wined3d: Fix texdepth instruction.

dlls/wined3d/arb_program_shader.c

@@ -884,6 +884,32 @@ void pshader_hw_texm3x3spec(SHADER_OPCODE_ARG* arg) {
     current_state->current_row = 0;
 }
 
+void pshader_hw_texdepth(SHADER_OPCODE_ARG* arg) {
+    SHADER_BUFFER* buffer = arg->buffer;
+    char dst_name[50];
+
+    /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
+     * which is essentially an input, is the destiantion register because it is the first
+     * param. According to the msdn, this must be register r5, but let's keep it more flexible
+     * here
+     */
+    pshader_get_register_name(arg->dst, dst_name);
+
+    /* According to the msdn, the source register(must be r5) is unusable after
+     * the texdepth instruction, so we're free to modify it
+     */
+    shader_addline(buffer, "MIN %s.g, %s.g, one.g;\n", dst_name, dst_name);
+
+    /* How to deal with the special case dst_name.g == 0? if r != 0, then
+     * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
+     * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
+     */
+    shader_addline(buffer, "RCP %s.g, %s.g;\n", dst_name, dst_name);
+    shader_addline(buffer, "MUL TMP.x, %s.r, %s.g;\n", dst_name, dst_name);
+    shader_addline(buffer, "MIN TMP.x, TMP.x, one.r;\n", dst_name, dst_name);
+    shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n", dst_name, dst_name);
+}
+
 /** Handles transforming all WINED3DSIO_M?x? opcodes for
     Vertex shaders to ARB_vertex_program codes */
 void vshader_hw_mnxn(SHADER_OPCODE_ARG* arg) {

dlls/wined3d/glsl_shader.c

@@ -1733,7 +1733,13 @@ void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
 
     shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
 
-    shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_param.reg_name, dst_param.reg_name);
+    /* Tests show that texdepth never returns anything below 0.0, and that r5.y is clamped to 1.0.
+     * Negative input is accepted, -0.25 / -0.5 returns 0.5. GL should clamp gl_FragDepth to [0;1], but
+     * this doesn't always work, so clamp the results manually. Wether or not the x value is clamped at 1
+     * too is irrelevant, since if x = 0, any y value < 1.0(and > 1.0 is not allowed) results in a result
+     * >= 1.0 or < 0.0
+     */
+    shader_addline(arg->buffer, "gl_FragDepth = (%s.y == 0.0) ? 1.0 : clamp((%s.x / min(%s.y, 1.0)), 0.0, 1.0);\n", dst_param.reg_name, dst_param.reg_name, dst_param.reg_name);
 }
 
 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:

dlls/wined3d/pixelshader.c

@@ -238,7 +238,7 @@ CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", GLNAME_REQUIRE_GLSL, 1, 2, NULL, pshader_glsl_texm3x2depth, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
     {WINED3DSIO_TEXDP3,   "texdp3",   GLNAME_REQUIRE_GLSL, 1, 2, NULL, pshader_glsl_texdp3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
     {WINED3DSIO_TEXM3x3,  "texm3x3",  GLNAME_REQUIRE_GLSL, 1, 2, NULL, pshader_glsl_texm3x3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
-    {WINED3DSIO_TEXDEPTH, "texdepth", GLNAME_REQUIRE_GLSL, 1, 1, NULL, pshader_glsl_texdepth, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
+    {WINED3DSIO_TEXDEPTH, "texdepth", NULL, 1, 1, pshader_hw_texdepth, pshader_glsl_texdepth, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
     {WINED3DSIO_BEM,      "bem",      "undefined",         1, 3, pshader_hw_bem, pshader_glsl_bem, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
     {WINED3DSIO_DSX,      "dsx",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},
     {WINED3DSIO_DSY,      "dsy",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},

dlls/wined3d/wined3d_private.h

@@ -1785,6 +1785,7 @@ extern void pshader_hw_texm3x3pad(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texm3x3tex(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texm3x3spec(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texm3x3vspec(SHADER_OPCODE_ARG* arg);
+extern void pshader_hw_texdepth(SHADER_OPCODE_ARG* arg);
 
 /* ARB vertex shader prototypes */
 extern void vshader_hw_map2gl(SHADER_OPCODE_ARG* arg);