Enable SSE2 for fragment shaders.

4 files changed, 29 insertions, 17 deletions

diff --git a/src/mesa/pipe/softpipe/sp_context.h b/src/mesa/pipe/softpipe/sp_context.h
index e51e81c7b83..548151b378e 100644
--- a/src/mesa/pipe/softpipe/sp_context.h
+++ b/src/mesa/pipe/softpipe/sp_context.h
@@ -79,7 +79,7 @@ struct softpipe_context {
    const struct pipe_sampler_state *sampler[PIPE_MAX_SAMPLERS];
    const struct pipe_depth_stencil_state   *depth_stencil;
    const struct pipe_rasterizer_state *rasterizer;
-   const struct pipe_shader_state *fs;
+   const struct sp_fragment_shader_state *fs;
    const struct sp_vertex_shader_state *vs;
 
    struct pipe_blend_color blend_color;
diff --git a/src/mesa/pipe/softpipe/sp_quad_fs.c b/src/mesa/pipe/softpipe/sp_quad_fs.c
index dc8d6e0e230..f079ab9583b 100644
--- a/src/mesa/pipe/softpipe/sp_quad_fs.c
+++ b/src/mesa/pipe/softpipe/sp_quad_fs.c
@@ -37,7 +37,10 @@
 
 #include "pipe/llvm/llvmtgsi.h"
 
+#include "x86/rtasm/x86sse.h"
+
 #include "sp_context.h"
+#include "sp_state.h"
 #include "sp_headers.h"
 #include "sp_quad.h"
 #include "sp_tex_sample.h"
@@ -100,8 +103,8 @@ shade_quad(
 
    /* run shader */
    /* XXX: Generated code effectively unusable until it handles quad->mask */
-   if( !quad->mask && softpipe->fs->executable != NULL ) {
-      codegen_function func = (codegen_function) softpipe->fs->executable;
+   if( !quad->mask ) {
+      codegen_function func = (codegen_function) x86_get_func( &softpipe->fs->sse2_program );
       func(
          machine->Inputs,
          machine->Outputs,
@@ -119,7 +122,7 @@ shade_quad(
    /* store result color */
    if (qss->colorOutSlot >= 0) {
       /* XXX need to handle multiple color outputs someday */
-      assert(qss->stage.softpipe->fs->output_semantic_name[qss->colorOutSlot]
+      assert(qss->stage.softpipe->fs->shader.output_semantic_name[qss->colorOutSlot]
              == TGSI_SEMANTIC_COLOR);
       memcpy(
              quad->outputs.color,
@@ -167,15 +170,15 @@ static void shade_begin(struct quad_stage *qs)
 
    /* XXX only do this if the fragment shader changes... */
    tgsi_exec_machine_init(&qss->machine,
-                          softpipe->fs->tokens,
+                          softpipe->fs->shader.tokens,
                           PIPE_MAX_SAMPLERS,
                           qss->samplers );
 
    /* find output slots for depth, color */
    qss->colorOutSlot = -1;
    qss->depthOutSlot = -1;
-   for (i = 0; i < qss->stage.softpipe->fs->num_outputs; i++) {
-      switch (qss->stage.softpipe->fs->output_semantic_name[i]) {
+   for (i = 0; i < qss->stage.softpipe->fs->shader.num_outputs; i++) {
+      switch (qss->stage.softpipe->fs->shader.output_semantic_name[i]) {
       case TGSI_SEMANTIC_POSITION:
          qss->depthOutSlot = i;
          break;
@@ -190,9 +193,9 @@ static void shade_begin(struct quad_stage *qs)
 }
 
 
-static void shade_destroy(struct quad_stage *qs)

-{

-   free( qs );

+static void shade_destroy(struct quad_stage *qs)
+{
+   free( qs );
 }
 
 
diff --git a/src/mesa/pipe/softpipe/sp_state.h b/src/mesa/pipe/softpipe/sp_state.h
index 61532bcdebb..676ad06bfc5 100644
--- a/src/mesa/pipe/softpipe/sp_state.h
+++ b/src/mesa/pipe/softpipe/sp_state.h
@@ -33,6 +33,17 @@
 
 #include "pipe/p_state.h"
 
+#include "x86/rtasm/x86sse.h"
+
+/**
+ * Softpipe fs state is derived from pipe_shader_state.
+ */
+struct sp_fragment_shader_state {
+   struct pipe_shader_state   shader;
+#if defined(__i386__) || defined(__386__)
+   struct x86_function        sse2_program;
+#endif
+};
 
 void *
 softpipe_create_alpha_test_state(struct pipe_context *,
diff --git a/src/mesa/pipe/softpipe/sp_state_fs.c b/src/mesa/pipe/softpipe/sp_state_fs.c
index 5547c849ac9..86aa80c0fca 100644
--- a/src/mesa/pipe/softpipe/sp_state_fs.c
+++ b/src/mesa/pipe/softpipe/sp_state_fs.c
@@ -43,16 +43,14 @@ void * softpipe_create_fs_state(struct pipe_context *pipe,
     * that now.
     */
 
-   struct pipe_shader_state *state = malloc(sizeof(struct pipe_shader_state));
-   memcpy(state, templ, sizeof(struct pipe_shader_state));
+   struct sp_fragment_shader_state *state = malloc(sizeof(struct sp_fragment_shader_state));
+   state->shader = *templ;
 
 #if defined(__i386__) || defined(__386__)
    if (softpipe->use_sse) {
       x86_init_func( &state->sse2_program );
 
-      tgsi_emit_sse2_fs( state->tokens, &state->sse2_program );
-
-      state->executable = (void *)x86_get_func( &state->sse2_program );
+      tgsi_emit_sse2_fs( state->shader.tokens, &state->sse2_program );
    }
 #endif
 
@@ -63,7 +61,7 @@ void softpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
 
-   softpipe->fs = (struct pipe_shader_state *)fs;
+   softpipe->fs = (struct sp_fragment_shader_state *) fs;
 
    softpipe->dirty |= SP_NEW_FS;
 }
@@ -72,7 +70,7 @@ void softpipe_delete_fs_state(struct pipe_context *pipe,
                               void *shader)
 {
 #if defined(__i386__) || defined(__386__)
-   struct pipe_shader_state *state = shader;
+   struct sp_fragment_shader_state *state = shader;
 
    x86_release_func( &state->sse2_program );
 #endif