?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Implemented pixel shaders using ARB_fragment_program extension.
[wine/hacks.git] / dlls / d3d8 / shader.c
blob8ca91bbe85e4e451a637c2d270e8ce690aaed52f
1 /*
2 * shaders implementation
3 *
4 * Copyright 2002-2004 Raphael Junqueira
5 * Copyright 2004 Jason Edmeades
6 * Copyright 2004 Christian Costa
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 */
22
23 #include "config.h"
24
25 #include <math.h>
26 #include <stdarg.h>
27 #include <stdio.h>
28
29 #include "windef.h"
30 #include "winbase.h"
31 #include "winuser.h"
32 #include "wingdi.h"
33 #include "wine/debug.h"
34
35 #include "d3d8_private.h"
36
37 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
38 WINE_DECLARE_DEBUG_CHANNEL(d3d_hw_shader);
39
40 /* Shader debugging - Change the following line to enable debugging of software
41 vertex shaders */
42 #if 0 /* Must not be 1 in cvs version */
43 # define VSTRACE(A) TRACE A
44 # define TRACE_VSVECTOR(name) TRACE( #name "=(%f, %f, %f, %f)\n", name.x, name.y, name.z, name.w)
45 #else
46 # define VSTRACE(A)
47 # define TRACE_VSVECTOR(name)
48 #endif
49
50 /**
51 * DirectX9 SDK download
52 * http://msdn.microsoft.com/library/default.asp?url=/downloads/list/directx.asp
53 *
54 * Exploring D3DX
55 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dndrive/html/directx07162002.asp
56 *
57 * Using Vertex Shaders
58 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dndrive/html/directx02192001.asp
59 *
60 * Dx9 New
61 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/whatsnew.asp
62 *
63 * Dx9 Shaders
64 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/Shaders/VertexShader2_0/VertexShader2_0.asp
65 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/Shaders/VertexShader2_0/Instructions/Instructions.asp
66 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/programmingguide/GettingStarted/VertexDeclaration/VertexDeclaration.asp
67 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/Shaders/VertexShader3_0/VertexShader3_0.asp
68 *
69 * Dx9 D3DX
70 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/programmingguide/advancedtopics/VertexPipe/matrixstack/matrixstack.asp
71 *
72 * FVF
73 * http://msdn.microsoft.com/library/en-us/directx9_c/directx/graphics/programmingguide/GettingStarted/VertexFormats/vformats.asp
74 *
75 * NVIDIA: DX8 Vertex Shader to NV Vertex Program
76 * http://developer.nvidia.com/view.asp?IO=vstovp
77 *
78 * NVIDIA: Memory Management with VAR
79 * http://developer.nvidia.com/view.asp?IO=var_memory_management
80 */
81
82 typedef void (*shader_fct_t)();
83
84 typedef struct SHADER_OPCODE {
85 CONST BYTE opcode;
86 const char* name;
87 CONST UINT num_params;
88 shader_fct_t soft_fct;
89 DWORD min_version;
90 DWORD max_version;
91 } SHADER_OPCODE;
92
93 /*******************************
94 * vshader functions software VM
95 */
96
97 void vshader_add(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
98 d->x = s0->x + s1->x;
99 d->y = s0->y + s1->y;
100 d->z = s0->z + s1->z;
101 d->w = s0->w + s1->w;
102 VSTRACE(("executing add: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
103 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
104 }
105
106 void vshader_dp3(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
107 d->x = d->y = d->z = d->w = s0->x * s1->x + s0->y * s1->y + s0->z * s1->z;
108 VSTRACE(("executing dp3: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
109 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
110 }
111
112 void vshader_dp4(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
113 d->x = d->y = d->z = d->w = s0->x * s1->x + s0->y * s1->y + s0->z * s1->z + s0->w * s1->w;
114 VSTRACE(("executing dp4: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
115 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
116 }
117
118 void vshader_dst(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
119 d->x = 1.0f;
120 d->y = s0->y * s1->y;
121 d->z = s0->z;
122 d->w = s1->w;
123 VSTRACE(("executing dst: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
124 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
125 }
126
127 void vshader_expp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
128 union {
129 float f;
130 DWORD d;
131 } tmp;
132
133 tmp.f = floorf(s0->w);
134 d->x = powf(2.0f, tmp.f);
135 d->y = s0->w - tmp.f;
136
137 tmp.f = powf(2.0f, s0->w);
138 tmp.d &= 0xFFFFFF00U;
139 d->z = tmp.f;
140 d->w = 1.0f;
141 VSTRACE(("executing exp: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
142 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
143 }
144
145 void vshader_lit(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
146 d->x = 1.0f;
147 d->y = (0.0f < s0->x) ? s0->x : 0.0f;
148 d->z = (0.0f < s0->x && 0.0f < s0->y) ? powf(s0->y, s0->w) : 0.0f;
149 d->w = 1.0f;
150 VSTRACE(("executing lit: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
151 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
152 }
153
154 void vshader_logp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
155 float tmp_f = fabsf(s0->w);
156 d->x = d->y = d->z = d->w = (0.0f != tmp_f) ? logf(tmp_f) / logf(2.0f) : -HUGE_VAL;
157 VSTRACE(("executing logp: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
158 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
159 }
160
161 void vshader_mad(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1, D3DSHADERVECTOR* s2) {
162 d->x = s0->x * s1->x + s2->x;
163 d->y = s0->y * s1->y + s2->y;
164 d->z = s0->z * s1->z + s2->z;
165 d->w = s0->w * s1->w + s2->w;
166 VSTRACE(("executing mad: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) s2=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
167 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, s2->x, s2->y, s2->z, s2->w, d->x, d->y, d->z, d->w));
168 }
169
170 void vshader_max(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
171 d->x = (s0->x >= s1->x) ? s0->x : s1->x;
172 d->y = (s0->y >= s1->y) ? s0->y : s1->y;
173 d->z = (s0->z >= s1->z) ? s0->z : s1->z;
174 d->w = (s0->w >= s1->w) ? s0->w : s1->w;
175 VSTRACE(("executing max: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
176 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
177 }
178
179 void vshader_min(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
180 d->x = (s0->x < s1->x) ? s0->x : s1->x;
181 d->y = (s0->y < s1->y) ? s0->y : s1->y;
182 d->z = (s0->z < s1->z) ? s0->z : s1->z;
183 d->w = (s0->w < s1->w) ? s0->w : s1->w;
184 VSTRACE(("executing min: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
185 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
186 }
187
188 void vshader_mov(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
189 d->x = s0->x;
190 d->y = s0->y;
191 d->z = s0->z;
192 d->w = s0->w;
193 VSTRACE(("executing mov: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
194 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
195 }
196
197 void vshader_mul(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
198 d->x = s0->x * s1->x;
199 d->y = s0->y * s1->y;
200 d->z = s0->z * s1->z;
201 d->w = s0->w * s1->w;
202 VSTRACE(("executing mul: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
203 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
204 }
205
206 void vshader_nop(void) {
207 /* NOPPPP ahhh too easy ;) */
208 }
209
210 void vshader_rcp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
211 d->x = d->y = d->z = d->w = (0.0f == s0->w) ? HUGE_VAL : 1.0f / s0->w;
212 VSTRACE(("executing rcp: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
213 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
214 }
215
216 void vshader_rsq(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
217 float tmp_f = fabsf(s0->w);
218 d->x = d->y = d->z = d->w = (0.0f == tmp_f) ? HUGE_VAL : ((1.0f != tmp_f) ? 1.0f / sqrtf(tmp_f) : 1.0f);
219 VSTRACE(("executing rsq: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
220 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
221 }
222
223 void vshader_sge(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
224 d->x = (s0->x >= s1->x) ? 1.0f : 0.0f;
225 d->y = (s0->y >= s1->y) ? 1.0f : 0.0f;
226 d->z = (s0->z >= s1->z) ? 1.0f : 0.0f;
227 d->w = (s0->w >= s1->w) ? 1.0f : 0.0f;
228 VSTRACE(("executing sge: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
229 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
230 }
231
232 void vshader_slt(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
233 d->x = (s0->x < s1->x) ? 1.0f : 0.0f;
234 d->y = (s0->y < s1->y) ? 1.0f : 0.0f;
235 d->z = (s0->z < s1->z) ? 1.0f : 0.0f;
236 d->w = (s0->w < s1->w) ? 1.0f : 0.0f;
237 VSTRACE(("executing slt: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
238 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
239 }
240
241 void vshader_sub(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
242 d->x = s0->x - s1->x;
243 d->y = s0->y - s1->y;
244 d->z = s0->z - s1->z;
245 d->w = s0->w - s1->w;
246 VSTRACE(("executing sub: s0=(%f, %f, %f, %f) s1=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
247 s0->x, s0->y, s0->z, s0->w, s1->x, s1->y, s1->z, s1->w, d->x, d->y, d->z, d->w));
248 }
249
250 /**
251 * Version 1.1 specific
252 */
253
254 void vshader_exp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
255 d->x = d->y = d->z = d->w = powf(2.0f, s0->w);
256 VSTRACE(("executing exp: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
257 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
258 }
259
260 void vshader_log(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
261 float tmp_f = fabsf(s0->w);
262 d->x = d->y = d->z = d->w = (0.0f != tmp_f) ? logf(tmp_f) / logf(2.0f) : -HUGE_VAL;
263 VSTRACE(("executing log: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
264 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
265 }
266
267 void vshader_frc(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
268 d->x = s0->x - floorf(s0->x);
269 d->y = s0->y - floorf(s0->y);
270 d->z = 0.0f;
271 d->w = 1.0f;
272 VSTRACE(("executing frc: s0=(%f, %f, %f, %f) => d=(%f, %f, %f, %f)\n",
273 s0->x, s0->y, s0->z, s0->w, d->x, d->y, d->z, d->w));
274 }
275
276 typedef FLOAT D3DMATRIX44[4][4];
277 typedef FLOAT D3DMATRIX43[4][3];
278 typedef FLOAT D3DMATRIX34[4][4];
279 typedef FLOAT D3DMATRIX33[4][3];
280 typedef FLOAT D3DMATRIX32[4][2];
281
282 void vshader_m4x4(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, /*D3DSHADERVECTOR* mat1*/ D3DMATRIX44 mat) {
283 /*
284 * BuGGY CODE: here only if cast not work for copy/paste
285 D3DSHADERVECTOR* mat2 = mat1 + 1;
286 D3DSHADERVECTOR* mat3 = mat1 + 2;
287 D3DSHADERVECTOR* mat4 = mat1 + 3;
288 d->x = mat1->x * s0->x + mat2->x * s0->y + mat3->x * s0->z + mat4->x * s0->w;
289 d->y = mat1->y * s0->x + mat2->y * s0->y + mat3->y * s0->z + mat4->y * s0->w;
290 d->z = mat1->z * s0->x + mat2->z * s0->y + mat3->z * s0->z + mat4->z * s0->w;
291 d->w = mat1->w * s0->x + mat2->w * s0->y + mat3->w * s0->z + mat4->w * s0->w;
292 */
293 d->x = mat[0][0] * s0->x + mat[0][1] * s0->y + mat[0][2] * s0->z + mat[0][3] * s0->w;
294 d->y = mat[1][0] * s0->x + mat[1][1] * s0->y + mat[1][2] * s0->z + mat[1][3] * s0->w;
295 d->z = mat[2][0] * s0->x + mat[2][1] * s0->y + mat[2][2] * s0->z + mat[2][3] * s0->w;
296 d->w = mat[3][0] * s0->x + mat[3][1] * s0->y + mat[3][2] * s0->z + mat[3][3] * s0->w;
297 VSTRACE(("executing m4x4(1): mat=(%f, %f, %f, %f) s0=(%f) d=(%f) \n", mat[0][0], mat[0][1], mat[0][2], mat[0][3], s0->x, d->x));
