/mesa/pipe/draw/draw_context.h b/src/mesa/pipe/draw/draw_context.h
index 0ccf5f6046..398e96d94d 100644
--- a/src/mesa/pipe/draw/draw_context.h
+++ b/src/mesa/pipe/draw/draw_context.h
@@ -92,9 +92,12 @@ void draw_set_rasterize_stage( struct draw_context *draw,
                                struct draw_stage *stage );
 
 
-void
-draw_set_vertex_shader(struct draw_context *draw,
-                       const struct pipe_shader_state *shader);
+void * draw_create_vertex_shader(struct draw_context *draw,
+                                 const struct pipe_shader_state *shader);
+void draw_bind_vertex_shader(struct draw_context *draw,
+                             void *vcso);
+void draw_delete_vertex_shader(struct draw_context *draw,
+                               void *vcso);
 
 
 void
diff --git a/src/mesa/pipe/draw/draw_private.h b/src/mesa/pipe/draw/draw_private.h
index ebef5347ab..2f52299f74 100644
--- a/src/mesa/pipe/draw/draw_private.h
+++ b/src/mesa/pipe/draw/draw_private.h
@@ -46,6 +46,7 @@
 
 #include "draw_vertex.h"
 
+#include "x86/rtasm/x86sse.h"
 
 /**
  * Basic vertex info.
@@ -104,7 +105,7 @@ struct draw_stage
 
    void (*tri)( struct draw_stage *,
 		struct prim_header * );
-   
+
    void (*end)( struct draw_stage * );
 
    void (*reset_stipple_counter)( struct draw_stage * );
@@ -116,6 +117,15 @@ struct draw_stage
 #define VCACHE_OVERFLOW        4
 #define VS_QUEUE_LENGTH        (VCACHE_SIZE + VCACHE_OVERFLOW + 1)	/* can never fill up */
 
+/**
+ * Private version of the compiled vertex_shader
+ */
+struct draw_vertex_shader {
+   const struct pipe_shader_state *state;
+#if defined(__i386__) || defined(__386__)
+   struct x86_function  sse2_program;
+#endif
+};
 
 /**
  * Private context for the drawing module.
@@ -145,7 +155,7 @@ struct draw_context
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffer[PIPE_ATTRIB_MAX];
    struct pipe_vertex_element vertex_element[PIPE_ATTRIB_MAX];
-   struct pipe_shader_state vertex_shader;
+   const struct draw_vertex_shader *vertex_shader;
    struct pipe_vertex_buffer feedback_buffer[PIPE_ATTRIB_MAX];
    struct pipe_vertex_element feedback_element[PIPE_ATTRIB_MAX];
 
diff --git a/src/mesa/pipe/draw/draw_vertex_fetch.c b/src/mesa/pipe/draw/draw_vertex_fetch.c
index 2b839d641e..bc37e18c34 100644
--- a/src/mesa/pipe/draw/draw_vertex_fetch.c
+++ b/src/mesa/pipe/draw/draw_vertex_fetch.c
@@ -85,7 +85,7 @@ void draw_vertex_fetch( struct draw_context *draw,
       /*printf("fetch vertex %u: \n", j);*/
 
       /* loop over vertex attributes (vertex shader inputs) */
-      for (attr = 0; attr < draw->vertex_shader.num_inputs; attr++) {
+      for (attr = 0; attr < draw->vertex_shader->state->num_inputs; attr++) {
 
          unsigned buf = draw->vertex_element[attr].vertex_buffer_index;
          const void *src
diff --git a/src/mesa/pipe/draw/draw_vertex_shader.c b/src/mesa/pipe/draw/draw_vertex_shader.c
index a2e1cdc472..f7ddbd1909 100644
--- a/src/mesa/pipe/draw/draw_vertex_shader.c
+++ b/src/mesa/pipe/draw/draw_vertex_shader.c
@@ -36,6 +36,8 @@
 #include "draw_context.h"
 #include "draw_vertex.h"
 
+#include "x86/rtasm/x86sse.h"
+
 #include "pipe/tgsi/exec/tgsi_core.h"
 
 static INLINE unsigned
@@ -70,6 +72,7 @@ typedef void (XSTDCALL *codegen_function) (
    float (*constant)[4],
    struct tgsi_exec_vector *temporary );
 
+
 /**
  * Transform vertices with the current vertex program/shader
  * Up to four vertices can be shaded at a time.
@@ -92,7 +95,7 @@ run_vertex_program(struct draw_context *draw,
    const float *trans = draw->viewport.translate;
 
    assert(count <= 4);
-   assert(draw->vertex_shader.output_semantic_name[0]
+   assert(draw->vertex_shader->state->output_semantic_name[0]
           == TGSI_SEMANTIC_POSITION);
 
