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:
34 * - triangle edge in/out testing
40 * - depth/stencil test (stencil TBI)
43 * This file has only the glue to assemble the fragment pipeline. The actual
44 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
45 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
46 * muster the LLVM JIT execution engine to create a function that follows an
47 * established binary interface and that can be called from C directly.
49 * A big source of complexity here is that we often want to run different
50 * stages with different precisions and data types and precisions. For example,
51 * the fragment shader needs typically to be done in floats, but the
52 * depth/stencil test and blending is better done in the type that most closely
53 * matches the depth/stencil and color buffer respectively.
55 * Since the width of a SIMD vector register stays the same regardless of the
56 * element type, different types imply different number of elements, so we must
57 * code generate more instances of the stages with larger types to be able to
58 * feed/consume the stages with smaller types.
60 * @author Jose Fonseca <jfonseca@vmware.com>
64 #include "pipe/p_defines.h"
65 #include "util/u_inlines.h"
66 #include "util/u_memory.h"
67 #include "util/u_format.h"
68 #include "util/u_dump.h"
69 #include "os/os_time.h"
70 #include "pipe/p_shader_tokens.h"
71 #include "draw/draw_context.h"
72 #include "tgsi/tgsi_dump.h"
73 #include "tgsi/tgsi_scan.h"
74 #include "tgsi/tgsi_parse.h"
75 #include "gallivm/lp_bld_type.h"
76 #include "gallivm/lp_bld_const.h"
77 #include "gallivm/lp_bld_conv.h"
78 #include "gallivm/lp_bld_intr.h"
79 #include "gallivm/lp_bld_logic.h"
80 #include "gallivm/lp_bld_depth.h"
81 #include "gallivm/lp_bld_interp.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_alpha.h"
84 #include "gallivm/lp_bld_blend.h"
85 #include "gallivm/lp_bld_swizzle.h"
86 #include "gallivm/lp_bld_flow.h"
87 #include "gallivm/lp_bld_debug.h"
88 #include "lp_buffer.h"
89 #include "lp_context.h"
92 #include "lp_screen.h"
95 #include "lp_tex_sample.h"
98 #include <llvm-c/Analysis.h>
101 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
102 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
106 * Derive from the quad's upper left scalar coordinates the coordinates for
107 * all other quad pixels
110 generate_pos0(LLVMBuilderRef builder,
116 LLVMTypeRef int_elem_type = LLVMInt32Type();
117 LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
118 LLVMTypeRef elem_type = LLVMFloatType();
119 LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
120 LLVMValueRef x_offsets[QUAD_SIZE];
121 LLVMValueRef y_offsets[QUAD_SIZE];
124 x = lp_build_broadcast(builder, int_vec_type, x);
125 y = lp_build_broadcast(builder, int_vec_type, y);
127 for(i = 0; i < QUAD_SIZE; ++i) {
128 x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
129 y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
132 x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
133 y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
135 *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
136 *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
141 * Generate the depth /stencil test code.
144 generate_depth_stencil(LLVMBuilderRef builder,
145 const struct lp_fragment_shader_variant_key *key,
146 struct lp_type src_type,
147 struct lp_build_mask_context *mask,
148 LLVMValueRef stencil_refs[2],
150 LLVMValueRef dst_ptr,
153 const struct util_format_description *format_desc;
154 struct lp_type dst_type;
156 if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
159 format_desc = util_format_description(key->zsbuf_format);
163 * Depths are expected to be between 0 and 1, even if they are stored in
164 * floats. Setting these bits here will ensure that the lp_build_conv() call
165 * below won't try to unnecessarily clamp the incoming values.
167 if(src_type.floating) {
168 src_type.sign = FALSE;
169 src_type.norm = TRUE;
172 assert(!src_type.sign);
173 assert(src_type.norm);
176 /* Pick the depth type. */
177 dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
179 /* FIXME: Cope with a depth test type with a different bit width. */
180 assert(dst_type.width == src_type.width);
181 assert(dst_type.length == src_type.length);
183 /* Convert fragment Z from float to integer */
184 lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
186 dst_ptr = LLVMBuildBitCast(builder,
188 LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
189 lp_build_depth_stencil_test(builder,
203 * Generate the code to do inside/outside triangle testing for the
204 * four pixels in a 2x2 quad. This will set the four elements of the
205 * quad mask vector to 0 or ~0.
