1 /**************************************************************************
2 *
3 * Copyright 2010-2021 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include <limits.h>
29 #include "util/simple_list.h"
30
31 #include "pipe/p_defines.h"
32 #include "util/u_inlines.h"
33 #include "util/u_memory.h"
34 #include "util/u_pointer.h"
35 #include "util/format/u_format.h"
36 #include "util/u_dump.h"
37 #include "util/u_string.h"
38 #include "util/os_time.h"
39 #include "pipe/p_shader_tokens.h"
40 #include "draw/draw_context.h"
41 #include "tgsi/tgsi_dump.h"
42 #include "tgsi/tgsi_scan.h"
43 #include "tgsi/tgsi_parse.h"
44 #include "gallivm/lp_bld_type.h"
45 #include "gallivm/lp_bld_const.h"
46 #include "gallivm/lp_bld_conv.h"
47 #include "gallivm/lp_bld_init.h"
48 #include "gallivm/lp_bld_intr.h"
49 #include "gallivm/lp_bld_logic.h"
50 #include "gallivm/lp_bld_tgsi.h"
51 #include "gallivm/lp_bld_swizzle.h"
52 #include "gallivm/lp_bld_flow.h"
53 #include "gallivm/lp_bld_printf.h"
54 #include "gallivm/lp_bld_debug.h"
55
56 #include "lp_bld_alpha.h"
57 #include "lp_bld_blend.h"
58 #include "lp_bld_depth.h"
59 #include "lp_bld_interp.h"
60 #include "lp_context.h"
61 #include "lp_debug.h"
62 #include "lp_perf.h"
63 #include "lp_screen.h"
64 #include "lp_setup.h"
65 #include "lp_state.h"
66 #include "lp_tex_sample.h"
67 #include "lp_flush.h"
68 #include "lp_state_fs.h"
69
70
71 /**
72 * Sampler.
73 */
74 struct linear_sampler
75 {
76 struct lp_build_sampler_aos base;
77
78 LLVMValueRef texels_ptrs[LP_MAX_LINEAR_TEXTURES];
79
80 LLVMValueRef counter;
81
82 unsigned instance;
83 };
84
85
86 /**
87 * Provide texels to the TGSI translation.
88 *
89 * We don't actually do any texture sampling here, but simply hand the
90 * precomputed row of texels.
91 */
92 static LLVMValueRef
emit_fetch_texel_linear(const struct lp_build_sampler_aos * base,struct lp_build_context * bld,unsigned target,unsigned unit,LLVMValueRef coords,const struct lp_derivatives derivs,enum lp_build_tex_modifier modifier)93 emit_fetch_texel_linear(const struct lp_build_sampler_aos *base,
94 struct lp_build_context *bld,
95 unsigned target, /* TGSI_TEXTURE_* */
96 unsigned unit,
97 LLVMValueRef coords,
98 const struct lp_derivatives derivs,
99 enum lp_build_tex_modifier modifier)
100 {
101 struct linear_sampler *sampler = (struct linear_sampler *)base;
102 LLVMValueRef texels_ptr;
103 LLVMValueRef texel;
104
105 if (sampler->instance >= LP_MAX_LINEAR_TEXTURES) {
106 assert(FALSE);
107 return bld->undef;
108 }
109
110 /* Pointer to a row of texels */
111 texels_ptr = sampler->texels_ptrs[sampler->instance];
112
113 texel = lp_build_pointer_get(bld->gallivm->builder, texels_ptr,
114 sampler->counter);
115 assert(LLVMTypeOf(texel) == bld->vec_type);
116
117 /*
118 * We have a struct lp_linear_sampler instance per TEX instruction,
119 * _not_ per unit, as each TEX intruction will need separate storage for the
120 * texels.