 #ifdef DEBUG
@@ -101,7 +104,7 @@ run_vertex_program(struct draw_context *draw,
 
    /* init machine state */
    tgsi_exec_machine_init(&machine,
-                          draw->vertex_shader.tokens,
+                          draw->vertex_shader->state->tokens,
                           PIPE_MAX_SAMPLERS,
                           NULL /*samplers*/ );
 
@@ -114,8 +117,8 @@ run_vertex_program(struct draw_context *draw,
    draw_vertex_fetch( draw, &machine, elts, count );
 
    /* run shader */
-   if( draw->vertex_shader.executable != NULL ) {
-      codegen_function func = (codegen_function) draw->vertex_shader.executable;
+   if( draw->vertex_shader->state->executable != NULL ) {
+      codegen_function func = (codegen_function) draw->vertex_shader->state->executable;
       func(
          machine.Inputs,
          machine.Outputs,
@@ -206,3 +209,42 @@ void draw_vertex_shader_queue_flush( struct draw_context *draw )
    draw->vs.queue_nr = 0;
 }
 
+
+void *
+draw_create_vertex_shader(struct draw_context *draw,
+                          const struct pipe_shader_state *shader)
+{
+   struct draw_vertex_shader *vs = calloc(1, sizeof(struct draw_vertex_shader));
+
+   vs->state = shader;
+#if defined(__i386__) || defined(__386__)
+   x86_init_func(&vs->sse2_program);
+
+   tgsi_emit_sse2(shader->tokens, &vs->sse2_program);
+
+   ((struct pipe_shader_state*)(vs->state))->executable =
+      x86_get_func(&vs->sse2_program);
+#endif
+
+   return vs;
+}
+
+void draw_bind_vertex_shader(struct draw_context *draw,
+                             void *vcso)
+{
+   draw_flush(draw);
+   draw->vertex_shader = (struct draw_vertex_shader*)(vcso);
+}
+
+void draw_delete_vertex_shader(struct draw_context *draw,
+                               void *vcso)
+{
+   struct draw_vertex_shader *vs = (struct draw_vertex_shader*)(vcso);
+#if defined(__i386__) || defined(__386__)
+   x86_release_func(&vs->sse2_program);
+#endif
+   free(vcso);
+}
+
+
+
diff --git a/src/mesa/pipe/i915simple/i915_state.c b/src/mesa/pipe/i915simple/i915_state.c
index 5c6d37a466..19ca5e575f 100644
--- a/src/mesa/pipe/i915simple/i915_state.c
+++ b/src/mesa/pipe/i915simple/i915_state.c
@@ -443,16 +443,13 @@ static void i915_set_polygon_stipple( struct pipe_context *pipe,
 {
 }
 
-
-static void *
-i915_create_shader_state(struct pipe_context *pipe,
-                         const struct pipe_shader_state *templ)
+static void * i915_create_fs_state(struct pipe_context *pipe,
+                                   const struct pipe_shader_state *templ)
 {
    return 0;
 }
 
-static void i915_bind_fs_state( struct pipe_context *pipe,
-                               void *fs )
+static void i915_bind_fs_state(struct pipe_context *pipe, void *fs)
 {
    struct i915_context *i915 = i915_context(pipe);
 
@@ -461,20 +458,35 @@ static void i915_bind_fs_state( struct pipe_context *pipe,
    i915->dirty |= I915_NEW_FS;
 }
 
+static void i915_delete_fs_state(struct pipe_context *pipe, void *shader)
+{
+   /*do nothing*/
+}
 
-static void i915_bind_vs_state(struct pipe_context *pipe,
-                               void *vs)
+static void *
+i915_create_vs_state(struct pipe_context *pipe,
+                     const struct pipe_shader_state *templ)
 {
    struct i915_context *i915 = i915_context(pipe);
 
    /* just pass-through to draw module */
-   draw_set_vertex_shader(i915->draw, (const struct pipe_shader_state *)vs);
+   return draw_create_vertex_shader(i915->draw, templ);
 }
 