206 * \param i which quad of the quad group to test, in [0,3]
209 generate_tri_edge_mask(LLVMBuilderRef builder,
211 LLVMValueRef *mask, /* ivec4, out */
212 LLVMValueRef c0, /* int32 */
213 LLVMValueRef c1, /* int32 */
214 LLVMValueRef c2, /* int32 */
215 LLVMValueRef step0_ptr, /* ivec4 */
216 LLVMValueRef step1_ptr, /* ivec4 */
217 LLVMValueRef step2_ptr) /* ivec4 */
219 #define OPTIMIZE_IN_OUT_TEST 0
220 #if OPTIMIZE_IN_OUT_TEST
221 struct lp_build_if_state ifctx;
222 LLVMValueRef not_draw_all;
224 struct lp_build_flow_context *flow;
225 struct lp_type i32_type;
226 LLVMTypeRef i32vec4_type, mask_type;
227 LLVMValueRef c0_vec, c1_vec, c2_vec;
228 LLVMValueRef in_out_mask;
232 /* int32 vector type */
233 memset(&i32_type, 0, sizeof i32_type);
234 i32_type.floating = FALSE; /* values are integers */
235 i32_type.sign = TRUE; /* values are signed */
236 i32_type.norm = FALSE; /* values are not normalized */
237 i32_type.width = 32; /* 32-bit int values */
238 i32_type.length = 4; /* 4 elements per vector */
240 i32vec4_type = lp_build_int32_vec4_type();
242 mask_type = LLVMIntType(32 * 4);
245 * Use a conditional here to do detailed pixel in/out testing.
246 * We only have to do this if c0 != INT_MIN.
248 flow = lp_build_flow_create(builder);
249 lp_build_flow_scope_begin(flow);
252 #if OPTIMIZE_IN_OUT_TEST
253 /* not_draw_all = (c0 != INT_MIN) */
254 not_draw_all = LLVMBuildICmp(builder,
257 LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
260 in_out_mask = lp_build_const_int_vec(i32_type, ~0);
263 lp_build_flow_scope_declare(flow, &in_out_mask);
265 /* if (not_draw_all) {... */
266 lp_build_if(&ifctx, flow, builder, not_draw_all);
269 LLVMValueRef step0_vec, step1_vec, step2_vec;
270 LLVMValueRef m0_vec, m1_vec, m2_vec;
271 LLVMValueRef index, m;
273 /* c0_vec = {c0, c0, c0, c0}
274 * Note that we emit this code four times but LLVM optimizes away
275 * three instances of it.
277 c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
278 c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
279 c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
280 lp_build_name(c0_vec, "edgeconst0vec");
281 lp_build_name(c1_vec, "edgeconst1vec");
282 lp_build_name(c2_vec, "edgeconst2vec");
284 /* load step0vec, step1, step2 vec from memory */
285 index = LLVMConstInt(LLVMInt32Type(), i, 0);
286 step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
287 step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
288 step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
289 lp_build_name(step0_vec, "step0vec");
290 lp_build_name(step1_vec, "step1vec");
291 lp_build_name(step2_vec, "step2vec");
293 /* m0_vec = step0_ptr[i] > c0_vec */
294 m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
295 m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
296 m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
298 /* in_out_mask = m0_vec & m1_vec & m2_vec */
299 m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
300 in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
301 lp_build_name(in_out_mask, "inoutmaskvec");
303 #if OPTIMIZE_IN_OUT_TEST
304 lp_build_endif(&ifctx);
308 lp_build_flow_scope_end(flow);
309 lp_build_flow_destroy(flow);
311 /* This is the initial alive/dead pixel mask for a quad of four pixels.
312 * It's an int[4] vector with each word set to 0 or ~0.
313 * Words will get cleared when pixels faile the Z test, etc.