121 */
122 (void)unit;
123 ++sampler->instance;
124
125 return texel;
126 }
127
128 /**
129 * Generates the main body of the fragment shader
130 * Supports generating code for 4 pixel blocks and individual pixels
131 */
132 static LLVMValueRef
llvm_fragment_body(struct lp_build_context * bld,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant * variant,struct linear_sampler * sampler,LLVMValueRef * inputs_ptrs,LLVMValueRef consts_ptr,LLVMValueRef blend_color,LLVMValueRef alpha_ref,struct lp_type fs_type,LLVMValueRef dst)133 llvm_fragment_body(struct lp_build_context *bld,
134 struct lp_fragment_shader *shader,
135 struct lp_fragment_shader_variant *variant,
136 struct linear_sampler* sampler,
137 LLVMValueRef *inputs_ptrs,
138 LLVMValueRef consts_ptr,
139 LLVMValueRef blend_color,
140 LLVMValueRef alpha_ref,
141 struct lp_type fs_type,
142 LLVMValueRef dst)
143 {
144 const unsigned char bgra_swizzles[4] = {2, 1, 0, 3};
145
146 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS];
147 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS];
148
149 LLVMBuilderRef builder = bld->gallivm->builder;
150 struct gallivm_state *gallivm = bld->gallivm;
151
152 LLVMValueRef src1 = lp_build_zero(gallivm, fs_type);
153
154 LLVMValueRef result = NULL;
155 unsigned i;
156
157 sampler->instance = 0;
158
159
160 /*
161 * Advance inputs
162 */
163
164 for (i = 0; i < shader->info.base.num_inputs; ++i) {
165 LLVMValueRef inputs_ptr;
166 LLVMValueRef input;
167
168 inputs_ptr = inputs_ptrs[i];
169
170 input = lp_build_pointer_get(builder, inputs_ptr, sampler->counter);
171 assert(LLVMTypeOf(input) == bld->vec_type);
172
173 inputs[i] = input;
174 }
175
176 for ( ; i < PIPE_MAX_SHADER_INPUTS; ++i) {
177 inputs[i] = bld->undef;
178 }
179
180
181 /*
182 * Translate the TGSI
183 */
184
185 for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; ++i) {
186 outputs[i] = bld->undef;
187 }
188
189 lp_build_tgsi_aos(gallivm, shader->base.tokens, fs_type,
190 bgra_swizzles,
191 consts_ptr, inputs, outputs,
192 &sampler->base,
193 &shader->info.base);
194
195 /*
196 * Blend output color
197 */
198
199 for (i = 0; i < shader->info.base.num_outputs; ++i) {
200 LLVMValueRef mask = NULL;
201 LLVMValueRef output;
202 unsigned cbuf;
203
204 if (!outputs[i])
205 continue;
206
207 output = LLVMBuildLoad(builder, outputs[i], "");
208 lp_build_name(output, "output%u", i);
209
210 cbuf = shader->info.base.output_semantic_index[i];
211 lp_build_name(output, "cbuf%u", cbuf);
212
213 if (shader->info.base.output_semantic_name[i] != TGSI_SEMANTIC_COLOR || cbuf != 0)
214 continue;
215
216 /* Perform alpha test if necessary */
217 if (variant->key.alpha.enabled) {
218 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, fs_type);
219 LLVMValueRef broadcast_alpha = lp_build_broadcast(gallivm, vec_type, alpha_ref);
220
221 mask = lp_build_cmp(bld, variant->key.alpha.func, output, broadcast_alpha);
222 /* XXX is 4 correct? */
223 mask = lp_build_swizzle_scalar_aos(bld, mask, bgra_swizzles[3], 4);
224
225 lp_build_name(mask, "alpha_test_mask");
226 }
227
228 result = lp_build_blend_aos(gallivm,
229 &variant->key.blend,
230 variant->key.cbuf_format[i],
231 fs_type,
232 cbuf, /* rt */
233 output, /* src */
234 NULL, /* src_alpha */
235 src1, /* src1 */
236 NULL, /* src1_alpha */
237 dst,
238 mask,
239 blend_color, /* const_ */
240 NULL, /* const_alpha */
241 bgra_swizzles,
242 4);
243 }
244
245 return result;
246 }
247
248 /**
249 * Generate a function that executes the fragment shader in a linear fashion.