298 VSTRACE(("executing m4x4(2): mat=(%f, %f, %f, %f) (%f) (%f) \n", mat[1][0], mat[1][1], mat[1][2], mat[1][3], s0->y, d->y));
299 VSTRACE(("executing m4x4(3): mat=(%f, %f, %f, %f) X (%f) = (%f) \n", mat[2][0], mat[2][1], mat[2][2], mat[2][3], s0->z, d->z));
300 VSTRACE(("executing m4x4(4): mat=(%f, %f, %f, %f) (%f) (%f) \n", mat[3][0], mat[3][1], mat[3][2], mat[3][3], s0->w, d->w));
301 }
302
303 void vshader_m4x3(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DMATRIX43 mat) {
304 d->x = mat[0][0] * s0->x + mat[0][1] * s0->y + mat[0][2] * s0->z + mat[0][3] * s0->w;
305 d->y = mat[1][0] * s0->x + mat[1][1] * s0->y + mat[1][2] * s0->z + mat[1][3] * s0->w;
306 d->z = mat[2][0] * s0->x + mat[2][1] * s0->y + mat[2][2] * s0->z + mat[2][3] * s0->w;
307 d->w = 1.0f;
308 VSTRACE(("executing m4x3(1): mat=(%f, %f, %f, %f) s0=(%f) d=(%f) \n", mat[0][0], mat[0][1], mat[0][2], mat[0][3], s0->x, d->x));
309 VSTRACE(("executing m4x3(2): mat=(%f, %f, %f, %f) (%f) (%f) \n", mat[1][0], mat[1][1], mat[1][2], mat[1][3], s0->y, d->y));
310 VSTRACE(("executing m4x3(3): mat=(%f, %f, %f, %f) X (%f) = (%f) \n", mat[2][0], mat[2][1], mat[2][2], mat[2][3], s0->z, d->z));
311 VSTRACE(("executing m4x3(4): (%f) (%f) \n", s0->w, d->w));
312 }
313
314 void vshader_m3x4(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DMATRIX34 mat) {
315 d->x = mat[0][0] * s0->x + mat[0][1] * s0->y + mat[0][2] * s0->z;
316 d->y = mat[1][0] * s0->x + mat[1][1] * s0->y + mat[1][2] * s0->z;
317 d->z = mat[2][0] * s0->x + mat[2][1] * s0->y + mat[2][2] * s0->z;
318 d->w = mat[3][0] * s0->x + mat[3][1] * s0->y + mat[3][2] * s0->z;
319 VSTRACE(("executing m3x4(1): mat=(%f, %f, %f) s0=(%f) d=(%f) \n", mat[0][0], mat[0][1], mat[0][2], s0->x, d->x));
320 VSTRACE(("executing m3x4(2): mat=(%f, %f, %f) (%f) (%f) \n", mat[1][0], mat[1][1], mat[1][2], s0->y, d->y));
321 VSTRACE(("executing m3x4(3): mat=(%f, %f, %f) X (%f) = (%f) \n", mat[2][0], mat[2][1], mat[2][2], s0->z, d->z));
322 VSTRACE(("executing m3x4(4): mat=(%f, %f, %f) (%f) (%f) \n", mat[3][0], mat[3][1], mat[3][2], s0->w, d->w));
323 }
324
325 void vshader_m3x3(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DMATRIX33 mat) {
326 d->x = mat[0][0] * s0->x + mat[0][1] * s0->y + mat[0][2] * s0->z;
327 d->y = mat[1][0] * s0->x + mat[1][1] * s0->y + mat[1][2] * s0->z;
328 d->z = mat[2][0] * s0->x + mat[2][1] * s0->y + mat[2][2] * s0->z;
329 d->w = 1.0f;
330 VSTRACE(("executing m3x3(1): mat=(%f, %f, %f) s0=(%f) d=(%f) \n", mat[0][0], mat[0][1], mat[0][2], s0->x, d->x));
331 VSTRACE(("executing m3x3(2): mat=(%f, %f, %f) (%f) (%f) \n", mat[1][0], mat[1][1], mat[1][2], s0->y, d->y));
332 VSTRACE(("executing m3x3(3): mat=(%f, %f, %f) X (%f) = (%f) \n", mat[2][0], mat[2][1], mat[2][2], s0->z, d->z));
333 VSTRACE(("executing m3x3(4): (%f) \n", d->w));
334 }
335
336 void vshader_m3x2(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DMATRIX32 mat) {
337 FIXME("check\n");
338 d->x = mat[0][0] * s0->x + mat[0][1] * s0->y + mat[0][2] * s0->z;
339 d->y = mat[1][0] * s0->x + mat[1][1] * s0->y + mat[1][2] * s0->z;
340 d->z = 0.0f;
341 d->w = 1.0f;
342 }
343
344 /**
345 * Version 2.0 specific
346 */
347 void vshader_lrp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1, D3DSHADERVECTOR* s2, D3DSHADERVECTOR* s3) {
348 d->x = s0->x * (s1->x - s2->x) + s2->x;
349 d->y = s0->y * (s1->y - s2->y) + s2->y;
350 d->z = s0->z * (s1->z - s2->z) + s2->z;
351 d->w = s0->w * (s1->w - s2->w) + s2->x;
352 }
353
354 /**
355 * log, exp, frc, m*x* seems to be macros ins ... to see
356 * Note opcode must be in uppercase if direct mapping to GL hw shaders
357 */
358 static CONST SHADER_OPCODE vshader_ins [] = {
359 {D3DSIO_NOP, "NOP", 0, vshader_nop, 0, 0},
360 {D3DSIO_MOV, "MOV", 2, vshader_mov, 0, 0},
361 {D3DSIO_ADD, "ADD", 3, vshader_add, 0, 0},
362 {D3DSIO_SUB, "SUB", 3, vshader_sub, 0, 0},
363 {D3DSIO_MAD, "MAD", 4, vshader_mad, 0, 0},
364 {D3DSIO_MUL, "MUL", 3, vshader_mul, 0, 0},
365 {D3DSIO_RCP, "RCP", 2, vshader_rcp, 0, 0},
366 {D3DSIO_RSQ, "RSQ", 2, vshader_rsq, 0, 0},
367 {D3DSIO_DP3, "DP3", 3, vshader_dp3, 0, 0},
368 {D3DSIO_DP4, "DP4", 3, vshader_dp4, 0, 0},
369 {D3DSIO_MIN, "MIN", 3, vshader_min, 0, 0},
370 {D3DSIO_MAX, "MAX", 3, vshader_max, 0, 0},
371 {D3DSIO_SLT, "SLT", 3, vshader_slt, 0, 0},
372 {D3DSIO_SGE, "SGE", 3, vshader_sge, 0, 0},
373 {D3DSIO_EXP, "EXP", 2, vshader_exp, 0, 0},
374 {D3DSIO_LOG, "LOG", 2, vshader_log, 0, 0},
375 {D3DSIO_LIT, "LIT", 2, vshader_lit, 0, 0},
376 {D3DSIO_DST, "DST", 3, vshader_dst, 0, 0},
377 {D3DSIO_LRP, "LRP", 5, vshader_lrp, 0, 0},
378 {D3DSIO_FRC, "FRC", 2, vshader_frc, 0, 0},
379 {D3DSIO_M4x4, "M4X4", 3, vshader_m4x4, 0, 0},
380 {D3DSIO_M4x3, "M4X3", 3, vshader_m4x3, 0, 0},
381 {D3DSIO_M3x4, "M3X4", 3, vshader_m3x4, 0, 0},
382 {D3DSIO_M3x3, "M3X3", 3, vshader_m3x3, 0, 0},
383 {D3DSIO_M3x2, "M3X2", 3, vshader_m3x2, 0, 0},
384 /** FIXME: use direct access so add the others opcodes as stubs */
385 {D3DSIO_EXPP, "EXPP", 2, vshader_expp, 0, 0},
386 {D3DSIO_LOGP, "LOGP", 2, vshader_logp, 0, 0},
387
388 {0, NULL, 0, NULL, 0, 0}
389 };
390
391
392 inline static const SHADER_OPCODE* vshader_program_get_opcode(const DWORD code) {
393 DWORD i = 0;
394 /** TODO: use dichotomic search */
395 while (NULL != vshader_ins[i].name) {
396 if ((code & D3DSI_OPCODE_MASK) == vshader_ins[i].opcode) {
397 return &vshader_ins[i];
398 }
399 ++i;
400 }
401 return NULL;
402 }
403
404 inline static BOOL vshader_is_version_token(DWORD token) {
405 return 0xFFFE0000 == (token & 0xFFFE0000);
406 }
407
408 inline static BOOL vshader_is_comment_token(DWORD token) {
409 return D3DSIO_COMMENT == (token & D3DSI_OPCODE_MASK);
410 }
411
412 inline static void vshader_program_dump_param(const DWORD param, int input) {
413 static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" };
414 static const char swizzle_reg_chars[] = "xyzw";
415
416 DWORD reg = param & 0x00001FFF;
417 DWORD regtype = ((param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT);
418
419 if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) TRACE("-");
420
421 switch (regtype << D3DSP_REGTYPE_SHIFT) {
422 case D3DSPR_TEMP:
423 TRACE("R[%lu]", reg);
424 break;
425 case D3DSPR_INPUT:
426 TRACE("V[%lu]", reg);
427 break;
428 case D3DSPR_CONST:
429 TRACE("C[%s%lu]", (param & D3DVS_ADDRMODE_RELATIVE) ? "a0.x + " : "", reg);
430 break;
431 case D3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
432 TRACE("a[%lu]", reg);
433 break;
434 case D3DSPR_RASTOUT:
435 TRACE("%s", rastout_reg_names[reg]);
436 break;
437 case D3DSPR_ATTROUT:
438 TRACE("oD[%lu]", reg);
439 break;
440 case D3DSPR_TEXCRDOUT:
441 TRACE("oT[%lu]", reg);
442 break;
443 default:
444 break;
445 }
446
447 if (!input) {
448 /** operand output */
449 if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
450 if (param & D3DSP_WRITEMASK_0) TRACE(".x");
451 if (param & D3DSP_WRITEMASK_1) TRACE(".y");
452 if (param & D3DSP_WRITEMASK_2) TRACE(".z");
453 if (param & D3DSP_WRITEMASK_3) TRACE(".w");
454 }
455 } else {
456 /** operand input */
457 DWORD swizzle = (param & D3DVS_SWIZZLE_MASK) >> D3DVS_SWIZZLE_SHIFT;
458 DWORD swizzle_x = swizzle & 0x03;
459 DWORD swizzle_y = (swizzle >> 2) & 0x03;
460 DWORD swizzle_z = (swizzle >> 4) & 0x03;
461 DWORD swizzle_w = (swizzle >> 6) & 0x03;
462 /**
463 * swizzle bits fields:
464 * WWZZYYXX
465 */
466 if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
467 if (swizzle_x == swizzle_y &&
468 swizzle_x == swizzle_z &&
469 swizzle_x == swizzle_w) {
470 TRACE(".%c", swizzle_reg_chars[swizzle_x]);
471 } else {
472 TRACE(".%c%c%c%c",
473 swizzle_reg_chars[swizzle_x],
474 swizzle_reg_chars[swizzle_y],
475 swizzle_reg_chars[swizzle_z],
476 swizzle_reg_chars[swizzle_w]);
477 }
478 }
479 }
480 }
481
482 inline static void vshader_program_add_param(const DWORD param, int input, char *hwLine) {
483 /*static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" }; */
484 static const char* hwrastout_reg_names[] = { "result.position", "result.fogcoord", "result.pointsize" };
485 static const char swizzle_reg_chars[] = "xyzw";
486
487 DWORD reg = param & 0x00001FFF;
488 DWORD regtype = ((param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT);
489 char tmpReg[255];
490
491 if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) {
492 strcat(hwLine, " -");
493 } else {
494 strcat(hwLine, " ");
495 }
496
497 switch (regtype << D3DSP_REGTYPE_SHIFT) {
498 case D3DSPR_TEMP:
499 sprintf(tmpReg, "T%lu", reg);
500 strcat(hwLine, tmpReg);
501 break;
502 case D3DSPR_INPUT:
503 sprintf(tmpReg, "vertex.attrib[%lu]", reg);
504 strcat(hwLine, tmpReg);
505 break;
506 case D3DSPR_CONST:
507 sprintf(tmpReg, "C[%s%lu]", (param & D3DVS_ADDRMODE_RELATIVE) ? "A0.x + " : "", reg);
508 strcat(hwLine, tmpReg);
509 break;
510 case D3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
511 sprintf(tmpReg, "A%lu", reg);
512 strcat(hwLine, tmpReg);
513 break;
514 case D3DSPR_RASTOUT:
515 sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
516 strcat(hwLine, tmpReg);
517 break;
518 case D3DSPR_ATTROUT:
519 if (reg==0) {
520 strcat(hwLine, "result.color.primary");
521 } else {
522 strcat(hwLine, "result.color.secondary");
523 }
524 break;
525 case D3DSPR_TEXCRDOUT:
526 sprintf(tmpReg, "result.texcoord[%lu]", reg);
527 strcat(hwLine, tmpReg);
528 break;
529 default:
530 break;
531 }
532
533 if (!input) {
534 /** operand output */
535 if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
536 strcat(hwLine, ".");
537 if (param & D3DSP_WRITEMASK_0) {
538 strcat(hwLine, "x");
539 }
540 if (param & D3DSP_WRITEMASK_1) {
541 strcat(hwLine, "y");
542 }
543 if (param & D3DSP_WRITEMASK_2) {
544 strcat(hwLine, "z");
545 }
546 if (param & D3DSP_WRITEMASK_3) {
547 strcat(hwLine, "w");
548 }
549 }
550 } else {
551 /** operand input */
552 DWORD swizzle = (param & D3DVS_SWIZZLE_MASK) >> D3DVS_SWIZZLE_SHIFT;
553 DWORD swizzle_x = swizzle & 0x03;
554 DWORD swizzle_y = (swizzle >> 2) & 0x03;
555 DWORD swizzle_z = (swizzle >> 4) & 0x03;
556 DWORD swizzle_w = (swizzle >> 6) & 0x03;
557 /**
558 * swizzle bits fields:
559 * WWZZYYXX
560 */
561 if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
562 if (swizzle_x == swizzle_y &&
563 swizzle_x == swizzle_z &&
564 swizzle_x == swizzle_w) {
565 sprintf(tmpReg, ".%c", swizzle_reg_chars[swizzle_x]);
566 strcat(hwLine, tmpReg);
567 } else {
568 sprintf(tmpReg, ".%c%c%c%c",
569 swizzle_reg_chars[swizzle_x],
570 swizzle_reg_chars[swizzle_y],
571 swizzle_reg_chars[swizzle_z],
572 swizzle_reg_chars[swizzle_w]);
573 strcat(hwLine, tmpReg);
574 }
575 }
576 }
577 }
578
579 DWORD MacroExpansion[4*4];
580
581 int ExpandMxMacro(DWORD macro_opcode, const DWORD* args) {
582 int i;
583 int nComponents = 0;
584 DWORD opcode =0;
585 switch(macro_opcode) {
586 case D3DSIO_M4x4:
587 nComponents = 4;
588 opcode = D3DSIO_DP4;
589 break;
590 case D3DSIO_M4x3:
591 nComponents = 3;
592 opcode = D3DSIO_DP4;
593 break;
594 case D3DSIO_M3x4:
595 nComponents = 4;
596 opcode = D3DSIO_DP3;
597 break;
598 case D3DSIO_M3x3:
599 nComponents = 3;
600 opcode = D3DSIO_DP3;
601 break;
602 case D3DSIO_M3x2:
603 nComponents = 2;
604 opcode = D3DSIO_DP3;
605 break;
606 default:
607 break;
608 }
609 for (i = 0; i < nComponents; i++) {
610 MacroExpansion[i*4+0] = opcode;
611 MacroExpansion[i*4+1] = ((*args) & ~D3DSP_WRITEMASK_ALL)|(D3DSP_WRITEMASK_0<<i);
612 MacroExpansion[i*4+2] = *(args+1);
613 MacroExpansion[i*4+3] = (*(args+2))+i;
614 }
615 return nComponents;
616 }
617
618 /**
619 * Function parser ...
