Stefan Dösinger : wined3d: Don' t enable the NV frag extensions if we don't need them.

[Alexandre Julliard]

Module: wine
Branch: master
Commit: de12f880ced7318ba0f8a62946fe0bd292ce66d9
URL:    http://source.winehq.org/git/wine.git/?a=commit;h=de12f880ced7318ba0f8a62946fe0bd292ce66d9

Author: Stefan Dösinger <stefan at codeweavers.com>
Date:   Fri May 29 17:11:04 2009 +0200

wined3d: Don't enable the NV frag extensions if we don't need them.

Enabling the NV extensions occupies a temp register for some reason. Avoid
needlessly enabling it.

---

 dlls/wined3d/arb_program_shader.c |   36 ++++++++++++++++++++++++++++++++++--
 dlls/wined3d/baseshader.c         |    4 ++++
 dlls/wined3d/wined3d_private.h    |    2 +-
 3 files changed, 39 insertions(+), 3 deletions(-)

diff --git a/dlls/wined3d/arb_program_shader.c b/dlls/wined3d/arb_program_shader.c
index 2a34567..0d32bbc 100644
--- a/dlls/wined3d/arb_program_shader.c
+++ b/dlls/wined3d/arb_program_shader.c
@@ -2058,6 +2058,7 @@ static GLuint shader_arb_generate_pshader(IWineD3DPixelShaderImpl *This,
     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
     struct shader_arb_ctx_priv priv_ctx;
     BOOL dcl_tmp = args->super.srgb_correction, dcl_td = FALSE;
+    BOOL want_nv_prog = FALSE;
 
     char srgbtmp[4][4];
     unsigned int i, found = 0;
@@ -2103,14 +2104,38 @@ static GLuint shader_arb_generate_pshader(IWineD3DPixelShaderImpl *This,
     priv_ctx.cur_ps_args = args;
     list_init(&priv_ctx.if_frames);
 
+    /* Avoid enabling NV_fragment_program* if we do not need it.
+     *
+     * Enabling GL_NV_fragment_program_option causes the driver to occupy a temporary register,
+     * and it slows down the shader execution noticeably(about 5%). Usually our instruction emulation
+     * is faster than what we gain from using higher native instructions. There are some things though
+     * that cannot be emulated. In that case enable the extensions.
+     * If the extension is enabled, instruction handlers that support both ways will use it.
+     *
+     * Testing shows no performance difference between OPTION NV_fragment_program2 and NV_fragment_program.
+     * So enable the best we can get.
+     */
+    if(reg_maps->usesdsx || reg_maps->usesdsy || reg_maps->loop_depth > 0)
+    {
+        want_nv_prog = TRUE;
+    }
+
     shader_addline(buffer, "!!ARBfp1.0\n");
-    if(GL_SUPPORT(NV_FRAGMENT_PROGRAM2)) {
+    if(want_nv_prog && GL_SUPPORT(NV_FRAGMENT_PROGRAM2)) {
         shader_addline(buffer, "OPTION NV_fragment_program2;\n");
         priv_ctx.target_version = NV3;
-    } else if(GL_SUPPORT(NV_FRAGMENT_PROGRAM_OPTION)) {
+    } else if(want_nv_prog && GL_SUPPORT(NV_FRAGMENT_PROGRAM_OPTION)) {
         shader_addline(buffer, "OPTION NV_fragment_program;\n");
         priv_ctx.target_version = NV2;
     } else {
+        if(want_nv_prog)
+        {
+            /* This is an error - either we're advertising the wrong shader version, or aren't enforcing some
+             * limits properly
+             */
+            ERR("The shader requires instructions that are not available in plain GL_ARB_fragment_program\n");
+            ERR("Try GLSL\n");
+        }
         priv_ctx.target_version = ARB;
     }
 
@@ -2131,6 +2156,10 @@ static GLuint shader_arb_generate_pshader(IWineD3DPixelShaderImpl *This,
         }
     }
 
+    /* For now always declare the temps. At least the Nvidia assembler optimizes completely
+     * unused temps away(but occupies them for the whole shader if they're used once). Always
+     * declaring them avoids tricky bookkeeping work
+     */
     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
@@ -2227,6 +2256,9 @@ static GLuint shader_arb_generate_vshader(IWineD3DVertexShaderImpl *This,
     /*  Create the hw ARB shader */
     shader_addline(buffer, "!!ARBvp1.0\n");
 
+    /* Always enable the NV extension if available. Unlike fragment shaders, there is no
+     * mesurable performance penalty, and we can always make use of it for clipplanes.
+     */
     if(GL_SUPPORT(NV_VERTEX_PROGRAM2_OPTION)) {
         shader_addline(buffer, "OPTION NV_vertex_program2;\n");
         priv_ctx.target_version = NV2;
diff --git a/dlls/wined3d/baseshader.c b/dlls/wined3d/baseshader.c
index 9dcdbe6..1a34fa2 100644
--- a/dlls/wined3d/baseshader.c
+++ b/dlls/wined3d/baseshader.c
@@ -683,6 +683,10 @@ HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3
             {
                 reg_maps->usesdsy = 1;
             }
+            else if (ins.handler_idx == WINED3DSIH_DSX)
+            {
+                reg_maps->usesdsx = 1;
+            }
             else if(ins.handler_idx == WINED3DSIH_TEXLDD)
             {
                 reg_maps->usestexldd = 1;
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 8b7bff8..f217366 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -632,7 +632,7 @@ typedef struct shader_reg_maps
 
     WINED3DSAMPLER_TEXTURE_TYPE sampler_type[max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS)];
     BOOL bumpmat[MAX_TEXTURES], luminanceparams[MAX_TEXTURES];
-    char usesnrm, vpos, usesdsy, usestexldd, usesmova;
+    char usesnrm, vpos, usesdsx, usesdsy, usestexldd, usesmova;
     char usesrelconstF;
 
     /* Whether or not loops are used in this shader, and nesting depth */