250 */
251 void
llvmpipe_fs_variant_linear_llvm(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant * variant)252 llvmpipe_fs_variant_linear_llvm(struct llvmpipe_context *lp,
253 struct lp_fragment_shader *shader,
254 struct lp_fragment_shader_variant *variant)
255 {
256 struct gallivm_state *gallivm = variant->gallivm;
257 char func_name[256];
258 struct lp_type fs_type;
259 LLVMTypeRef ret_type;
260 LLVMTypeRef arg_types[4];
261 LLVMTypeRef func_type;
262 LLVMValueRef context_ptr;
263 LLVMValueRef x;
264 LLVMValueRef y;
265 LLVMValueRef width;
266 LLVMValueRef excess;
267 LLVMValueRef function;
268 LLVMBasicBlockRef block;
269 LLVMBuilderRef builder;
270 struct lp_build_context bld;
271 struct linear_sampler sampler;
272 struct lp_build_if_state ifstate;
273 struct lp_build_for_loop_state loop;
274
275 LLVMValueRef consts_ptr;
276 LLVMValueRef interpolators_ptr;
277 LLVMValueRef samplers_ptr;
278 LLVMValueRef color0_ptr;
279 LLVMValueRef blend_color;
280 LLVMValueRef alpha_ref;
281
282 LLVMValueRef inputs_ptrs[LP_MAX_LINEAR_INPUTS];
283
284 LLVMTypeRef int8t = LLVMInt8TypeInContext(gallivm->context);
285 LLVMTypeRef int32t = LLVMInt32TypeInContext(gallivm->context);
286 LLVMTypeRef pint8t = LLVMPointerType(int8t, 0);
287 LLVMTypeRef pixelt = LLVMVectorType(int32t, 4);
288
289 unsigned attrib;
290 unsigned i;
291
292 memset(&fs_type, 0, sizeof fs_type);
293 fs_type.floating = FALSE;
294 fs_type.sign = FALSE;
295 fs_type.norm = TRUE;
296 fs_type.width = 8;
297 fs_type.length = 16;
298
299 if (LP_DEBUG & DEBUG_TGSI) {
300 tgsi_dump(shader->base.tokens, 0);
301 }
302
303 /*
304 * Generate the function prototype. Any change here must be reflected in
305 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
306 */
307
308 snprintf(func_name, sizeof(func_name), "fs%u_variant%u_linear",
309 shader->no, variant->no);
310
311 ret_type = pint8t;
312 arg_types[0] = variant->jit_linear_context_ptr_type; /* context */
313 arg_types[1] = int32t; /* x */
314 arg_types[2] = int32t; /* y */
315 arg_types[3] = int32t; /* width */
316
317 func_type = LLVMFunctionType(ret_type, arg_types, ARRAY_SIZE(arg_types), 0);
318
319 function = LLVMAddFunction(gallivm->module, func_name, func_type);
320 LLVMSetFunctionCallConv(function, LLVMCCallConv);
321
322 variant->linear_function = function;
323
324 /* XXX: need to propagate noalias down into color param now we are
325 * passing a pointer-to-pointer?
326 */
327 for (i = 0; i < ARRAY_SIZE(arg_types); ++i) {
328 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
329 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
330 }
331 }
332
333 context_ptr = LLVMGetParam(function, 0);
334 x = LLVMGetParam(function, 1);
335 y = LLVMGetParam(function, 2);
336 width = LLVMGetParam(function, 3);
337
338 lp_build_name(context_ptr, "context");
339 lp_build_name(x, "x");
340 lp_build_name(y, "y");
341 lp_build_name(width, "width");
342
343 /*
344 * Function body
345 */
346
347 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
348 builder = gallivm->builder;
349
350 LLVMPositionBuilderAtEnd(builder, block);
351
352 lp_build_context_init(&bld, gallivm, fs_type);
353
354 /*
355 * Get context data
356 */
357