-static void i915_delete_shader_state(struct pipe_context *pipe,
-                                     void *shader)
+static void i915_bind_vs_state(struct pipe_context *pipe, void *vs)
 {
-   /*do nothing*/
+   struct i915_context *i915 = i915_context(pipe);
+
+   /* just pass-through to draw module */
+   draw_bind_vertex_shader(i915->draw, vs);
+}
+
+static void i915_delete_vs_state(struct pipe_context *pipe, void *shader)
+{
+   struct i915_context *i915 = i915_context(pipe);
+
+   /* just pass-through to draw module */
+   draw_delete_vertex_shader(i915->draw, shader);
 }
 
 static void i915_set_constant_buffer(struct pipe_context *pipe,
@@ -707,12 +719,12 @@ i915_init_state_functions( struct i915_context *i915 )
    i915->pipe.create_rasterizer_state = i915_create_rasterizer_state;
    i915->pipe.bind_rasterizer_state = i915_bind_rasterizer_state;
    i915->pipe.delete_rasterizer_state = i915_delete_rasterizer_state;
-   i915->pipe.create_fs_state = i915_create_shader_state;
+   i915->pipe.create_fs_state = i915_create_fs_state;
    i915->pipe.bind_fs_state = i915_bind_fs_state;
-   i915->pipe.delete_fs_state = i915_delete_shader_state;
-   i915->pipe.create_vs_state = i915_create_shader_state;
+   i915->pipe.delete_fs_state = i915_delete_fs_state;
+   i915->pipe.create_vs_state = i915_create_vs_state;
    i915->pipe.bind_vs_state = i915_bind_vs_state;
-   i915->pipe.delete_vs_state = i915_delete_shader_state;
+   i915->pipe.delete_vs_state = i915_delete_vs_state;
 
    i915->pipe.set_blend_color = i915_set_blend_color;
    i915->pipe.set_clip_state = i915_set_clip_state;
diff --git a/src/mesa/pipe/softpipe/sp_context.c b/src/mesa/pipe/softpipe/sp_context.c
index e415966dae..695bf1a9e0 100644
--- a/src/mesa/pipe/softpipe/sp_context.c
+++ b/src/mesa/pipe/softpipe/sp_context.c
@@ -275,12 +275,12 @@ struct pipe_context *softpipe_create( struct pipe_winsys *pipe_winsys,
    softpipe->pipe.create_rasterizer_state = softpipe_create_rasterizer_state;
    softpipe->pipe.bind_rasterizer_state   = softpipe_bind_rasterizer_state;
    softpipe->pipe.delete_rasterizer_state = softpipe_delete_rasterizer_state;
-   softpipe->pipe.create_fs_state = softpipe_create_shader_state;
+   softpipe->pipe.create_fs_state = softpipe_create_fs_state;
    softpipe->pipe.bind_fs_state   = softpipe_bind_fs_state;
-   softpipe->pipe.delete_fs_state = softpipe_delete_shader_state;
-   softpipe->pipe.create_vs_state = softpipe_create_shader_state;
+   softpipe->pipe.delete_fs_state = softpipe_delete_fs_state;
+   softpipe->pipe.create_vs_state = softpipe_create_vs_state;
    softpipe->pipe.bind_vs_state   = softpipe_bind_vs_state;
-   softpipe->pipe.delete_vs_state = softpipe_delete_shader_state;
+   softpipe->pipe.delete_vs_state = softpipe_delete_vs_state;
 
    softpipe->pipe.set_blend_color = softpipe_set_blend_color;
    softpipe->pipe.set_clip_state = softpipe_set_clip_state;
diff --git a/src/mesa/pipe/softpipe/sp_context.h b/src/mesa/pipe/softpipe/sp_context.h
index c0a681f3d6..ccf29b5683 100644
--- a/src/mesa/pipe/softpipe/sp_context.h
+++ b/src/mesa/pipe/softpipe/sp_context.h
@@ -62,8 +62,12 @@ struct draw_stage;
 #define SP_NEW_VS            0x2000
 #define SP_NEW_CONSTANTS     0x4000
 
+struct sp_vertex_shader_state {
+   const struct pipe_shader_state *state;
+   void  *draw_data;
+};
 
-struct softpipe_context {     
+struct softpipe_context {
    struct pipe_context pipe;  /**< base class */
    struct softpipe_winsys *winsys;	/**< window system interface */
 
@@ -76,7 +80,7 @@ struct softpipe_context {
    const struct pipe_depth_stencil_state   *depth_stencil;
    const struct pipe_rasterizer_state *rasterizer;
    const struct pipe_shader_state *fs;
-   const struct pipe_shader_state *vs;
+   const struct sp_vertex_shader_state *vs;
 
    struct pipe_blend_color blend_color;
    struct pipe_clear_color_state clear_color;
diff --git a/src/mesa/pipe/softpipe/sp_state.h b/src/mesa/pipe/softpipe/sp_state.h
index f0e1461d25..f9061e86e5 100644
--- a/src/mesa/pipe/softpipe/sp_state.h
+++ b/src/mesa/pipe/softpipe/sp_state.h
@@ -87,12 +87,14 @@ void softpipe_set_constant_buffer(struct pipe_context *,
 void softpipe_set_feedback_state( struct pipe_context *,
                                   const struct pipe_feedback_state * );
 