620 */
621 inline static VOID IDirect3DVertexShaderImpl_GenerateProgramArbHW(IDirect3DVertexShaderImpl* vshader, CONST DWORD* pFunction) {
622 const DWORD* pToken = pFunction;
623 const DWORD* pSavedToken = NULL;
624 const SHADER_OPCODE* curOpcode = NULL;
625 int nRemInstr = -1;
626 DWORD i;
627 unsigned lineNum = 0;
628 char *pgmStr = NULL;
629 char tmpLine[255];
630 DWORD nUseAddressRegister = 0;
631 DWORD nUseTempRegister = 0;
632 DWORD regtype;
633 DWORD reg;
634 IDirect3DDevice8Impl* This = vshader->device;
635
636 pgmStr = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, 65535); /* 64kb should be enough */
637
638 /**
639 * First pass to determine what we need to declare:
640 * - Temporary variables
641 * - Address variables
642 */
643 if (NULL != pToken) {
644 while (D3DVS_END() != *pToken) {
645 if (vshader_is_version_token(*pToken)) {
646 /** skip version */
647 ++pToken;
648 continue;
649 }
650 if (vshader_is_comment_token(*pToken)) { /** comment */
651 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
652 ++pToken;
653 pToken += comment_len;
654 continue;
655 }
656 curOpcode = vshader_program_get_opcode(*pToken);
657 ++pToken;
658 if (NULL == curOpcode) {
659 while (*pToken & 0x80000000) {
660 /* skip unrecognized opcode */
661 ++pToken;
662 }
663 } else {
664 if (curOpcode->num_params > 0) {
665 regtype = ((((*pToken) & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT) << D3DSP_REGTYPE_SHIFT);
666 reg = ((*pToken) & 0x00001FFF);
667 /** we should validate GL_MAX_PROGRAM_ADDRESS_REGISTERS_AR limits here */
668 if (D3DSPR_ADDR == regtype && nUseAddressRegister <= reg) nUseAddressRegister = reg + 1;
669 /** we should validate GL_MAX_PROGRAM_LOCAL_PARAMETERS_ARB limits here */
670 if (D3DSPR_TEMP == regtype && nUseTempRegister <= reg) nUseTempRegister = reg + 1;
671 ++pToken;
672 for (i = 1; i < curOpcode->num_params; ++i) {
673 regtype = ((((*pToken) & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT) << D3DSP_REGTYPE_SHIFT);
674 reg = ((*pToken) & 0x00001FFF);
675 /** we should validate GL_MAX_PROGRAM_LOCAL_PARAMETERS_ARB limits here */
676 if (D3DSPR_TEMP == regtype && nUseTempRegister <= reg) nUseTempRegister = reg + 1;
677 ++pToken;
678 }
679 }
680 }
681 }
682 }
683
684 /** second pass, now generate */
685 pToken = pFunction;
686
687 if (NULL != pToken) {
688 while (D3DVS_END() != *pToken) {
689 tmpLine[0] = 0;
690
691 if ((nRemInstr >= 0) && (--nRemInstr == -1))
692 /* Macro is finished, continue normal path */
693 pToken = pSavedToken;
694
695 if (vshader_is_version_token(*pToken)) { /** version */
696
697 /* Extract version *10 into integer value (ie. 1.0 == 10, 1.1==11 etc */
698 int version = (((*pToken >> 8) & 0x0F) * 10) + (*pToken & 0x0F);
699 int numTemps;
700 int numConstants;
701
702 TRACE_(d3d_hw_shader)("vs.%lu.%lu;\n", (*pToken >> 8) & 0x0F, (*pToken & 0x0F));
703
704 /* Each release of vertex shaders has had different numbers of temp registers */
705 switch (version) {
706 case 10:
707 case 11: numTemps=12;
708 numConstants=96;
709 strcpy(tmpLine, "!!ARBvp1.0\n");
710 TRACE_(d3d_hw_shader)("GL HW (%u) : %s", strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
711 break;
712 case 20: numTemps=12;
713 numConstants=256;
714 strcpy(tmpLine, "!!ARBvp2.0\n");
715 FIXME_(d3d_hw_shader)("No work done yet to support vs2.0 in hw\n");
716 TRACE_(d3d_hw_shader)("GL HW (%u) : %s", strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
717 break;
718 case 30: numTemps=32;
719 numConstants=256;
720 strcpy(tmpLine, "!!ARBvp3.0\n");
721 FIXME_(d3d_hw_shader)("No work done yet to support vs3.0 in hw\n");
722 TRACE_(d3d_hw_shader)("GL HW (%u) : %s", strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
723 break;
724 default:
725 numTemps=12;
726 numConstants=96;
727 strcpy(tmpLine, "!!ARBvp1.0\n");
728 FIXME_(d3d_hw_shader)("Unrecognized vertex shader version!\n");
729 }
730 strcat(pgmStr,tmpLine);
731 ++lineNum;
732
733 for (i = 0; i < nUseTempRegister/*we should check numTemps here*/; i++) {
734 sprintf(tmpLine, "TEMP T%ld;\n", i);
735 ++lineNum;
736 TRACE_(d3d_hw_shader)("GL HW (%u, %u) : %s", lineNum, strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
737 strcat(pgmStr,tmpLine);
738 }
739 for (i = 0; i < nUseAddressRegister; i++) {
740 sprintf(tmpLine, "ADDRESS A%ld;\n", i);
741 ++lineNum;
742 TRACE_(d3d_hw_shader)("GL HW (%u, %u) : %s", lineNum, strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
743 strcat(pgmStr,tmpLine);
744 }
745 /* Due to the dynamic constants binding mechanism, we need to declare
746 * all the constants for relative addressing. */
747 /* Mesa supports only 95 constants for VS1.X although we should have at least 96.
