1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
38 * - depth/stencil test (stencil TBI)
41 * This file has only the glue to assembly the fragment pipeline. The actual
42 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
43 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
44 * muster the LLVM JIT execution engine to create a function that follows an
45 * established binary interface and that can be called from C directly.
47 * A big source of complexity here is that we often want to run different
48 * stages with different precisions and data types and precisions. For example,
49 * the fragment shader needs typically to be done in floats, but the
50 * depth/stencil test and blending is better done in the type that most closely
51 * matches the depth/stencil and color buffer respectively.
53 * Since the width of a SIMD vector register stays the same regardless of the
54 * element type, different types imply different number of elements, so we must
55 * code generate more instances of the stages with larger types to be able to
56 * feed/consume the stages with smaller types.
58 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_memory.h"
63 #include "util/u_format.h"
64 #include "util/u_debug_dump.h"
65 #include "pipe/internal/p_winsys_screen.h"
66 #include "pipe/p_shader_tokens.h"
67 #include "draw/draw_context.h"
68 #include "tgsi/tgsi_dump.h"
69 #include "tgsi/tgsi_scan.h"
70 #include "tgsi/tgsi_parse.h"
71 #include "lp_bld_type.h"
72 #include "lp_bld_const.h"
73 #include "lp_bld_conv.h"
74 #include "lp_bld_intr.h"
75 #include "lp_bld_logic.h"
76 #include "lp_bld_depth.h"
77 #include "lp_bld_interp.h"
78 #include "lp_bld_tgsi.h"
79 #include "lp_bld_alpha.h"
80 #include "lp_bld_blend.h"
81 #include "lp_bld_swizzle.h"
82 #include "lp_bld_flow.h"
83 #include "lp_bld_debug.h"
84 #include "lp_screen.h"
85 #include "lp_context.h"
86 #include "lp_buffer.h"
89 #include "lp_tex_sample.h"
93 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
94 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
98 * Derive from the quad's upper left scalar coordinates the coordinates for
99 * all other quad pixels
102 generate_pos0(LLVMBuilderRef builder,
108 LLVMTypeRef int_elem_type = LLVMInt32Type();
109 LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
110 LLVMTypeRef elem_type = LLVMFloatType();
111 LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
112 LLVMValueRef x_offsets[QUAD_SIZE];
113 LLVMValueRef y_offsets[QUAD_SIZE];
116 x = lp_build_broadcast(builder, int_vec_type, x);
117 y = lp_build_broadcast(builder, int_vec_type, y);
119 for(i = 0; i < QUAD_SIZE; ++i) {
120 x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
121 y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
124 x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
125 y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
127 *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
128 *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
133 * Generate the depth test.
136 generate_depth(LLVMBuilderRef builder,
137 const struct lp_fragment_shader_variant_key *key,
138 struct lp_type src_type,
139 struct lp_build_mask_context *mask,
141 LLVMValueRef dst_ptr)
143 const struct util_format_description *format_desc;
144 struct lp_type dst_type;
146 if(!key->depth.enabled)
149 format_desc = util_format_description(key->zsbuf_format);
153 * Depths are expected to be between 0 and 1, even if they are stored in
154 * floats. Setting these bits here will ensure that the lp_build_conv() call
155 * below won't try to unnecessarily clamp the incoming values.
157 if(src_type.floating) {
158 src_type.sign = FALSE;
159 src_type.norm = TRUE;
162 assert(!src_type.sign);
163 assert(src_type.norm);
166 /* Pick the depth type. */
167 dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
169 /* FIXME: Cope with a depth test type with a different bit width. */
170 assert(dst_type.width == src_type.width);
171 assert(dst_type.length == src_type.length);
173 lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
175 dst_ptr = LLVMBuildBitCast(builder,
177 LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
179 lp_build_depth_test(builder,
190 * Generate the code to do inside/outside triangle testing for the
191 * four pixels in a 2x2 quad. This will set the four elements of the
192 * quad mask vector to 0 or ~0.