-void *
-softpipe_create_shader_state( struct pipe_context *,
-                              const struct pipe_shader_state * );
-void softpipe_bind_fs_state( struct pipe_context *, void * );
-void softpipe_bind_vs_state( struct pipe_context *, void * );
-void softpipe_delete_shader_state( struct pipe_context *, void * );
+void *softpipe_create_fs_state(struct pipe_context *,
+                               const struct pipe_shader_state *);
+void softpipe_bind_fs_state(struct pipe_context *, void *);
+void softpipe_delete_fs_state(struct pipe_context *, void *);
+void *softpipe_create_vs_state(struct pipe_context *,
+                               const struct pipe_shader_state *);
+void softpipe_bind_vs_state(struct pipe_context *, void *);
+void softpipe_delete_vs_state(struct pipe_context *, void *);
 
 void softpipe_set_polygon_stipple( struct pipe_context *,
 				  const struct pipe_poly_stipple * );
diff --git a/src/mesa/pipe/softpipe/sp_state_derived.c b/src/mesa/pipe/softpipe/sp_state_derived.c
index 6c6e798069..b6145df8e2 100644
--- a/src/mesa/pipe/softpipe/sp_state_derived.c
+++ b/src/mesa/pipe/softpipe/sp_state_derived.c
@@ -43,7 +43,7 @@
  */
 static void calculate_vertex_layout( struct softpipe_context *softpipe )
 {
-   const struct pipe_shader_state *vs = softpipe->vs;
+   const struct pipe_shader_state *vs = softpipe->vs->state;
    const struct pipe_shader_state *fs = softpipe->fs;
    const interp_mode colorInterp
       = softpipe->rasterizer->flatshade ? INTERP_CONSTANT : INTERP_LINEAR;
diff --git a/src/mesa/pipe/softpipe/sp_state_fs.c b/src/mesa/pipe/softpipe/sp_state_fs.c
index 8306a95f44..f1bec2c73a 100644
--- a/src/mesa/pipe/softpipe/sp_state_fs.c
+++ b/src/mesa/pipe/softpipe/sp_state_fs.c
@@ -33,10 +33,13 @@
 #include "pipe/draw/draw_context.h"
 
 
-void * softpipe_create_shader_state(struct pipe_context *pipe,
-                                    const struct pipe_shader_state *templ)
+void * softpipe_create_fs_state(struct pipe_context *pipe,
+                                const struct pipe_shader_state *templ)
 {
-   /* we just want the pipe_shader_state template in the bind calls */
+   /* Decide whether we'll be codegenerating this shader and if so do
+    * that now.
+    */
+
    return 0;
 }
 
@@ -49,25 +52,51 @@ void softpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
    softpipe->dirty |= SP_NEW_FS;
 }
 
+void softpipe_delete_fs_state(struct pipe_context *pipe,
+                              void *shader)
+{
+}
+
+
+void * softpipe_create_vs_state(struct pipe_context *pipe,
+                                const struct pipe_shader_state *templ)
+{
+   struct softpipe_context *softpipe = softpipe_context(pipe);
+   struct sp_vertex_shader_state *state =
+      malloc(sizeof(struct sp_vertex_shader_state));
+
+   state->state = templ;
+   state->draw_data = draw_create_vertex_shader(softpipe->draw,
+                                                state->state);
+
+   return state;
+}
 
 void softpipe_bind_vs_state(struct pipe_context *pipe, void *vs)
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
 
-   softpipe->vs = (struct pipe_shader_state *)vs;
+   softpipe->vs = (const struct sp_vertex_shader_state *)vs;
 
-   softpipe->dirty |= SP_NEW_VS;
+   draw_bind_vertex_shader(softpipe->draw, softpipe->vs->draw_data);
 
-   draw_set_vertex_shader(softpipe->draw, (struct pipe_shader_state *)vs);
+   softpipe->dirty |= SP_NEW_VS;
 }
 
-
-void softpipe_delete_shader_state( struct pipe_context *pipe,
-                                   void *shader )
+void softpipe_delete_vs_state(struct pipe_context *pipe,
+                              void *vs)
 {
-   /* do nothing */
+   struct softpipe_context *softpipe = softpipe_context(pipe);
+
+   struct sp_vertex_shader_state *state =
+      (struct sp_vertex_shader_state *)vs;
+
+   draw_delete_vertex_shader(softpipe->draw, state->draw_data);
+   free(state);
 }
 
+
+
 void softpipe_set_constant_buffer(struct pipe_context *pipe,
                                   uint shader, uint index,
                                   const struct pipe_constant_buffer *buf)