748 * Let's declare max constants minus one for now. */
749 sprintf(tmpLine, "PARAM C[%d] = { program.env[0..%d] };\n", numConstants-1, numConstants-2);
750 TRACE("GL HW (%u) : %s", strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
751 strcat(pgmStr, tmpLine);
752
753 ++pToken;
754 continue;
755 }
756 if (vshader_is_comment_token(*pToken)) { /** comment */
757 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
758 ++pToken;
759 /*TRACE("comment[%ld] ;%s\n", comment_len, (char*)pToken);*/
760 pToken += comment_len;
761 continue;
762 }
763 curOpcode = vshader_program_get_opcode(*pToken);
764 ++pToken;
765 if (NULL == curOpcode) {
766 /* unkown current opcode ... */
767 while (*pToken & 0x80000000) {
768 TRACE_(d3d_hw_shader)("unrecognized opcode: %08lx\n", *pToken);
769 ++pToken;
770 }
771 } else {
772 /* Build opcode for GL vertex_program */
773 switch (curOpcode->opcode) {
774 case D3DSIO_MOV:
775 /* Address registers must be loaded with the ARL instruction */
776 if (((*pToken) & D3DSP_REGTYPE_MASK) == D3DSPR_ADDR) {
777 if (0 < nUseAddressRegister) {
778 strcpy(tmpLine, "ARL");
779 break;
780 } else
781 FIXME_(d3d_hw_shader)("Try to load an undeclared address register!\n");
782 }
783 /* fall through */
784 case D3DSIO_ADD:
785 case D3DSIO_SUB:
786 case D3DSIO_MAD:
787 case D3DSIO_MUL:
788 case D3DSIO_RCP:
789 case D3DSIO_RSQ:
790 case D3DSIO_DP3:
791 case D3DSIO_DP4:
792 case D3DSIO_MIN:
793 case D3DSIO_MAX:
794 case D3DSIO_SLT:
795 case D3DSIO_SGE:
796 case D3DSIO_LIT:
797 case D3DSIO_DST:
798 case D3DSIO_FRC:
799 strcpy(tmpLine, curOpcode->name);
800 break;
801
802 case D3DSIO_EXPP:
803 strcpy(tmpLine, "EXP");
804 break;
805 case D3DSIO_LOGP:
806 strcpy(tmpLine, "LOG");
807 break;
808 case D3DSIO_EXP:
809 strcpy(tmpLine, "EX2");
810 break;
811 case D3DSIO_LOG:
812 strcpy(tmpLine, "LG2");
813 break;
814
815 case D3DSIO_M4x4:
816 case D3DSIO_M4x3:
817 case D3DSIO_M3x4:
818 case D3DSIO_M3x3:
819 case D3DSIO_M3x2:
820 /* Expand the macro and get number of generated instruction */
821 nRemInstr = ExpandMxMacro(curOpcode->opcode, pToken);
822 /* Save point to next instruction */
823 pSavedToken = pToken + 3;
824 /* Execute expanded macro */
825 pToken = MacroExpansion;
826 continue;
827
828 default:
829 FIXME_(d3d_hw_shader)("Can't handle opcode %s in hwShader\n", curOpcode->name);
830 }
831
832 if (curOpcode->num_params > 0) {
833 vshader_program_add_param(*pToken, 0, tmpLine);
834
835 ++pToken;
836 for (i = 1; i < curOpcode->num_params; ++i) {
837 strcat(tmpLine, ",");
838 vshader_program_add_param(*pToken, 1, tmpLine);
839 ++pToken;
840 }
841 }
842 strcat(tmpLine,";\n");
843 ++lineNum;
844 TRACE_(d3d_hw_shader)("GL HW (%u, %u) : %s", lineNum, strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
845 strcat(pgmStr, tmpLine);
846 }
847 }
848 strcpy(tmpLine, "END\n");
849 ++lineNum;
850 TRACE_(d3d_hw_shader)("GL HW (%u, %u) : %s", lineNum, strlen(pgmStr), tmpLine); /* Don't add \n to this line as already in tmpLine */
851 strcat(pgmStr, tmpLine);
852 }
853
854 /* Create the hw shader */
855 GL_EXTCALL(glGenProgramsARB(1, &vshader->prgId));
856 TRACE_(d3d_hw_shader)("Creating a hw vertex shader, prg=%d\n", vshader->prgId);
857
858 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vshader->prgId));
859
860 /* Create the program and check for errors */
861 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(pgmStr), pgmStr));
862 if (glGetError() == GL_INVALID_OPERATION) {
863 GLint errPos;
864 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
865 FIXME_(d3d_hw_shader)("HW VertexShader Error at position: %d\n%s\n", errPos, glGetString(GL_PROGRAM_ERROR_STRING_ARB));
866 vshader->prgId = -1;
867 }
868
869 HeapFree(GetProcessHeap(), 0, pgmStr);
870 }
871
872 inline static VOID IDirect3DVertexShaderImpl_ParseProgram(IDirect3DVertexShaderImpl* vshader, CONST DWORD* pFunction, int useHW) {
873 const DWORD* pToken = pFunction;
874 const SHADER_OPCODE* curOpcode = NULL;
875 DWORD len = 0;
876 DWORD i;
877
878 if (NULL != pToken) {
879 while (D3DVS_END() != *pToken) {
880 if (vshader_is_version_token(*pToken)) { /** version */
881 TRACE("vs.%lu.%lu\n", (*pToken >> 8) & 0x0F, (*pToken & 0x0F));
882 ++pToken;
883 ++len;
884 continue;
885 }
886 if (vshader_is_comment_token(*pToken)) { /** comment */
887 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
888 ++pToken;
889 /*TRACE("comment[%ld] ;%s\n", comment_len, (char*)pToken);*/
890 pToken += comment_len;
891 len += comment_len + 1;
892 continue;
893 }
894 curOpcode = vshader_program_get_opcode(*pToken);
895 ++pToken;
896 ++len;
897 if (NULL == curOpcode) {
898 /* unkown current opcode ... */
899 while (*pToken & 0x80000000) {
900 TRACE("unrecognized opcode: %08lx\n", *pToken);
901 ++pToken;
902 ++len;
903 }
904 } else {
905 TRACE("%s ", curOpcode->name);
906 if (curOpcode->num_params > 0) {
907 vshader_program_dump_param(*pToken, 0);
908 ++pToken;
909 ++len;
910 for (i = 1; i < curOpcode->num_params; ++i) {
911 TRACE(", ");
912 vshader_program_dump_param(*pToken, 1);
913 ++pToken;
914 ++len;
915 }
916 }
917 TRACE("\n");
918 }
919 }
920 vshader->functionLength = (len + 1) * sizeof(DWORD);
921 } else {
922 vshader->functionLength = 1; /* no Function defined use fixed function vertex processing */
923 }
924
925 /* Generate HW shader in needed */
926 if (useHW && NULL != pFunction) {
927 IDirect3DVertexShaderImpl_GenerateProgramArbHW(vshader, pFunction);
928 }
929
930 /* copy the function ... because it will certainly be released by application */
931 if (NULL != pFunction) {
932 vshader->function = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, vshader->functionLength);
933 memcpy(vshader->function, pFunction, vshader->functionLength);
934 } else {
935 vshader->function = NULL;
936 }
937 }
938
939 HRESULT WINAPI IDirect3DDeviceImpl_CreateVertexShader(IDirect3DDevice8Impl* This, CONST DWORD* pFunction, DWORD Usage, IDirect3DVertexShaderImpl** ppVertexShader) {
940 IDirect3DVertexShaderImpl* object;
941 int useHW;
942
943 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(IDirect3DVertexShaderImpl));
944 if (NULL == object) {
945 *ppVertexShader = NULL;
946 return D3DERR_OUTOFVIDEOMEMORY;
947 }
948 /*object->lpVtbl = &Direct3DVextexShader9_Vtbl;*/
949 object->device = This; /* FIXME: AddRef(This) */
950 object->ref = 1;
951
952 object->usage = Usage;
953 object->data = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(VSHADERDATA8));
954
955 useHW = (((vs_mode == VS_HW) && GL_SUPPORT(ARB_VERTEX_PROGRAM)) &&
956 This->devType != D3DDEVTYPE_REF &&
957 object->usage != D3DUSAGE_SOFTWAREPROCESSING);
958
959 IDirect3DVertexShaderImpl_ParseProgram(object, pFunction, useHW);
960
961 *ppVertexShader = object;
962 return D3D_OK;
963 }
964
965 BOOL IDirect3DVertexShaderImpl_ExecuteHAL(IDirect3DVertexShaderImpl* vshader, VSHADERINPUTDATA8* input, VSHADEROUTPUTDATA8* output) {
966 /**
967 * TODO: use the NV_vertex_program (or 1_1) extension
968 * and specifics vendors (ARB_vertex_program??) variants for it
969 */
970 return TRUE;
971 }
972
973 HRESULT WINAPI IDirect3DVertexShaderImpl_ExecuteSW(IDirect3DVertexShaderImpl* vshader, VSHADERINPUTDATA8* input, VSHADEROUTPUTDATA8* output) {
974 /** Vertex Shader Temporary Registers */
975 D3DSHADERVECTOR R[12];
976 /*D3DSHADERSCALAR A0;*/
977 D3DSHADERVECTOR A[1];
978 /** temporary Vector for modifier management */
979 D3DSHADERVECTOR d;
980 D3DSHADERVECTOR s[3];
981 /** parser datas */
982 const DWORD* pToken = vshader->function;
983 const SHADER_OPCODE* curOpcode = NULL;
984 /** functions parameters */
985 D3DSHADERVECTOR* p[4];
986 D3DSHADERVECTOR* p_send[4];
987 DWORD i;
988
989 /** init temporary register */
990 memset(R, 0, 12 * sizeof(D3DSHADERVECTOR));
991
992 /* vshader_program_parse(vshader); */
993 #if 0 /* Must not be 1 in cvs */
994 TRACE("Input:\n");
995 TRACE_VSVECTOR(vshader->data->C[0]);
996 TRACE_VSVECTOR(vshader->data->C[1]);
997 TRACE_VSVECTOR(vshader->data->C[2]);
998 TRACE_VSVECTOR(vshader->data->C[3]);
999 TRACE_VSVECTOR(vshader->data->C[4]);
1000 TRACE_VSVECTOR(vshader->data->C[5]);
1001 TRACE_VSVECTOR(vshader->data->C[6]);
1002 TRACE_VSVECTOR(vshader->data->C[7]);
1003 TRACE_VSVECTOR(vshader->data->C[8]);
1004 TRACE_VSVECTOR(vshader->data->C[64]);
1005 TRACE_VSVECTOR(input->V[D3DVSDE_POSITION]);
1006 TRACE_VSVECTOR(input->V[D3DVSDE_BLENDWEIGHT]);
1007 TRACE_VSVECTOR(input->V[D3DVSDE_BLENDINDICES]);
1008 TRACE_VSVECTOR(input->V[D3DVSDE_NORMAL]);
1009 TRACE_VSVECTOR(input->V[D3DVSDE_PSIZE]);
1010 TRACE_VSVECTOR(input->V[D3DVSDE_DIFFUSE]);
1011 TRACE_VSVECTOR(input->V[D3DVSDE_SPECULAR]);
1012 TRACE_VSVECTOR(input->V[D3DVSDE_TEXCOORD0]);
1013 TRACE_VSVECTOR(input->V[D3DVSDE_TEXCOORD1]);
1014 #endif
1015
1016 TRACE_VSVECTOR(vshader->data->C[64]);
1017
1018 /* the first dword is the version tag */
1019 /* TODO: parse it */
1020
1021 if (vshader_is_version_token(*pToken)) { /** version */
1022 ++pToken;
1023 }
1024 while (D3DVS_END() != *pToken) {
1025 if (vshader_is_comment_token(*pToken)) { /** comment */
1026 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
1027 ++pToken;
1028 pToken += comment_len;
1029 continue ;
1030 }
1031 curOpcode = vshader_program_get_opcode(*pToken);
1032 ++pToken;
1033 if (NULL == curOpcode) {
1034 i = 0;
1035 /* unkown current opcode ... */
1036 while (*pToken & 0x80000000) {
1037 if (i == 0) {
1038 TRACE("unrecognized opcode: pos=%d token=%08lX\n", (pToken - 1) - vshader->function, *(pToken - 1));
1039 }
1040 TRACE("unrecognized opcode param: pos=%d token=%08lX what=", pToken - vshader->function, *pToken);
1041 vshader_program_add_param(*pToken, i, NULL); /* Add function just used for trace error scenario */
1042 TRACE("\n");
1043 ++i;
1044 ++pToken;
1045 }
1046 /*return FALSE;*/
1047 } else {
1048 if (curOpcode->num_params > 0) {
1049 /*TRACE(">> execting opcode: pos=%d opcode_name=%s token=%08lX\n", pToken - vshader->function, curOpcode->name, *pToken);*/
1050 for (i = 0; i < curOpcode->num_params; ++i) {
1051 DWORD reg = pToken[i] & 0x00001FFF;
1052 DWORD regtype = ((pToken[i] & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT);