193 * \param i which quad of the quad group to test, in [0,3]
196 generate_tri_edge_mask(LLVMBuilderRef builder,
198 LLVMValueRef *mask, /* ivec4, out */
199 LLVMValueRef c0, /* int32 */
200 LLVMValueRef c1, /* int32 */
201 LLVMValueRef c2, /* int32 */
202 LLVMValueRef step0_ptr, /* ivec4 */
203 LLVMValueRef step1_ptr, /* ivec4 */
204 LLVMValueRef step2_ptr) /* ivec4 */
210 m0_vec = step0_ptr[i] > c0_vec
211 m1_vec = step1_ptr[i] > c1_vec
212 m2_vec = step2_ptr[i] > c2_vec
213 mask = m0_vec & m1_vec & m2_vec
215 struct lp_type i32_type;
216 LLVMTypeRef i32vec4_type;
219 LLVMValueRef c0_vec, c1_vec, c2_vec;
220 LLVMValueRef step0_vec, step1_vec, step2_vec;
221 LLVMValueRef m0_vec, m1_vec, m2_vec;
226 /* int32 vector type */
227 memset(&i32_type, 0, sizeof i32_type);
228 i32_type.floating = FALSE; /* values are integers */
229 i32_type.sign = TRUE; /* values are signed */
230 i32_type.norm = FALSE; /* values are not normalized */
231 i32_type.width = 32; /* 32-bit int values */
232 i32_type.length = 4; /* 4 elements per vector */
234 i32vec4_type = lp_build_int32_vec4_type();
236 /* c0_vec = {c0, c0, c0, c0}
237 * Note that we emit this code four times but LLVM optimizes away
238 * three instances of it.
240 c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
241 c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
242 c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
244 lp_build_name(c0_vec, "edgeconst0vec");
245 lp_build_name(c1_vec, "edgeconst1vec");
246 lp_build_name(c2_vec, "edgeconst2vec");
248 index = LLVMConstInt(LLVMInt32Type(), i, 0);
249 step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
250 step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
251 step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
253 lp_build_name(step0_vec, "step0vec");
254 lp_build_name(step1_vec, "step1vec");
255 lp_build_name(step2_vec, "step2vec");
257 m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
258 m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
259 m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
261 m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
262 m = LLVMBuildAnd(builder, m, m2_vec, "");
264 lp_build_name(m, "inoutmaskvec");
269 * if mask = {0,0,0,0} skip quad
275 * Generate the fragment shader, depth/stencil test, and alpha tests.
276 * \param i which quad in the tile, in range [0,3]
279 generate_fs(struct llvmpipe_context *lp,
280 struct lp_fragment_shader *shader,
281 const struct lp_fragment_shader_variant_key *key,
282 LLVMBuilderRef builder,
284 LLVMValueRef context_ptr,
286 const struct lp_build_interp_soa_context *interp,
287 struct lp_build_sampler_soa *sampler,
290 LLVMValueRef depth_ptr,
294 LLVMValueRef step0_ptr,
295 LLVMValueRef step1_ptr,
296 LLVMValueRef step2_ptr)
298 const struct tgsi_token *tokens = shader->base.tokens;
299 LLVMTypeRef elem_type;
300 LLVMTypeRef vec_type;
301 LLVMTypeRef int_vec_type;
302 LLVMValueRef consts_ptr;
303 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
304 LLVMValueRef z = interp->pos[2];
305 struct lp_build_flow_context *flow;
306 struct lp_build_mask_context mask;
307 boolean early_depth_test;
313 elem_type = lp_build_elem_type(type);
314 vec_type = lp_build_vec_type(type);
315 int_vec_type = lp_build_int_vec_type(type);
317 consts_ptr = lp_jit_context_constants(builder, context_ptr);
319 flow = lp_build_flow_create(builder);
321 memset(outputs, 0, sizeof outputs);
323 lp_build_flow_scope_begin(flow);
325 /* Declare the color and z variables */
326 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
327 color[chan] = LLVMGetUndef(vec_type);
328 lp_build_flow_scope_declare(flow, &color[chan]);
330 lp_build_flow_scope_declare(flow, &z);
332 /* do triangle edge testing */
333 generate_tri_edge_mask(builder, i, pmask,
334 c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
336 /* 'mask' will control execution based on quad's pixel alive/killed state */
337 lp_build_mask_begin(&mask, flow, type, *pmask);
341 key->depth.enabled &&
342 !key->alpha.enabled &&
343 !shader->info.uses_kill &&
344 !shader->info.writes_z;
347 generate_depth(builder, key,
351 lp_build_tgsi_soa(builder, tokens, type, &mask,
352 consts_ptr, interp->pos, interp->inputs,
355 for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
356 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
357 if(outputs[attrib][chan]) {
358 lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
360 switch (shader->info.output_semantic_name[attrib]) {
361 case TGSI_SEMANTIC_COLOR:
363 unsigned cbuf = shader->info.output_semantic_index[attrib];
365 lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
368 /* XXX: should the alpha reference value be passed separately? */
369 if(cbuf == 0 && chan == 3) {
370 LLVMValueRef alpha = outputs[attrib][chan];
371 LLVMValueRef alpha_ref_value;
372 alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
373 alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
374 lp_build_alpha_test(builder, &key->alpha, type,
375 &mask, alpha, alpha_ref_value);
379 color[chan] = outputs[attrib][chan];
384 case TGSI_SEMANTIC_POSITION:
386 z = outputs[attrib][chan];
393 if(!early_depth_test)
394 generate_depth(builder, key,
398 lp_build_mask_end(&mask);
400 lp_build_flow_scope_end(flow);
402 lp_build_flow_destroy(flow);
410 * Generate color blending and color output.