320 generate_scissor_test(LLVMBuilderRef builder,
321 LLVMValueRef context_ptr,
322 const struct lp_build_interp_soa_context *interp,
325 LLVMTypeRef vec_type = lp_build_vec_type(type);
326 LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
327 LLVMValueRef xmin, ymin, xmax, ymax;
328 LLVMValueRef m0, m1, m2, m3, m;
330 /* xpos, ypos contain the window coords for the four pixels in the quad */
334 /* get the current scissor bounds, convert to vectors */
335 xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
336 xmin = lp_build_broadcast(builder, vec_type, xmin);
338 ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
339 ymin = lp_build_broadcast(builder, vec_type, ymin);
341 xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
342 xmax = lp_build_broadcast(builder, vec_type, xmax);
344 ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
345 ymax = lp_build_broadcast(builder, vec_type, ymax);
347 /* compare the fragment's position coordinates against the scissor bounds */
348 m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
349 m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
350 m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
351 m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
353 /* AND all the masks together */
354 m = LLVMBuildAnd(builder, m0, m1, "");
355 m = LLVMBuildAnd(builder, m, m2, "");
356 m = LLVMBuildAnd(builder, m, m3, "");
358 lp_build_name(m, "scissormask");
365 build_int32_vec_const(int value)
367 struct lp_type i32_type;
369 memset(&i32_type, 0, sizeof i32_type);
370 i32_type.floating = FALSE; /* values are integers */
371 i32_type.sign = TRUE; /* values are signed */
372 i32_type.norm = FALSE; /* values are not normalized */
373 i32_type.width = 32; /* 32-bit int values */
374 i32_type.length = 4; /* 4 elements per vector */
375 return lp_build_const_int_vec(i32_type, value);
381 * Generate the fragment shader, depth/stencil test, and alpha tests.
382 * \param i which quad in the tile, in range [0,3]
383 * \param do_tri_test if 1, do triangle edge in/out testing
386 generate_fs(struct llvmpipe_context *lp,
387 struct lp_fragment_shader *shader,
388 const struct lp_fragment_shader_variant_key *key,
389 LLVMBuilderRef builder,
391 LLVMValueRef context_ptr,
393 const struct lp_build_interp_soa_context *interp,
394 struct lp_build_sampler_soa *sampler,
396 LLVMValueRef (*color)[4],
397 LLVMValueRef depth_ptr,
399 unsigned do_tri_test,
403 LLVMValueRef step0_ptr,
404 LLVMValueRef step1_ptr,
405 LLVMValueRef step2_ptr)
407 const struct tgsi_token *tokens = shader->base.tokens;
408 LLVMTypeRef elem_type;
409 LLVMTypeRef vec_type;
410 LLVMTypeRef int_vec_type;
411 LLVMValueRef consts_ptr;
412 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
413 LLVMValueRef z = interp->pos[2];
414 LLVMValueRef stencil_refs[2];
415 struct lp_build_flow_context *flow;
416 struct lp_build_mask_context mask;
417 boolean early_depth_stencil_test;
424 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
425 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
427 elem_type = lp_build_elem_type(type);
428 vec_type = lp_build_vec_type(type);
429 int_vec_type = lp_build_int_vec_type(type);
431 consts_ptr = lp_jit_context_constants(builder, context_ptr);
433 flow = lp_build_flow_create(builder);
435 memset(outputs, 0, sizeof outputs);
437 lp_build_flow_scope_begin(flow);
439 /* Declare the color and z variables */
440 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
441 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
442 color[cbuf][chan] = LLVMGetUndef(vec_type);
443 lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
446 lp_build_flow_scope_declare(flow, &z);
448 /* do triangle edge testing */
450 generate_tri_edge_mask(builder, i, pmask,
451 c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
454 *pmask = build_int32_vec_const(~0);
457 /* 'mask' will control execution based on quad's pixel alive/killed state */
458 lp_build_mask_begin(&mask, flow, type, *pmask);
462 generate_scissor_test(builder, context_ptr, interp, type);
463 lp_build_mask_update(&mask, smask);
466 early_depth_stencil_test =
467 (key->depth.enabled || key->stencil[0].enabled) &&
468 !key->alpha.enabled &&