1053
1054 switch (regtype << D3DSP_REGTYPE_SHIFT) {
1055 case D3DSPR_TEMP:
1056 /*TRACE("p[%d]=R[%d]\n", i, reg);*/
1057 p[i] = &R[reg];
1058 break;
1059 case D3DSPR_INPUT:
1060 /*TRACE("p[%d]=V[%s]\n", i, VertexShaderDeclRegister[reg]);*/
1061 p[i] = &input->V[reg];
1062 break;
1063 case D3DSPR_CONST:
1064 if (pToken[i] & D3DVS_ADDRMODE_RELATIVE) {
1065 p[i] = &vshader->data->C[(DWORD) A[0].x + reg];
1066 } else {
1067 p[i] = &vshader->data->C[reg];
1068 }
1069 break;
1070 case D3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
1071 if (0 != reg) {
1072 ERR("cannot handle address registers != a0, forcing use of a0\n");
1073 reg = 0;
1074 }
1075 /*TRACE("p[%d]=A[%d]\n", i, reg);*/
1076 p[i] = &A[reg];
1077 break;
1078 case D3DSPR_RASTOUT:
1079 switch (reg) {
1080 case D3DSRO_POSITION:
1081 p[i] = &output->oPos;
1082 break;
1083 case D3DSRO_FOG:
1084 p[i] = &output->oFog;
1085 break;
1086 case D3DSRO_POINT_SIZE:
1087 p[i] = &output->oPts;
1088 break;
1089 }
1090 break;
1091 case D3DSPR_ATTROUT:
1092 /*TRACE("p[%d]=oD[%d]\n", i, reg);*/
1093 p[i] = &output->oD[reg];
1094 break;
1095 case D3DSPR_TEXCRDOUT:
1096 /*TRACE("p[%d]=oT[%d]\n", i, reg);*/
1097 p[i] = &output->oT[reg];
1098 break;
1099 default:
1100 break;
1101 }
1102
1103 if (i > 0) { /* input reg */
1104 DWORD swizzle = (pToken[i] & D3DVS_SWIZZLE_MASK) >> D3DVS_SWIZZLE_SHIFT;
1105 UINT isNegative = ((pToken[i] & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG);
1106
1107 if (!isNegative && (D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) == swizzle) {
1108 /*TRACE("p[%d] not swizzled\n", i);*/
1109 p_send[i] = p[i];
1110 } else {
1111 DWORD swizzle_x = swizzle & 0x03;
1112 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1113 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1114 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1115 /*TRACE("p[%d] swizzled\n", i);*/
1116 float* tt = (float*) p[i];
1117 s[i].x = (isNegative) ? -tt[swizzle_x] : tt[swizzle_x];
1118 s[i].y = (isNegative) ? -tt[swizzle_y] : tt[swizzle_y];
1119 s[i].z = (isNegative) ? -tt[swizzle_z] : tt[swizzle_z];
1120 s[i].w = (isNegative) ? -tt[swizzle_w] : tt[swizzle_w];
1121 p_send[i] = &s[i];
1122 }
1123 } else { /* output reg */
1124 if ((pToken[i] & D3DSP_WRITEMASK_ALL) == D3DSP_WRITEMASK_ALL) {
1125 p_send[i] = p[i];
1126 } else {
1127 p_send[i] = &d; /* to be post-processed for modifiers management */
1128 }
1129 }
1130 }
1131 }
1132
1133 switch (curOpcode->num_params) {
1134 case 0:
1135 curOpcode->soft_fct();
1136 break;
1137 case 1:
1138 curOpcode->soft_fct(p_send[0]);
1139 break;
1140 case 2:
1141 curOpcode->soft_fct(p_send[0], p_send[1]);
1142 break;
1143 case 3:
1144 curOpcode->soft_fct(p_send[0], p_send[1], p_send[2]);
1145 break;
1146 case 4:
1147 curOpcode->soft_fct(p_send[0], p_send[1], p_send[2], p_send[3]);
1148 break;
1149 case 5:
1150 curOpcode->soft_fct(p_send[0], p_send[1], p_send[2], p_send[3], p_send[4]);
1151 break;
1152 default:
1153 ERR("%s too many params: %u\n", curOpcode->name, curOpcode->num_params);
1154 }
1155
1156 /* check if output reg modifier post-process */
1157 if (curOpcode->num_params > 0 && (pToken[0] & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
1158 if (pToken[0] & D3DSP_WRITEMASK_0) p[0]->x = d.x;
1159 if (pToken[0] & D3DSP_WRITEMASK_1) p[0]->y = d.y;
1160 if (pToken[0] & D3DSP_WRITEMASK_2) p[0]->z = d.z;
1161 if (pToken[0] & D3DSP_WRITEMASK_3) p[0]->w = d.w;
1162 }
1163
1164 #if 0
1165 TRACE_VSVECTOR(output->oPos);
1166 TRACE_VSVECTOR(output->oD[0]);
1167 TRACE_VSVECTOR(output->oD[1]);
1168 TRACE_VSVECTOR(output->oT[0]);
1169 TRACE_VSVECTOR(output->oT[1]);
1170 TRACE_VSVECTOR(R[0]);
1171 TRACE_VSVECTOR(R[1]);
1172 TRACE_VSVECTOR(R[2]);
1173 TRACE_VSVECTOR(R[3]);
1174 TRACE_VSVECTOR(R[4]);
1175 TRACE_VSVECTOR(R[5]);
1176 #endif
1177
1178 /* to next opcode token */
1179 pToken += curOpcode->num_params;
1180 }
1181 #if 0
1182 TRACE("End of current instruction:\n");
1183 TRACE_VSVECTOR(output->oPos);
1184 TRACE_VSVECTOR(output->oD[0]);
1185 TRACE_VSVECTOR(output->oD[1]);
1186 TRACE_VSVECTOR(output->oT[0]);
1187 TRACE_VSVECTOR(output->oT[1]);
1188 TRACE_VSVECTOR(R[0]);
1189 TRACE_VSVECTOR(R[1]);
1190 TRACE_VSVECTOR(R[2]);
1191 TRACE_VSVECTOR(R[3]);
1192 TRACE_VSVECTOR(R[4]);
1193 TRACE_VSVECTOR(R[5]);
1194 #endif
1195 }
1196 #if 0 /* Must not be 1 in cvs */
1197 TRACE("Output:\n");
1198 TRACE_VSVECTOR(output->oPos);
1199 TRACE_VSVECTOR(output->oD[0]);
1200 TRACE_VSVECTOR(output->oD[1]);
1201 TRACE_VSVECTOR(output->oT[0]);
1202 TRACE_VSVECTOR(output->oT[1]);
1203 #endif
1204 return D3D_OK;
1205 }
1206
1207 HRESULT WINAPI IDirect3DVertexShaderImpl_GetFunction(IDirect3DVertexShaderImpl* This, VOID* pData, UINT* pSizeOfData) {
1208 if (NULL == pData) {
1209 *pSizeOfData = This->functionLength;
1210 return D3D_OK;
1211 }
1212 if (*pSizeOfData < This->functionLength) {
1213 *pSizeOfData = This->functionLength;
1214 return D3DERR_MOREDATA;
1215 }
1216 if (NULL == This->function) { /* no function defined */
1217 TRACE("(%p) : GetFunction no User Function defined using NULL to %p\n", This, pData);
1218 (*(DWORD **) pData) = NULL;
1219 } else {
1220 TRACE("(%p) : GetFunction copying to %p\n", This, pData);
1221 memcpy(pData, This->function, This->functionLength);
1222 }
1223 return D3D_OK;
1224 }
1225
1226 HRESULT WINAPI IDirect3DVertexShaderImpl_SetConstantF(IDirect3DVertexShaderImpl* This, UINT StartRegister, CONST FLOAT* pConstantData, UINT Vector4fCount) {
1227 if (StartRegister + Vector4fCount > D3D8_VSHADER_MAX_CONSTANTS) {
1228 return D3DERR_INVALIDCALL;
1229 }
1230 if (NULL == This->data) { /* temporary while datas not supported */
1231 FIXME("(%p) : VertexShader_SetConstant not fully supported yet\n", This);
1232 return D3DERR_INVALIDCALL;
1233 }
1234 memcpy(&This->data->C[StartRegister], pConstantData, Vector4fCount * 4 * sizeof(FLOAT));
1235 return D3D_OK;
1236 }
1237
1238 HRESULT WINAPI IDirect3DVertexShaderImpl_GetConstantF(IDirect3DVertexShaderImpl* This, UINT StartRegister, FLOAT* pConstantData, UINT Vector4fCount) {
1239 if (StartRegister + Vector4fCount > D3D8_VSHADER_MAX_CONSTANTS) {
1240 return D3DERR_INVALIDCALL;
1241 }
1242 if (NULL == This->data) { /* temporary while datas not supported */
1243 return D3DERR_INVALIDCALL;
1244 }
1245 memcpy(pConstantData, &This->data->C[StartRegister], Vector4fCount * 4 * sizeof(FLOAT));
1246 return D3D_OK;
1247 }
1248
1249
1250 /**********************************************************************************************************************************************
1251 **********************************************************************************************************************************************
1252 **********************************************************************************************************************************************
1253 **********************************************************************************************************************************************
1254 **********************************************************************************************************************************************/
1255
1256 void pshader_texcoord(D3DSHADERVECTOR* d) {
1257 }
1258
1259 void pshader_texkill(D3DSHADERVECTOR* d) {
1260 }
1261
1262 void pshader_tex(D3DSHADERVECTOR* d) {
1263 }
1264
1265 void pshader_texbem(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1266 }
1267
1268 void pshader_texbeml(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1269 }
1270
1271 void pshader_texreg2ar(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1272 }
1273
1274 void pshader_texreg2gb(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1275 }
1276
1277 void pshader_texm3x2pad(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1278 }
1279
1280 void pshader_texm3x2tex(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1281 }
1282
1283 void pshader_texm3x3pad(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1284 }
1285
1286 void pshader_texm3x3tex(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1287 }
1288
1289 void pshader_texm3x3diff(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1290 }
1291
1292 void pshader_texm3x3spec(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
1293 }
1294
1295 void pshader_texm3x3vspec(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1296 }
1297
1298 void pshader_cnd(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1, D3DSHADERVECTOR* s2) {
1299 }
1300
1301 void pshader_def(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1, D3DSHADERVECTOR* s2, D3DSHADERVECTOR* s3) {
1302 }
1303
1304 void pshader_texreg2rgb(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1305 }
1306
1307 void pshader_texdp3tex(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1308 }
1309
1310 void pshader_texm3x2depth(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1311 }
1312
1313 void pshader_texdp3(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1314 }
1315
1316 void pshader_texm3x3(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0) {
1317 }
1318
1319 void pshader_texdepth(D3DSHADERVECTOR* d) {
1320 }
1321
1322 void pshader_cmp(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1, D3DSHADERVECTOR* s2) {
1323 }
1324
1325 void pshader_bem(D3DSHADERVECTOR* d, D3DSHADERVECTOR* s0, D3DSHADERVECTOR* s1) {
1326 }
1327
1328 static CONST SHADER_OPCODE pshader_ins [] = {
1329 {D3DSIO_NOP, "nop", 0, vshader_nop, 0, 0},
1330 {D3DSIO_MOV, "mov", 2, vshader_mov, 0, 0},
1331 {D3DSIO_ADD, "add", 3, vshader_add, 0, 0},
1332 {D3DSIO_SUB, "sub", 3, vshader_sub, 0, 0},
1333 {D3DSIO_MAD, "mad", 4, vshader_mad, 0, 0},
1334 {D3DSIO_MUL, "mul", 3, vshader_mul, 0, 0},
1335 {D3DSIO_RCP, "rcp", 2, vshader_rcp, 0, 0},
1336 {D3DSIO_RSQ, "rsq", 2, vshader_rsq, 0, 0},
1337 {D3DSIO_DP3, "dp3", 3, vshader_dp3, 0, 0},
1338 {D3DSIO_DP4, "dp4", 3, vshader_dp4, 0, 0},
1339 {D3DSIO_MIN, "min", 3, vshader_min, 0, 0},
1340 {D3DSIO_MAX, "max", 3, vshader_max, 0, 0},
1341 {D3DSIO_SLT, "slt", 3, vshader_slt, 0, 0},