413 generate_blend(const struct pipe_blend_state *blend,
414 LLVMBuilderRef builder,
416 LLVMValueRef context_ptr,
419 LLVMValueRef dst_ptr)
421 struct lp_build_context bld;
422 struct lp_build_flow_context *flow;
423 struct lp_build_mask_context mask_ctx;
424 LLVMTypeRef vec_type;
425 LLVMTypeRef int_vec_type;
426 LLVMValueRef const_ptr;
432 lp_build_context_init(&bld, builder, type);
434 flow = lp_build_flow_create(builder);
435 lp_build_mask_begin(&mask_ctx, flow, type, mask);
437 vec_type = lp_build_vec_type(type);
438 int_vec_type = lp_build_int_vec_type(type);
440 const_ptr = lp_jit_context_blend_color(builder, context_ptr);
441 const_ptr = LLVMBuildBitCast(builder, const_ptr,
442 LLVMPointerType(vec_type, 0), "");
444 for(chan = 0; chan < 4; ++chan) {
445 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
446 con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
448 dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
450 lp_build_name(con[chan], "con.%c", "rgba"[chan]);
451 lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
454 lp_build_blend_soa(builder, blend, type, src, dst, con, res);
456 for(chan = 0; chan < 4; ++chan) {
457 if(blend->colormask & (1 << chan)) {
458 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
459 lp_build_name(res[chan], "res.%c", "rgba"[chan]);
460 res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
461 LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
465 lp_build_mask_end(&mask_ctx);
466 lp_build_flow_destroy(flow);
471 * Generate the runtime callable function for the whole fragment pipeline.
472 * Note that the function which we generate operates on a block of 16
473 * pixels at at time. The block contains 2x2 quads. Each quad contains
476 static struct lp_fragment_shader_variant *
477 generate_fragment(struct llvmpipe_context *lp,
478 struct lp_fragment_shader *shader,
479 const struct lp_fragment_shader_variant_key *key)
481 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
482 struct lp_fragment_shader_variant *variant;
483 struct lp_type fs_type;
484 struct lp_type blend_type;
485 LLVMTypeRef fs_elem_type;
486 LLVMTypeRef fs_vec_type;
487 LLVMTypeRef fs_int_vec_type;
488 LLVMTypeRef blend_vec_type;
489 LLVMTypeRef blend_int_vec_type;
490 LLVMTypeRef arg_types[14];
491 LLVMTypeRef func_type;
492 LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
493 LLVMValueRef context_ptr;
497 LLVMValueRef dadx_ptr;
498 LLVMValueRef dady_ptr;
499 LLVMValueRef color_ptr;
500 LLVMValueRef depth_ptr;
501 LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
502 LLVMBasicBlockRef block;
503 LLVMBuilderRef builder;
506 struct lp_build_sampler_soa *sampler;
507 struct lp_build_interp_soa_context interp;
508 LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
509 LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
510 LLVMValueRef blend_mask;
511 LLVMValueRef blend_in_color[NUM_CHANNELS];
516 if (LP_DEBUG & DEBUG_JIT) {
517 tgsi_dump(shader->base.tokens, 0);
518 if(key->depth.enabled) {
519 debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
520 debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
521 debug_printf("depth.writemask = %u\n", key->depth.writemask);
523 if(key->alpha.enabled) {
524 debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
525 debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
527 if(key->blend.logicop_enable) {
528 debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
530 else if(key->blend.blend_enable) {
531 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE));
532 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
533 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
534 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE));
535 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
536 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
538 debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
539 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
540 if(key->sampler[i].format) {
541 debug_printf("sampler[%u] = \n", i);
542 debug_printf(" .format = %s\n",
543 pf_name(key->sampler[i].format));
544 debug_printf(" .target = %s\n",
545 debug_dump_tex_target(key->sampler[i].target, TRUE));
546 debug_printf(" .pot = %u %u %u\n",
547 key->sampler[i].pot_width,
548 key->sampler[i].pot_height,
549 key->sampler[i].pot_depth);
550 debug_printf(" .wrap = %s %s %s\n",
551 debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
552 debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
553 debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
554 debug_printf(" .min_img_filter = %s\n",
555 debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
556 debug_printf(" .min_mip_filter = %s\n",
557 debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
558 debug_printf(" .mag_img_filter = %s\n",
559 debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
560 if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