358 consts_ptr = lp_jit_linear_context_constants(gallivm, context_ptr);
359 interpolators_ptr = lp_jit_linear_context_inputs(gallivm, context_ptr);
360 samplers_ptr = lp_jit_linear_context_tex(gallivm, context_ptr);
361
362 color0_ptr = lp_jit_linear_context_color0(gallivm, context_ptr);
363 color0_ptr = LLVMBuildLoad(builder, color0_ptr, "");
364 color0_ptr = LLVMBuildBitCast(builder, color0_ptr, LLVMPointerType(bld.vec_type, 0), "");
365
366 blend_color = lp_jit_linear_context_blend_color(gallivm, context_ptr);
367 blend_color = LLVMBuildLoad(builder, blend_color, "");
368 blend_color = lp_build_broadcast(gallivm, LLVMVectorType(int32t, 4), blend_color);
369 blend_color = LLVMBuildBitCast(builder, blend_color, LLVMVectorType(int8t, 16), "");
370
371 alpha_ref = lp_jit_linear_context_alpha_ref(gallivm, context_ptr);
372 alpha_ref = LLVMBuildLoad(builder, alpha_ref, "");
373
374 /*
375 * Invoke the input interpolators
376 */
377
378 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
379 LLVMValueRef index;
380 LLVMValueRef elem;
381 LLVMValueRef fetch_ptr;
382 LLVMValueRef inputs_ptr;
383
384 assert(attrib < LP_MAX_LINEAR_INPUTS);
385 if (attrib >= LP_MAX_LINEAR_INPUTS) {
386 break;
387 }
388
389 index = LLVMConstInt(int32t, attrib, 0);
390
391 elem = lp_build_array_get(bld.gallivm, interpolators_ptr, index);
392 assert(LLVMGetTypeKind(LLVMTypeOf(elem)) == LLVMPointerTypeKind);
393
394 fetch_ptr = lp_build_pointer_get(builder, elem,
395 LLVMConstInt(int32t, 0, 0));
396 assert(LLVMGetTypeKind(LLVMTypeOf(fetch_ptr)) == LLVMPointerTypeKind);
397
398 /* Pointer to a row of interpolated inputs */
399 elem = LLVMBuildBitCast(builder, elem, pint8t, "");
400 inputs_ptr = LLVMBuildCall(builder, fetch_ptr, &elem, 1, "");
401 assert(LLVMGetTypeKind(LLVMTypeOf(inputs_ptr)) == LLVMPointerTypeKind);
402
403 /* Mark the function read-only so that LLVM can optimize it away */
404 lp_add_function_attr(inputs_ptr, -1, LP_FUNC_ATTR_READONLY);
405 lp_add_function_attr(inputs_ptr, -1, LP_FUNC_ATTR_NOUNWIND);
406
407 lp_build_name(inputs_ptr, "input%u_ptr", attrib);
408
409 inputs_ptrs[attrib] = inputs_ptr;
410 }
411
412 /*
413 * Invoke and hook up the texture samplers.
414 */
415
416 memset(&sampler, 0, sizeof sampler);
417 sampler.base.emit_fetch_texel = &emit_fetch_texel_linear;
418
419 for (attrib = 0; attrib < shader->info.num_texs; ++attrib) {
420 LLVMValueRef index;
421 LLVMValueRef elem;
422 LLVMValueRef fetch_ptr;
423 LLVMValueRef texels_ptr;
424
425 assert(attrib < LP_MAX_LINEAR_TEXTURES);
426 if (attrib >= LP_MAX_LINEAR_TEXTURES) {
427 break;
428 }
429
430 index = LLVMConstInt(int32t, attrib, 0);
431
432 elem = lp_build_array_get(bld.gallivm, samplers_ptr, index);
433 assert(LLVMGetTypeKind(LLVMTypeOf(elem)) == LLVMPointerTypeKind);
434
435 fetch_ptr = lp_build_pointer_get(builder, elem, LLVMConstInt(int32t, 0, 0));
436 assert(LLVMGetTypeKind(LLVMTypeOf(fetch_ptr)) == LLVMPointerTypeKind);
437
438 /* Pointer to a row of texels */
439 elem = LLVMBuildBitCast(builder, elem, pint8t, "");
440 texels_ptr = LLVMBuildCall(builder, fetch_ptr, &elem, 1, "");
441 assert(LLVMGetTypeKind(LLVMTypeOf(texels_ptr)) == LLVMPointerTypeKind);
442