1342 {D3DSIO_SGE, "sge", 3, vshader_sge, 0, 0},
1343 {D3DSIO_EXP, "exp", 2, vshader_exp, 0, 0},
1344 {D3DSIO_LOG, "log", 2, vshader_log, 0, 0},
1345 {D3DSIO_LIT, "lit", 2, vshader_lit, 0, 0},
1346 {D3DSIO_DST, "dst", 3, vshader_dst, 0, 0},
1347 {D3DSIO_LRP, "lrp", 5, vshader_lrp, 0, 0},
1348 {D3DSIO_FRC, "frc", 2, vshader_frc, 0, 0},
1349 {D3DSIO_M4x4, "m4x4", 3, vshader_m4x4, 0, 0},
1350 {D3DSIO_M4x3, "m4x3", 3, vshader_m4x3, 0, 0},
1351 {D3DSIO_M3x4, "m3x4", 3, vshader_m3x4, 0, 0},
1352 {D3DSIO_M3x3, "m3x3", 3, vshader_m3x3, 0, 0},
1353 {D3DSIO_M3x2, "m3x2", 3, vshader_m3x2, 0, 0},
1354
1355 {D3DSIO_TEXCOORD, "texcoord", 1, pshader_texcoord, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1356 {D3DSIO_TEXKILL, "texkill", 1, pshader_texkill, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1357 {D3DSIO_TEX, "tex", 1, pshader_tex, D3DPS_VERSION(1,1), D3DPS_VERSION(1,4)},
1358 {D3DSIO_TEXBEM, "texbem", 2, pshader_texbem, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1359 {D3DSIO_TEXBEML, "texbeml", 2, pshader_texbeml, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1360 {D3DSIO_TEXREG2AR, "texreg2ar", 2, pshader_texreg2ar, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1361 {D3DSIO_TEXREG2GB, "texreg2gb", 2, pshader_texreg2gb, D3DPS_VERSION(1,2), D3DPS_VERSION(1,3)},
1362 {D3DSIO_TEXM3x2PAD, "texm3x2pad", 2, pshader_texm3x2pad, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1363 {D3DSIO_TEXM3x2TEX, "texm3x2tex", 2, pshader_texm3x2tex, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1364 {D3DSIO_TEXM3x3PAD, "texm3x3pad", 2, pshader_texm3x3pad, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1365 {D3DSIO_TEXM3x3TEX, "texm3x3tex", 2, pshader_texm3x3tex, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1366 {D3DSIO_TEXM3x3DIFF, "texm3x3diff", 2, pshader_texm3x3diff, D3DPS_VERSION(0,0), D3DPS_VERSION(0,0)},
1367 {D3DSIO_TEXM3x3SPEC, "texm3x3spec", 3, pshader_texm3x3spec, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1368 {D3DSIO_TEXM3x3VSPEC, "texm3x3vspec", 2, pshader_texm3x3vspec, D3DPS_VERSION(1,1), D3DPS_VERSION(1,3)},
1369
1370 {D3DSIO_EXPP, "expp", 2, vshader_expp, 0, 0},
1371 {D3DSIO_LOGP, "logp", 2, vshader_logp, 0, 0},
1372
1373 {D3DSIO_CND, "cnd", 4, pshader_cnd, D3DPS_VERSION(1,1), D3DPS_VERSION(1,4)},
1374 {D3DSIO_DEF, "def", 5, pshader_def, D3DPS_VERSION(1,1), D3DPS_VERSION(3,0)},
1375 {D3DSIO_TEXREG2RGB, "texbreg2rgb", 2, pshader_texreg2rgb, D3DPS_VERSION(1,2), D3DPS_VERSION(1,3)},
1376
1377 {D3DSIO_TEXDP3TEX, "texdp3tex", 2, pshader_texdp3tex, D3DPS_VERSION(1,2), D3DPS_VERSION(1,3)},
1378 {D3DSIO_TEXM3x2DEPTH, "texm3x2depth", 2, pshader_texm3x2depth, D3DPS_VERSION(1,3), D3DPS_VERSION(1,3)},
1379 {D3DSIO_TEXDP3, "texdp3", 2, pshader_texdp3, D3DPS_VERSION(1,2), D3DPS_VERSION(1,3)},
1380 {D3DSIO_TEXM3x3, "texm3x3", 2, pshader_texm3x3, D3DPS_VERSION(1,2), D3DPS_VERSION(1,3)},
1381 {D3DSIO_TEXDEPTH, "texdepth", 1, pshader_texdepth, D3DPS_VERSION(1,4), D3DPS_VERSION(1,4)},
1382 {D3DSIO_CMP, "cmp", 4, pshader_cmp, D3DPS_VERSION(1,1), D3DPS_VERSION(3,0)},
1383 {D3DSIO_BEM, "bem", 3, pshader_bem, D3DPS_VERSION(1,4), D3DPS_VERSION(1,4)},
1384
1385 {0, NULL, 0, NULL}
1386 };
1387
1388 inline static const SHADER_OPCODE* pshader_program_get_opcode(const DWORD code) {
1389 DWORD i = 0;
1390 /** TODO: use dichotomic search */
1391 while (NULL != pshader_ins[i].name) {
1392 if ((code & D3DSI_OPCODE_MASK) == pshader_ins[i].opcode) {
1393 return &pshader_ins[i];
1394 }
1395 ++i;
1396 }
1397 return NULL;
1398 }
1399
1400 inline static BOOL pshader_is_version_token(DWORD token) {
1401 return 0xFFFF0000 == (token & 0xFFFF0000);
1402 }
1403
1404 inline static BOOL pshader_is_comment_token(DWORD token) {
1405 return D3DSIO_COMMENT == (token & D3DSI_OPCODE_MASK);
1406 }
1407
1408 inline static void pshader_program_dump_opcode(const SHADER_OPCODE* curOpcode, const DWORD code, const DWORD output) {
1409 if (0 != (code & ~D3DSI_OPCODE_MASK)) {
1410 DWORD mask = (code & ~D3DSI_OPCODE_MASK);
1411 switch (mask) {
1412 case 0x40000000: TRACE("+"); break;
1413 default:
1414 TRACE(" unhandled modifier(0x%08lx) ", mask);
1415 }
1416 }
1417 TRACE("%s", curOpcode->name);
1418 /**
1419 * normally this is a destination reg modifier
1420 * but in pixel shaders asm code its specified as:
1421 * dp3_x4 t1.rgba, r1, c1
1422 * or
1423 * dp3_x2_sat r0, t0_bx2, v0_bx2
1424 * so for better debbuging i use the same norm
1425 */
1426 if (0 != (output & D3DSP_DSTSHIFT_MASK)) {
1427 DWORD shift = (output & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
1428 if (shift < 8) {
1429 TRACE("_x%u", 1 << shift);
1430 } else {
1431 TRACE("_d%u", 1 << (16-shift));
1432 }
1433 }
1434 if (0 != (output & D3DSP_DSTMOD_MASK)) {
1435 DWORD mask = output & D3DSP_DSTMOD_MASK;
1436 switch (mask) {
1437 case D3DSPDM_SATURATE: TRACE("_sat"); break;
1438 default:
1439 TRACE("_unhandled_modifier(0x%08lx)", mask);
1440 }
1441 }
1442 TRACE(" ");
1443 }
1444
1445 inline static void pshader_program_dump_param(const DWORD param, int input) {
1446 static const char* rastout_reg_names[] = { "oC0", "oC1", "oC2", "oC3", "oDepth" };
1447 static const char swizzle_reg_chars[] = "rgba";
1448
1449 DWORD reg = param & 0x00001FFF;
1450 DWORD regtype = ((param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT);
1451
1452 if (input) {
1453 if ( ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) ||
1454 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_BIASNEG) ||
1455 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_SIGNNEG) ||
1456 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_X2NEG) )
1457 TRACE("-");
1458 else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_COMP)
1459 TRACE("1-");
1460 }
1461
1462 switch (regtype << D3DSP_REGTYPE_SHIFT) {
1463 case D3DSPR_TEMP:
1464 TRACE("R[%lu]", reg);
1465 break;
1466 case D3DSPR_INPUT:
1467 TRACE("V[%lu]", reg);
1468 break;
1469 case D3DSPR_CONST:
1470 TRACE("C[%s%lu]", (param & D3DVS_ADDRMODE_RELATIVE) ? "a0.x + " : "", reg);
1471 break;
1472 case D3DSPR_TEXTURE: /* case D3DSPR_ADDR: */
1473 TRACE("t[%lu]", reg);
1474 break;
1475 case D3DSPR_RASTOUT:
1476 TRACE("%s", rastout_reg_names[reg]);
1477 break;
1478 case D3DSPR_ATTROUT:
1479 TRACE("oD[%lu]", reg);
1480 break;
1481 case D3DSPR_TEXCRDOUT:
1482 TRACE("oT[%lu]", reg);
1483 break;
1484 default:
1485 break;
1486 }
1487
1488 if (!input) {
1489 /** operand output */
1490 /**
1491 * for better debugging traces it's done into opcode dump code
1492 * @see pshader_program_dump_opcode
1493 if (0 != (param & D3DSP_DSTMOD_MASK)) {
1494 DWORD mask = param & D3DSP_DSTMOD_MASK;
1495 switch (mask) {
1496 case D3DSPDM_SATURATE: TRACE("_sat"); break;
1497 default:
1498 TRACE("_unhandled_modifier(0x%08lx)", mask);
1499 }
1500 }
1501 if (0 != (param & D3DSP_DSTSHIFT_MASK)) {
1502 DWORD shift = (param & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
1503 if (shift > 0) {
1504 TRACE("_x%u", 1 << shift);
1505 }
1506 }
1507 */
1508 if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
1509 if (param & D3DSP_WRITEMASK_0) TRACE(".r");
1510 if (param & D3DSP_WRITEMASK_1) TRACE(".g");
1511 if (param & D3DSP_WRITEMASK_2) TRACE(".b");
1512 if (param & D3DSP_WRITEMASK_3) TRACE(".a");
1513 }
1514 } else {
1515 /** operand input */
1516 DWORD swizzle = (param & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
1517 DWORD swizzle_x = swizzle & 0x03;
1518 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1519 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1520 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1521 /**
1522 * swizzle bits fields:
1523 * WWZZYYXX
1524 */
1525 if ((D3DSP_NOSWIZZLE >> D3DSP_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
1526 if (swizzle_x == swizzle_y &&
1527 swizzle_x == swizzle_z &&
1528 swizzle_x == swizzle_w) {
1529 TRACE(".%c", swizzle_reg_chars[swizzle_x]);
1530 } else {
1531 TRACE(".%c%c%c%c",
1532 swizzle_reg_chars[swizzle_x],
1533 swizzle_reg_chars[swizzle_y],
1534 swizzle_reg_chars[swizzle_z],
1535 swizzle_reg_chars[swizzle_w]);
1536 }
1537 }
1538 if (0 != (param & D3DSP_SRCMOD_MASK)) {
1539 DWORD mask = param & D3DSP_SRCMOD_MASK;
1540 /*TRACE("_modifier(0x%08lx) ", mask);*/
1541 switch (mask) {
1542 case D3DSPSM_NONE: break;
1543 case D3DSPSM_NEG: break;
1544 case D3DSPSM_BIAS: TRACE("_bias"); break;
1545 case D3DSPSM_BIASNEG: TRACE("_bias"); break;
1546 case D3DSPSM_SIGN: TRACE("_bx2"); break;
1547 case D3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1548 case D3DSPSM_COMP: break;
1549 case D3DSPSM_X2: TRACE("_x2"); break;
1550 case D3DSPSM_X2NEG: TRACE("_x2"); break;
1551 case D3DSPSM_DZ: TRACE("_dz"); break;
1552 case D3DSPSM_DW: TRACE("_dw"); break;
1553 default:
1554 TRACE("_unknown(0x%08lx)", mask);
1555 }
1556 }
1557 }
1558 }
1559
1560 static int constants[D3D8_PSHADER_MAX_CONSTANTS];
1561
1562 inline static void get_register_name(const DWORD param, char* regstr)
1563 {
1564 static const char* rastout_reg_names[] = { "oC0", "oC1", "oC2", "oC3", "oDepth" };
1565
1566 DWORD reg = param & 0x00001FFF;
1567 DWORD regtype = ((param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT);
1568
1569 switch (regtype << D3DSP_REGTYPE_SHIFT) {
1570 case D3DSPR_TEMP:
1571 sprintf(regstr, "R%lu", reg);
1572 break;
1573 case D3DSPR_INPUT:
1574 if (reg==0) {
1575 strcpy(regstr, "fragment.color.primary");
1576 } else {
1577 strcpy(regstr, "fragment.color.secondary");
1578 }
1579 break;
1580 case D3DSPR_CONST:
1581 if (constants[reg])
1582 sprintf(regstr, "C%lu", reg);
1583 else
1584 sprintf(regstr, "program.env[%lu]", reg);
1585 break;
1586 case D3DSPR_TEXTURE: /* case D3DSPR_ADDR: */
1587 sprintf(regstr,"T%lu", reg);
1588 break;
1589 case D3DSPR_RASTOUT:
1590 sprintf(regstr, "%s", rastout_reg_names[reg]);
1591 break;
1592 case D3DSPR_ATTROUT:
1593 sprintf(regstr, "oD[%lu]", reg);
1594 break;
1595 case D3DSPR_TEXCRDOUT:
1596 sprintf(regstr, "oT[%lu]", reg);
1597 break;
1598 default:
1599 break;
1600 }
1601 }
1602
1603 inline static void addline(int* lineNum, char* pgm, char* line)
1604 {
1605 ++(*lineNum);