561 debug_printf(" .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
562 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
563 debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter);
568 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
572 variant->shader = shader;
573 memcpy(&variant->key, key, sizeof *key);
575 /* TODO: actually pick these based on the fs and color buffer
576 * characteristics. */
578 memset(&fs_type, 0, sizeof fs_type);
579 fs_type.floating = TRUE; /* floating point values */
580 fs_type.sign = TRUE; /* values are signed */
581 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
582 fs_type.width = 32; /* 32-bit float */
583 fs_type.length = 4; /* 4 elements per vector */
584 num_fs = 4; /* number of quads per block */
586 memset(&blend_type, 0, sizeof blend_type);
587 blend_type.floating = FALSE; /* values are integers */
588 blend_type.sign = FALSE; /* values are unsigned */
589 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
590 blend_type.width = 8; /* 8-bit ubyte values */
591 blend_type.length = 16; /* 16 elements per vector */
594 * Generate the function prototype. Any change here must be reflected in
595 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
598 fs_elem_type = lp_build_elem_type(fs_type);
599 fs_vec_type = lp_build_vec_type(fs_type);
600 fs_int_vec_type = lp_build_int_vec_type(fs_type);
602 blend_vec_type = lp_build_vec_type(blend_type);
603 blend_int_vec_type = lp_build_int_vec_type(blend_type);
605 arg_types[0] = screen->context_ptr_type; /* context */
606 arg_types[1] = LLVMInt32Type(); /* x */
607 arg_types[2] = LLVMInt32Type(); /* y */
608 arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
609 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
610 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
611 arg_types[6] = LLVMPointerType(blend_vec_type, 0); /* color */
612 arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
613 arg_types[8] = LLVMInt32Type(); /* c0 */
614 arg_types[9] = LLVMInt32Type(); /* c1 */
615 arg_types[10] = LLVMInt32Type(); /* c2 */
616 /* Note: the step arrays are built as int32[16] but we interpret
617 * them here as int32_vec4[4].
619 arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
620 arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
621 arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
623 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
625 variant->function = LLVMAddFunction(screen->module, "shader", func_type);
626 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
627 for(i = 0; i < Elements(arg_types); ++i)
628 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
629 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
631 context_ptr = LLVMGetParam(variant->function, 0);
632 x = LLVMGetParam(variant->function, 1);
633 y = LLVMGetParam(variant->function, 2);
634 a0_ptr = LLVMGetParam(variant->function, 3);
635 dadx_ptr = LLVMGetParam(variant->function, 4);
636 dady_ptr = LLVMGetParam(variant->function, 5);
637 color_ptr = LLVMGetParam(variant->function, 6);
638 depth_ptr = LLVMGetParam(variant->function, 7);
639 c0 = LLVMGetParam(variant->function, 8);
640 c1 = LLVMGetParam(variant->function, 9);
641 c2 = LLVMGetParam(variant->function, 10);
642 step0_ptr = LLVMGetParam(variant->function, 11);
643 step1_ptr = LLVMGetParam(variant->function, 12);
644 step2_ptr = LLVMGetParam(variant->function, 13);
646 lp_build_name(context_ptr, "context");
647 lp_build_name(x, "x");
648 lp_build_name(y, "y");
649 lp_build_name(a0_ptr, "a0");
650 lp_build_name(dadx_ptr, "dadx");
651 lp_build_name(dady_ptr, "dady");
652 lp_build_name(color_ptr, "color");
653 lp_build_name(depth_ptr, "depth");
654 lp_build_name(c0, "c0");
655 lp_build_name(c1, "c1");
656 lp_build_name(c2, "c2");
657 lp_build_name(step0_ptr, "step0");
658 lp_build_name(step1_ptr, "step1");
659 lp_build_name(step2_ptr, "step2");
665 block = LLVMAppendBasicBlock(variant->function, "entry");
666 builder = LLVMCreateBuilder();
667 LLVMPositionBuilderAtEnd(builder, block);
669 generate_pos0(builder, x, y, &x0, &y0);
671 lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
672 a0_ptr, dadx_ptr, dady_ptr,
675 /* code generated texture sampling */
676 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
678 /* loop over quads in the block */
679 for(i = 0; i < num_fs; ++i) {
680 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
681 LLVMValueRef out_color[NUM_CHANNELS];
682 LLVMValueRef depth_ptr_i;
685 lp_build_interp_soa_update(&interp, i);
687 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
689 generate_fs(lp, shader, key,
696 &fs_mask[i], /* output */
700 step0_ptr, step1_ptr, step2_ptr);
702 for(chan = 0; chan < NUM_CHANNELS; ++chan)
703 fs_out_color[chan][i] = out_color[chan];
706 sampler->destroy(sampler);
709 * Convert the fs's output color and mask to fit to the blending type.
