wined3d: Add dp2add support to arb shaders.

dlls/wined3d/arb_program_shader.c

@@ -878,6 +878,33 @@ void pshader_hw_cmp(SHADER_OPCODE_ARG* arg) {
         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 }
 
+/** Process the WINED3DSIO_DP2ADD instruction in ARB.
+ * dst = dot2(src0, src1) + src2 */
+void pshader_hw_dp2add(SHADER_OPCODE_ARG* arg) {
+    SHADER_BUFFER* buffer = arg->buffer;
+    char dst_wmask[20];
+    char dst_name[50];
+    char src_name[3][50];
+    DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
+    BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
+
+    pshader_get_register_name(arg->dst, dst_name);
+    shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
+
+    pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
+    pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
+    pshader_gen_input_modifier_line(buffer, arg->src[2], 2, src_name[2]);
+
+    /* Emulate a DP2 with a DP3 and 0.0 */
+    shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
+    shader_addline(buffer, "MOV TMP.z, 0.0;\n");
+    shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
+    shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
+
+    if (shift != 0)
+        pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
+}
+
 /* Map the opcode 1-to-1 to the GL code */
 void pshader_hw_map2gl(SHADER_OPCODE_ARG* arg) {
 

dlls/wined3d/pixelshader.c

@@ -189,8 +189,7 @@ CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
     {WINED3DSIO_NRM,      "nrm",      NULL, 1, 2, NULL, shader_glsl_map2gl, 0, 0},
     {WINED3DSIO_SINCOS,   "sincos",   NULL, 1, 4, NULL, shader_glsl_sincos, WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
     {WINED3DSIO_SINCOS,   "sincos",   NULL, 1, 2, NULL, shader_glsl_sincos, WINED3DPS_VERSION(3,0), -1},
-    /* TODO: dp2add can be made out of multiple instuctions */
-    {WINED3DSIO_DP2ADD,   "dp2add",   GLNAME_REQUIRE_GLSL,  1, 4, NULL, pshader_glsl_dp2add, WINED3DPS_VERSION(2,0), -1},
+    {WINED3DSIO_DP2ADD,   "dp2add",   NULL, 1, 4, pshader_hw_dp2add, pshader_glsl_dp2add, WINED3DPS_VERSION(2,0), -1},
     /* Matrix */
     {WINED3DSIO_M4x4, "m4x4", "undefined", 1, 3, NULL, shader_glsl_mnxn, 0, 0},
     {WINED3DSIO_M4x3, "m4x3", "undefined", 1, 3, NULL, shader_glsl_mnxn, 0, 0},

dlls/wined3d/wined3d_private.h

@@ -1786,6 +1786,7 @@ extern void pshader_hw_texdp3tex(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texdp3(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texm3x3(SHADER_OPCODE_ARG* arg);
 extern void pshader_hw_texm3x2depth(SHADER_OPCODE_ARG* arg);
+extern void pshader_hw_dp2add(SHADER_OPCODE_ARG* arg);
 
 /* ARB vertex shader prototypes */
 extern void vshader_hw_map2gl(SHADER_OPCODE_ARG* arg);