1606 TRACE_(d3d_hw_shader)("GL HW (%u, %u) : %s\n", *lineNum, strlen(pgm), line);
1607 strcat(pgm, line);
1608 strcat(pgm, "\n");
1609 }
1610
1611 char* shift_tab[] = {
1612 "dummy", /* 0 (none) */
1613 "coefmul.x", /* 1 (x2) */
1614 "coefmul.y", /* 2 (x4) */
1615 "coefmul.z", /* 3 (x8) */
1616 "coefmul.w", /* 4 (x16) */
1617 "dummy", /* 5 (x32) */
1618 "dummy", /* 6 (x64) */
1619 "dummy", /* 7 (x128) */
1620 "dummy", /* 8 (d256) */
1621 "dummy", /* 9 (d128) */
1622 "dummy", /* 10 (d64) */
1623 "dummy", /* 11 (d32) */
1624 "coefdiv.w", /* 12 (d16) */
1625 "coefdiv.z", /* 13 (d8) */
1626 "coefdiv.y", /* 14 (d4) */
1627 "coefdiv.x" /* 15 (d2) */
1628 };
1629
1630 inline static void get_write_mask(const DWORD output_reg, char* write_mask)
1631 {
1632 *write_mask = 0;
1633 if ((output_reg & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
1634 if (output_reg & D3DSP_WRITEMASK_0) strcat(write_mask, ".r");
1635 if (output_reg & D3DSP_WRITEMASK_1) strcat(write_mask, ".g");
1636 if (output_reg & D3DSP_WRITEMASK_2) strcat(write_mask, ".b");
1637 if (output_reg & D3DSP_WRITEMASK_3) strcat(write_mask, ".a");
1638 }
1639 }
1640
1641 inline static void get_input_register_swizzle(const DWORD instr, char* swzstring)
1642 {
1643 static const char swizzle_reg_chars[] = "rgba";
1644 DWORD swizzle = (instr & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
1645 DWORD swizzle_x = swizzle & 0x03;
1646 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1647 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1648 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1649 /**
1650 * swizzle bits fields:
1651 * WWZZYYXX
1652 */
1653 *swzstring = 0;
1654 if ((D3DSP_NOSWIZZLE >> D3DSP_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
1655 if (swizzle_x == swizzle_y &&
1656 swizzle_x == swizzle_z &&
1657 swizzle_x == swizzle_w) {
1658 sprintf(swzstring, ".%c", swizzle_reg_chars[swizzle_x]);
1659 } else {
1660 sprintf(swzstring, ".%c%c%c%c",
1661 swizzle_reg_chars[swizzle_x],
1662 swizzle_reg_chars[swizzle_y],
1663 swizzle_reg_chars[swizzle_z],
1664 swizzle_reg_chars[swizzle_w]);
1665 }
1666 }
1667 }
1668
1669 inline static void gen_output_modifier_line(int saturate, char* write_mask, int shift, char *regstr, char* line)
1670 {
1671 /* Generate a line that does the output modifier computation */
1672 sprintf(line, "MUL%s %s%s, %s, %s;", saturate ? "_SAT" : "", regstr, write_mask, regstr, shift_tab[shift]);
1673 }
1674
1675 inline static int gen_input_modifier_line(const DWORD instr, int tmpreg, char* outregstr, char* line)
1676 {
1677 /* Generate a line that does the input modifier computation and return the input register to use */
1678 static char regstr[256];
1679 int insert_line;
1680
1681 /* Assume a new line will be added */
1682 insert_line = 1;
1683
1684 /* Get register name */
1685 get_register_name(instr, regstr);
1686
1687 switch (instr & D3DSP_SRCMOD_MASK) {
1688 case D3DSPSM_NONE:
1689 strcpy(outregstr, regstr);
1690 insert_line = 0;
1691 break;
1692 case D3DSPSM_NEG:
1693 sprintf(outregstr, "-%s", regstr);
1694 insert_line = 0;
1695 break;
1696 case D3DSPSM_BIAS:
1697 sprintf(line, "ADD T%c, %s, -coefdiv.x;", 'A' + tmpreg, regstr);
1698 break;
1699 case D3DSPSM_BIASNEG:
1700 sprintf(line, "ADD T%c, -%s, coefdiv.x;", 'A' + tmpreg, regstr);
1701 break;
1702 case D3DSPSM_SIGN:
1703 sprintf(line, "MAD T%c, %s, coefmul.x, -one.x;", 'A' + tmpreg, regstr);
1704 break;
1705 case D3DSPSM_SIGNNEG:
1706 sprintf(line, "MAD T%c, %s, -coefmul.x, one.x;", 'A' + tmpreg, regstr);
1707 break;
1708 case D3DSPSM_COMP:
1709 sprintf(line, "SUB T%c, one.x, %s;", 'A' + tmpreg, regstr);
1710 break;
1711 case D3DSPSM_X2:
1712 sprintf(line, "ADD T%c, %s, %s", 'A' + tmpreg, regstr, regstr);
1713 break;
1714 case D3DSPSM_X2NEG:
1715 sprintf(line, "ADD T%c, %s, %s", 'A' + tmpreg, regstr, regstr);
1716 break;
1717 case D3DSPSM_DZ:
1718 case D3DSPSM_DW:
1719 default:
1720 strcpy(outregstr, regstr);
1721 insert_line = 0;
1722 }
1723
1724 if (insert_line) {
1725 /* Substitute the register name */
1726 sprintf(outregstr, "T%c", 'A' + tmpreg);
1727 }
1728
1729 return insert_line;
1730 }
1731
1732 /**
1733 * Pixel Shaders
1734 *
1735 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/Shaders/PixelShader1_X/modifiers/sourceregistermodifiers.asp
1736 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/Shaders/PixelShader2_0/Registers/Registers.asp
1737 * http://msdn.microsoft.com/library/default.asp?url=/library/en-us/directx9_c/directx/graphics/reference/d3d/interfaces/IDirect3DPixelShader9/_IDirect3DPixelShader9.asp
1738 *
1739 */
1740 inline static VOID IDirect3DPixelShaderImpl_GenerateProgramArbHW(IDirect3DPixelShaderImpl* pshader, CONST DWORD* pFunction) {
1741 const DWORD* pToken = pFunction;
1742 const SHADER_OPCODE* curOpcode = NULL;
1743 const DWORD* pInstr;
1744 DWORD code;
1745 DWORD i;
1746 int autoparam;
1747 unsigned lineNum = 0;
1748 char *pgmStr = NULL;
1749 char tmpLine[255];
1750 BOOL saturate;
1751 IDirect3DDevice8Impl* This = pshader->device;
1752
1753 for(i = 0; i < D3D8_PSHADER_MAX_CONSTANTS; i++)
1754 constants[i] = 0;
1755
1756 pgmStr = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, 65535); /* 64kb should be enough */
1757
1758 if (NULL != pToken) {
1759 while (D3DPS_END() != *pToken) {
1760 if (pshader_is_version_token(*pToken)) { /** version */
1761
1762 /* Extract version *10 into integer value (ie. 1.0 == 10, 1.1==11 etc */
1763 int version = (((*pToken >> 8) & 0x0F) * 10) + (*pToken & 0x0F);
1764 int numTemps;
1765 int numConstants;
1766
1767 TRACE_(d3d_hw_shader)("ps.%lu.%lu;\n", (*pToken >> 8) & 0x0F, (*pToken & 0x0F));
1768
1769 /* Each release of pixel shaders has had different numbers of temp registers */
1770 switch (version) {
1771 case 10: numTemps=12;
1772 numConstants=8;
1773 strcpy(tmpLine, "!!ARBfp1.0");
1774 break;
1775 case 11: numTemps=12;
1776 numConstants=8;
1777 strcpy(tmpLine, "!!ARBfp1.0");
1778 break;
1779 case 20: numTemps=12;
1780 numConstants=8;
1781 strcpy(tmpLine, "!!ARBfp2.0");
1782 FIXME_(d3d_hw_shader)("No work done yet to support ps2.0 in hw\n");
1783 break;
1784 case 30: numTemps=32;
1785 numConstants=8;
1786 strcpy(tmpLine, "!!ARBfp3.0");
1787 FIXME_(d3d_hw_shader)("No work done yet to support ps3.0 in hw\n");
1788 break;
1789 default:
1790 numTemps=12;
1791 numConstants=8;
1792 strcpy(tmpLine, "!!ARBfp1.0");
1793 FIXME_(d3d_hw_shader)("Unrecognized pixel shader version!\n");
1794 }
1795 addline(&lineNum, pgmStr, tmpLine);
1796
1797 for(i = 0; i < 4; i++) {
1798 sprintf(tmpLine, "TEMP T%lu;", i);
1799 addline(&lineNum, pgmStr, tmpLine);
1800 }
1801 for(i = 0; i < 2; i++) {
1802 sprintf(tmpLine, "TEMP R%lu;", i);
1803 addline(&lineNum, pgmStr, tmpLine);
1804 }
1805
1806 sprintf(tmpLine, "TEMP TMP;");
1807 addline(&lineNum, pgmStr, tmpLine);
1808 sprintf(tmpLine, "TEMP TA;");
1809 addline(&lineNum, pgmStr, tmpLine);
1810 sprintf(tmpLine, "TEMP TB;");
1811 addline(&lineNum, pgmStr, tmpLine);
1812 sprintf(tmpLine, "TEMP TC;");
1813 addline(&lineNum, pgmStr, tmpLine);
1814
1815 strcpy(tmpLine, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };");
1816 addline(&lineNum, pgmStr, tmpLine);
1817 strcpy(tmpLine, "PARAM coefmul = { 2, 4, 8, 16 };");
1818 addline(&lineNum, pgmStr, tmpLine);
1819 strcpy(tmpLine, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };");
1820 addline(&lineNum, pgmStr, tmpLine);
1821
1822 for(i = 0; i < 4; i++) {
1823 sprintf(tmpLine, "MOV T%lu, fragment.texcoord[%lu];", i, i);
1824 addline(&lineNum, pgmStr, tmpLine);
1825 }
1826
1827 ++pToken;
1828 continue;
1829 }
1830 if (pshader_is_comment_token(*pToken)) { /** comment */
1831 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
1832 ++pToken;
1833 /*TRACE("comment[%ld] ;%s\n", comment_len, (char*)pToken);*/
1834 pToken += comment_len;
1835 continue;
1836 }
1837 code = *pToken;
1838 pInstr = pToken;
1839 curOpcode = pshader_program_get_opcode(code);
1840 ++pToken;
1841 if (NULL == curOpcode) {
1842 /* unkown current opcode ... */
1843 while (*pToken & 0x80000000) {
1844 TRACE("unrecognized opcode: %08lx\n", *pToken);
1845 ++pToken;
1846 }
1847 } else {
1848 autoparam = 1;
1849 saturate = FALSE;
1850 /* Build opcode for GL vertex_program */
1851 switch (curOpcode->opcode) {
1852 case D3DSIO_DEF:
1853 {
1854 DWORD reg = *pToken & 0x00001FFF;
1855 sprintf(tmpLine, "PARAM C%lu = { %f, %f, %f, %f };", reg,
1856 *((float*)(pToken+1)),
1857 *((float*)(pToken+2)),
1858 *((float*)(pToken+3)),
1859 *((float*)(pToken+4)) );
1860 addline(&lineNum, pgmStr, tmpLine);
1861 constants[reg] = 1;
1862 autoparam = 0;
1863 pToken+=5;
1864 }
1865 break;
1866 case D3DSIO_TEXKILL:
1867 strcpy(tmpLine, "KIL");
1868 break;
1869 case D3DSIO_TEX:
1870 {
1871 DWORD reg = *pToken & 0x00001FFF;
1872 sprintf(tmpLine,"TEX T%lu, T%lu, texture[%lu], 2D;", reg, reg, reg);
1873 addline(&lineNum, pgmStr, tmpLine);
1874 autoparam = 0;
1875 pToken++;
1876 }
1877 break;
1878 case D3DSIO_TEXCOORD:
1879 {
1880 DWORD reg = *pToken & 0x00001FFF;
1881 sprintf(tmpLine, "MOV T%lu, fragment.texcoord[%lu];", reg, reg);
1882 addline(&lineNum, pgmStr, tmpLine);
1883 autoparam = 0;
1884 pToken++;
1885 }
1886 break;
1887 case D3DSIO_TEXM3x2PAD:
1888 {
1889 DWORD reg = *pToken & 0x00001FFF;
1890 char buf[50];
1891 if (gen_input_modifier_line(*(pToken+1), 0, buf, tmpLine))
1892 addline(&lineNum, pgmStr, tmpLine);
1893 sprintf(tmpLine, "DP3 TMP.x, T%lu, %s;", reg, buf);
1894 addline(&lineNum, pgmStr, tmpLine);
1895 autoparam = 0;
1896 pToken += 2;
1897 }
1898 break;
1899 case D3DSIO_TEXM3x2TEX:
1900 {
1901 DWORD reg = *pToken & 0x00001FFF;
1902 char buf[50];
1903 if (gen_input_modifier_line(*(pToken+1), 0, buf, tmpLine))
1904 addline(&lineNum, pgmStr, tmpLine);
1905 sprintf(tmpLine, "DP3 TMP.y, T%lu, %s;", reg, buf);
1906 addline(&lineNum, pgmStr, tmpLine);
1907 sprintf(tmpLine, "TEX T%lu, TMP, texture[%lu], 2D;", reg, reg);
1908 addline(&lineNum, pgmStr, tmpLine);
1909 autoparam = 0;
1910 pToken += 2;
1911 }
1912 break;
1913 case D3DSIO_TEXREG2AR:
1914 {
1915 DWORD reg1 = *pToken & 0x00001FFF;
1916 DWORD reg2 = *(pToken+1) & 0x00001FFF;