712 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
713 lp_build_conv(builder, fs_type, blend_type,
714 fs_out_color[chan], num_fs,
715 &blend_in_color[chan], 1);
716 lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
720 lp_build_conv_mask(builder, fs_type, blend_type,
728 generate_blend(&key->blend,
736 LLVMBuildRetVoid(builder);
738 LLVMDisposeBuilder(builder);
741 * Translate the LLVM IR into machine code.
745 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
746 LLVMDumpValue(variant->function);
751 LLVMRunFunctionPassManager(screen->pass, variant->function);
753 if (LP_DEBUG & DEBUG_JIT) {
754 LLVMDumpValue(variant->function);
758 variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
760 if (LP_DEBUG & DEBUG_ASM)
761 lp_disassemble(variant->jit_function);
763 variant->next = shader->variants;
764 shader->variants = variant;
771 llvmpipe_create_fs_state(struct pipe_context *pipe,
772 const struct pipe_shader_state *templ)
774 struct lp_fragment_shader *shader;
776 shader = CALLOC_STRUCT(lp_fragment_shader);
780 /* get/save the summary info for this shader */
781 tgsi_scan_shader(templ->tokens, &shader->info);
783 /* we need to keep a local copy of the tokens */
784 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
791 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
793 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
795 if (llvmpipe->fs == fs)
798 draw_flush(llvmpipe->draw);
802 llvmpipe->dirty |= LP_NEW_FS;
807 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
809 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
810 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
811 struct lp_fragment_shader *shader = fs;
812 struct lp_fragment_shader_variant *variant;
814 assert(fs != llvmpipe->fs);
817 variant = shader->variants;
819 struct lp_fragment_shader_variant *next = variant->next;
821 if(variant->function) {
822 if(variant->jit_function)
823 LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
824 LLVMDeleteFunction(variant->function);
832 FREE((void *) shader->base.tokens);
839 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
840 uint shader, uint index,
841 const struct pipe_constant_buffer *constants)
843 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
844 struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
845 unsigned size = buffer ? buffer->size : 0;
846 const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
848 assert(shader < PIPE_SHADER_TYPES);
851 if(llvmpipe->constants[shader].buffer == buffer)
854 draw_flush(llvmpipe->draw);
856 /* note: reference counting */
857 pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
859 if(shader == PIPE_SHADER_VERTEX) {
860 draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
864 llvmpipe->dirty |= LP_NEW_CONSTANTS;
869 * We need to generate several variants of the fragment pipeline to match
870 * all the combinations of the contributing state atoms.
872 * TODO: there is actually no reason to tie this to context state -- the
873 * generated code could be cached globally in the screen.
876 make_variant_key(struct llvmpipe_context *lp,
877 struct lp_fragment_shader *shader,
878 struct lp_fragment_shader_variant_key *key)
882 memset(key, 0, sizeof *key);
884 if(lp->framebuffer.zsbuf &&
885 lp->depth_stencil->depth.enabled) {
886 key->zsbuf_format = lp->framebuffer.zsbuf->format;
887 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
890 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
891 if(key->alpha.enabled)
892 key->alpha.func = lp->depth_stencil->alpha.func;
893 /* alpha.ref_value is passed in jit_context */
895 if(lp->framebuffer.cbufs[0]) {
896 const struct util_format_description *format_desc;
899 memcpy(&key->blend, lp->blend, sizeof key->blend);
901 format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
902 assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
903 format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
905 /* mask out color channels not present in the color buffer */
906 for(chan = 0; chan < 4; ++chan) {
907 enum util_format_swizzle swizzle = format_desc->swizzle[chan];
909 key->blend.colormask &= ~(1 << chan);
913 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
914 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
915 lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
920 llvmpipe_update_fs(struct llvmpipe_context *lp)
922 struct lp_fragment_shader *shader = lp->fs;
923 struct lp_fragment_shader_variant_key key;
924 struct lp_fragment_shader_variant *variant;
926 make_variant_key(lp, shader, &key);
928 variant = shader->variants;
930 if(memcmp(&variant->key, &key, sizeof key) == 0)
933 variant = variant->next;
937 variant = generate_fragment(lp, shader, &key);
939 shader->current = variant;
941 lp_setup_set_fs_function(lp->setup,
942 shader->current->jit_function);