443 /* Mark the function read-only so that LLVM can optimize it away */
444 lp_add_function_attr(texels_ptr, -1, LP_FUNC_ATTR_READONLY);
445 lp_add_function_attr(texels_ptr, -1, LP_FUNC_ATTR_NOUNWIND);
446
447 lp_build_name(texels_ptr, "tex%u_ptr", attrib);
448
449 sampler.texels_ptrs[attrib] = texels_ptr;
450 }
451
452 excess = LLVMBuildAnd(builder, width, LLVMConstInt(int32t, 3, 0), "");
453 width = LLVMBuildLShr(builder, width, LLVMConstInt(int32t, 2, 0), "");
454
455 /* Loop over blocks of 4 pixels */
456 lp_build_for_loop_begin(&loop, gallivm, LLVMConstInt(int32t, 0, 0), LLVMIntULT, width, LLVMConstInt(int32t, 1, 0));
457 {
458 LLVMValueRef value;
459 sampler.counter = loop.counter;
460
461 /* Read 4 pixels */
462 value = lp_build_pointer_get_unaligned(builder, color0_ptr, loop.counter, 4);
463
464 /* Perform fragment shader body */
465 value = llvm_fragment_body(&bld, shader, variant, &sampler, inputs_ptrs, consts_ptr, blend_color, alpha_ref, fs_type, value);
466
467 /* Write 4 pixels */
468 lp_build_pointer_set_unaligned(builder, color0_ptr, loop.counter, value, 4);
469 }
470 lp_build_for_loop_end(&loop);
471
472 /* Compute the edge pixels (width % 4) */
473 lp_build_if(&ifstate, gallivm, LLVMBuildICmp(builder, LLVMIntNE, excess, LLVMConstInt(int32t, 0, 0), ""));
474 {
475 struct lp_build_loop_state loop_read, loop_write;
476 LLVMValueRef buf, elem, result, pixel_ptr;
477 LLVMValueRef buf_ptr = lp_build_alloca(gallivm, pixelt, "");
478
479 sampler.counter = width;
480
481 /* Get the i32* pixel pointer from the <i16x8>* element pointer */
482 pixel_ptr = LLVMBuildGEP(gallivm->builder, color0_ptr, &width, 1, "");
483 pixel_ptr = LLVMBuildBitCast(gallivm->builder, pixel_ptr, LLVMPointerType(int32t, 0), "");
484
485 /* Copy individual pixels from memory to local buffer */
486 lp_build_loop_begin(&loop_read, gallivm, LLVMConstInt(int32t, 0, 0));
487 {
488 elem = lp_build_pointer_get(gallivm->builder, pixel_ptr, loop_read.counter);
489
490 buf = LLVMBuildLoad(gallivm->builder, buf_ptr, "");
491 buf = LLVMBuildInsertElement(builder, buf, elem, loop_read.counter, "");
492 LLVMBuildStore(builder, buf, buf_ptr);
493 }
494 lp_build_loop_end_cond(&loop_read, excess, LLVMConstInt(int32t, 1, 0), LLVMIntUGE);
495
496 /* Perform fragment shader body */
497 buf = LLVMBuildLoad(gallivm->builder, buf_ptr, "");
498 buf = LLVMBuildBitCast(builder, buf, bld.vec_type, "");
499
500 result = llvm_fragment_body(&bld, shader, variant, &sampler, inputs_ptrs, consts_ptr, blend_color, alpha_ref, fs_type, buf);
501 result = LLVMBuildBitCast(builder, result, pixelt, "");
502
503 /* Write individual pixels from local buffer to the memory */
504 lp_build_loop_begin(&loop_write, gallivm, LLVMConstInt(int32t, 0, 0));
505 {
506 elem = LLVMBuildExtractElement(builder, result, loop_write.counter, "");
507
508 lp_build_pointer_set(gallivm->builder, pixel_ptr, loop_write.counter, elem);
509 }
510 lp_build_loop_end_cond(&loop_write, excess, LLVMConstInt(int32t, 1, 0), LLVMIntUGE);
511 }
512 lp_build_endif(&ifstate);
513
514
515 color0_ptr = LLVMBuildBitCast(builder, color0_ptr, pint8t, "");
516
517 LLVMBuildRet(builder, color0_ptr);
518
519 gallivm_verify_function(gallivm, function);
520 }
521
522
523