1917 sprintf(tmpLine, "MOV TMP.r, T%lu.a;", reg2);
1918 addline(&lineNum, pgmStr, tmpLine);
1919 sprintf(tmpLine, "MOV TMP.g, T%lu.r;", reg2);
1920 addline(&lineNum, pgmStr, tmpLine);
1921 sprintf(tmpLine, "TEX T%lu, TMP, texture[%lu], 2D;", reg1, reg1);
1922 addline(&lineNum, pgmStr, tmpLine);
1923 autoparam = 0;
1924 pToken+=2;
1925 }
1926 break;
1927 case D3DSIO_TEXREG2GB:
1928 {
1929 DWORD reg1 = *pToken & 0x00001FFF;
1930 DWORD reg2 = *(pToken+1) & 0x00001FFF;
1931 sprintf(tmpLine, "MOV TMP.r, T%lu.g;", reg2);
1932 addline(&lineNum, pgmStr, tmpLine);
1933 sprintf(tmpLine, "MOV TMP.g, T%lu.b;", reg2);
1934 addline(&lineNum, pgmStr, tmpLine);
1935 sprintf(tmpLine, "TEX T%lu, TMP, texture[%lu], 2D;", reg1, reg1);
1936 addline(&lineNum, pgmStr, tmpLine);
1937 autoparam = 0;
1938 pToken+=2;
1939 }
1940 break;
1941 case D3DSIO_MOV:
1942 strcpy(tmpLine, "MOV");
1943 break;
1944 case D3DSIO_MUL:
1945 strcpy(tmpLine, "MUL");
1946 break;
1947 case D3DSIO_DP3:
1948 strcpy(tmpLine, "DP3");
1949 break;
1950 case D3DSIO_MAD:
1951 strcpy(tmpLine, "MAD");
1952 break;
1953 case D3DSIO_ADD:
1954 strcpy(tmpLine, "ADD");
1955 break;
1956 case D3DSIO_SUB:
1957 strcpy(tmpLine, "SUB");
1958 break;
1959 default:
1960 FIXME_(d3d_hw_shader)("Can't handle opcode %s in hwShader\n", curOpcode->name);
1961 }
1962 if (0 != (*pToken & D3DSP_DSTMOD_MASK)) {
1963 DWORD mask = *pToken & D3DSP_DSTMOD_MASK;
1964 switch (mask) {
1965 case D3DSPDM_SATURATE: saturate = TRUE; break;
1966 default:
1967 TRACE("_unhandled_modifier(0x%08lx)", mask);
1968 }
1969 }
1970 if (autoparam && (curOpcode->num_params > 0)) {
1971 char regs[3][50];
1972 char tmp[256];
1973 char swzstring[20];
1974 /* Generate lines that handle input modifier computation */
1975 for (i = 1; i < curOpcode->num_params; i++) {
1976 if (gen_input_modifier_line(*(pToken+i), i-1, regs[i-1], tmp))
1977 addline(&lineNum, pgmStr, tmp);
1978 }
1979 /* Handle saturation only when no shift is present in the output modifier */
1980 if ((*pToken & D3DSPDM_SATURATE) && (0 == (*pToken & D3DSP_DSTSHIFT_MASK)))
1981 strcat(tmpLine,"_SAT");
1982 strcat(tmpLine, " ");
1983 /* Handle output register */
1984 get_register_name(*pToken, tmp);
1985 strcat(tmpLine, tmp);
1986 get_write_mask(*pToken, tmp);
1987 strcat(tmpLine, tmp);
1988 /* Handle input registers */
1989 for (i = 1; i < curOpcode->num_params; i++) {
1990 strcat(tmpLine, ", ");
1991 strcat(tmpLine, regs[i-1]);
1992 get_input_register_swizzle(*(pToken+i), swzstring);
1993 strcat(tmpLine, swzstring);
1994 }
1995 strcat(tmpLine,";");
1996 addline(&lineNum, pgmStr, tmpLine);
1997 pToken += curOpcode->num_params;
1998 }
1999 if (curOpcode->num_params > 0) {
2000 DWORD param = *(pInstr+1);
2001 if (0 != (param & D3DSP_DSTSHIFT_MASK)) {
2002 /* Generate a line that handle the output modifier computation */
2003 char regstr[100];
2004 char write_mask[20];
2005 DWORD shift = (param & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
2006 get_register_name(param, regstr);
2007 get_write_mask(param, write_mask);
2008 gen_output_modifier_line(saturate, write_mask, shift, regstr, tmpLine);
2009 addline(&lineNum, pgmStr, tmpLine);
2010 }
2011 }
2012 }
2013 }
2014 strcpy(tmpLine, "MOV result.color, R0;");
2015 addline(&lineNum, pgmStr, tmpLine);
2016
2017 strcpy(tmpLine, "END");
2018 addline(&lineNum, pgmStr, tmpLine);
2019 }
2020
2021 /* Create the hw shader */
2022 GL_EXTCALL(glGenProgramsARB(1, &pshader->prgId));
2023 TRACE_(d3d_hw_shader)("Creating a hw pixel shader, prg=%d\n", pshader->prgId);
2024
2025 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pshader->prgId));
2026
2027 /* Create the program and check for errors */
2028 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(pgmStr), pgmStr));
2029 if (glGetError() == GL_INVALID_OPERATION) {
2030 GLint errPos;
2031 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2032 FIXME_(d3d_hw_shader)("HW PixelShader Error at position: %d\n%s\n", errPos, glGetString(GL_PROGRAM_ERROR_STRING_ARB));
2033 pshader->prgId = -1;
2034 }
2035
2036 HeapFree(GetProcessHeap(), 0, pgmStr);
2037 }
2038
2039 inline static VOID IDirect3DPixelShaderImpl_ParseProgram(IDirect3DPixelShaderImpl* pshader, CONST DWORD* pFunction) {
2040 const DWORD* pToken = pFunction;
2041 const SHADER_OPCODE* curOpcode = NULL;
2042 DWORD code;
2043 DWORD len = 0;
2044 DWORD i;
2045
2046 if (NULL != pToken) {
2047 while (D3DPS_END() != *pToken) {
2048 if (pshader_is_version_token(*pToken)) { /** version */
2049 TRACE("ps.%lu.%lu\n", (*pToken >> 8) & 0x0F, (*pToken & 0x0F));
2050 ++pToken;
2051 ++len;
2052 continue;
2053 }
2054 if (pshader_is_comment_token(*pToken)) { /** comment */
2055 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
2056 ++pToken;
2057 /*TRACE("comment[%ld] ;%s\n", comment_len, (char*)pToken);*/
2058 pToken += comment_len;
2059 len += comment_len + 1;
2060 continue;
2061 }
2062 code = *pToken;
2063 curOpcode = pshader_program_get_opcode(code);
2064 ++pToken;
2065 ++len;
2066 if (NULL == curOpcode) {
2067 /* unkown current opcode ... */
2068 while (*pToken & 0x80000000) {
2069 TRACE("unrecognized opcode: %08lx\n", *pToken);
2070 ++pToken;
2071 ++len;
2072 }
2073 } else {
2074 TRACE(" ");
2075 pshader_program_dump_opcode(curOpcode, code, *pToken);
2076 if (curOpcode->num_params > 0) {
2077 pshader_program_dump_param(*pToken, 0);
2078 ++pToken;
2079 ++len;
2080 for (i = 1; i < curOpcode->num_params; ++i) {
2081 TRACE(", ");
2082 if (D3DSIO_DEF != code) {
2083 pshader_program_dump_param(*pToken, 1);
2084 } else {
2085 TRACE("%f", *((float*) pToken));
2086 }
2087 ++pToken;
2088 ++len;
2089 }
2090 }
2091 TRACE("\n");
2092 }
2093 pshader->functionLength = (len + 1) * sizeof(DWORD);
2094 }
2095 } else {
2096 pshader->functionLength = 1; /* no Function defined use fixed function vertex processing */
2097 }
2098
2099 if (NULL != pFunction) {
2100 IDirect3DPixelShaderImpl_GenerateProgramArbHW(pshader, pFunction);
2101 }
2102
2103 if (NULL != pFunction) {
2104 pshader->function = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, pshader->functionLength);
2105 memcpy(pshader->function, pFunction, pshader->functionLength);
2106 } else {
2107 pshader->function = NULL;
2108 }
2109 }
2110
2111 HRESULT WINAPI IDirect3DDeviceImpl_CreatePixelShader(IDirect3DDevice8Impl* This, CONST DWORD* pFunction, IDirect3DPixelShaderImpl** ppPixelShader) {
2112 IDirect3DPixelShaderImpl* object;
2113
2114 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(IDirect3DPixelShaderImpl));
2115 if (NULL == object) {
2116 *ppPixelShader = NULL;
2117 return D3DERR_OUTOFVIDEOMEMORY;
2118 }
2119 /*object->lpVtbl = &Direct3DPixelShader9_Vtbl;*/
2120 object->device = This;
2121 object->ref = 1;
2122
2123 object->data = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(PSHADERDATA8));
2124
2125 IDirect3DPixelShaderImpl_ParseProgram(object, pFunction);
2126
2127 *ppPixelShader = object;
2128 return D3D_OK;
2129 }
2130
2131 HRESULT WINAPI IDirect3DPixelShaderImpl_GetFunction(IDirect3DPixelShaderImpl* This, VOID* pData, UINT* pSizeOfData) {
2132 if (NULL == pData) {
2133 *pSizeOfData = This->functionLength;
2134 return D3D_OK;
2135 }
2136 if (*pSizeOfData < This->functionLength) {
2137 *pSizeOfData = This->functionLength;
2138 return D3DERR_MOREDATA;
2139 }
2140 if (NULL == This->function) { /* no function defined */
2141 TRACE("(%p) : GetFunction no User Function defined using NULL to %p\n", This, pData);
2142 (*(DWORD **) pData) = NULL;
2143 } else {
2144 TRACE("(%p) : GetFunction copying to %p\n", This, pData);
2145 memcpy(pData, This->function, This->functionLength);
2146 }
2147 return D3D_OK;
2148 }
2149
2150 HRESULT WINAPI IDirect3DPixelShaderImpl_SetConstantF(IDirect3DPixelShaderImpl* This, UINT StartRegister, CONST FLOAT* pConstantData, UINT Vector4fCount) {
2151 if (StartRegister + Vector4fCount > D3D8_VSHADER_MAX_CONSTANTS) {
2152 return D3DERR_INVALIDCALL;
2153 }
2154 if (NULL == This->data) { /* temporary while datas not supported */
2155 FIXME("(%p) : VertexShader_SetConstant not fully supported yet\n", This);
2156 return D3DERR_INVALIDCALL;
2157 }
2158 memcpy(&This->data->C[StartRegister], pConstantData, Vector4fCount * 4 * sizeof(FLOAT));
2159 return D3D_OK;
2160 }
2161
2162 HRESULT WINAPI IDirect3DPixelShaderImpl_GetConstantF(IDirect3DPixelShaderImpl* This, UINT StartRegister, FLOAT* pConstantData, UINT Vector4fCount) {
2163 if (StartRegister + Vector4fCount > D3D8_VSHADER_MAX_CONSTANTS) {
2164 return D3DERR_INVALIDCALL;
2165 }
2166 if (NULL == This->data) { /* temporary while datas not supported */
2167 return D3DERR_INVALIDCALL;
2168 }
2169 memcpy(pConstantData, &This->data->C[StartRegister], Vector4fCount * 4 * sizeof(FLOAT));
2170 return D3D_OK;
2171 }
2172
2173
2174 /**********************************************************************************************************************************************
2175 **********************************************************************************************************************************************
2176 **********************************************************************************************************************************************
2177 **********************************************************************************************************************************************
2178 **********************************************************************************************************************************************/
2179
2180 /***********************************************************************
2181 * ValidateVertexShader (D3D8.@)
2182 */
2183 BOOL WINAPI ValidateVertexShader(LPVOID what, LPVOID toto) {
2184 FIXME("(void): stub: %p %p\n", what, toto);
2185 return TRUE;
2186 }
2187
2188 /***********************************************************************
2189 * ValidatePixelShader (D3D8.@)
2190 */
2191 BOOL WINAPI ValidatePixelShader(LPVOID what, LPVOID toto) {
2192 FIXME("(void): stub: %p %p\n", what, toto);
2193 return TRUE;
2194 }
useful hacks and other patches not (yet) suitable for mainline