469 !shader->info.uses_kill &&
470 !shader->info.writes_z;
472 if (early_depth_stencil_test)
473 generate_depth_stencil(builder, key,
475 stencil_refs, z, depth_ptr, facing);
477 lp_build_tgsi_soa(builder, tokens, type, &mask,
478 consts_ptr, interp->pos, interp->inputs,
481 for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
482 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
483 if(outputs[attrib][chan]) {
484 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
485 lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
487 switch (shader->info.output_semantic_name[attrib]) {
488 case TGSI_SEMANTIC_COLOR:
490 unsigned cbuf = shader->info.output_semantic_index[attrib];
492 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
495 /* XXX: should the alpha reference value be passed separately? */
496 /* XXX: should only test the final assignment to alpha */
497 if(cbuf == 0 && chan == 3) {
498 LLVMValueRef alpha = out;
499 LLVMValueRef alpha_ref_value;
500 alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
501 alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
502 lp_build_alpha_test(builder, &key->alpha, type,
503 &mask, alpha, alpha_ref_value);
506 color[cbuf][chan] = out;
510 case TGSI_SEMANTIC_POSITION:
519 if (!early_depth_stencil_test)
520 generate_depth_stencil(builder, key,
522 stencil_refs, z, depth_ptr, facing);
524 lp_build_mask_end(&mask);
526 lp_build_flow_scope_end(flow);
528 lp_build_flow_destroy(flow);
536 * Generate color blending and color output.
539 generate_blend(const struct pipe_blend_state *blend,
540 LLVMBuilderRef builder,
542 LLVMValueRef context_ptr,
545 LLVMValueRef dst_ptr)
547 struct lp_build_context bld;
548 struct lp_build_flow_context *flow;
549 struct lp_build_mask_context mask_ctx;
550 LLVMTypeRef vec_type;
551 LLVMTypeRef int_vec_type;
552 LLVMValueRef const_ptr;
558 lp_build_context_init(&bld, builder, type);
560 flow = lp_build_flow_create(builder);
562 /* we'll use this mask context to skip blending if all pixels are dead */
563 lp_build_mask_begin(&mask_ctx, flow, type, mask);
565 vec_type = lp_build_vec_type(type);
566 int_vec_type = lp_build_int_vec_type(type);
568 const_ptr = lp_jit_context_blend_color(builder, context_ptr);
569 const_ptr = LLVMBuildBitCast(builder, const_ptr,
570 LLVMPointerType(vec_type, 0), "");
572 for(chan = 0; chan < 4; ++chan) {
573 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
574 con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
576 dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
578 lp_build_name(con[chan], "con.%c", "rgba"[chan]);
579 lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
582 lp_build_blend_soa(builder, blend, type, src, dst, con, res);
584 for(chan = 0; chan < 4; ++chan) {
585 if(blend->rt[0].colormask & (1 << chan)) {
586 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
587 lp_build_name(res[chan], "res.%c", "rgba"[chan]);
588 res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
589 LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
593 lp_build_mask_end(&mask_ctx);
594 lp_build_flow_destroy(flow);
598 /** casting function to avoid compiler warnings */
599 static lp_jit_frag_func
600 cast_voidptr_to_lp_jit_frag_func(void *p)
606 assert(sizeof(tmp.v) == sizeof(tmp.f));
613 * Generate the runtime callable function for the whole fragment pipeline.
614 * Note that the function which we generate operates on a block of 16
615 * pixels at at time. The block contains 2x2 quads. Each quad contains
619 generate_fragment(struct llvmpipe_context *lp,
620 struct lp_fragment_shader *shader,
621 struct lp_fragment_shader_variant *variant,
622 unsigned do_tri_test)
624 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
625 const struct lp_fragment_shader_variant_key *key = &variant->key;
626 struct lp_type fs_type;
627 struct lp_type blend_type;
628 LLVMTypeRef fs_elem_type;
629 LLVMTypeRef fs_vec_type;
630 LLVMTypeRef fs_int_vec_type;
631 LLVMTypeRef blend_vec_type;
632 LLVMTypeRef blend_int_vec_type;
633 LLVMTypeRef arg_types[15];
634 LLVMTypeRef func_type;
635 LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
636 LLVMValueRef context_ptr;
640 LLVMValueRef dadx_ptr;
641 LLVMValueRef dady_ptr;
642 LLVMValueRef color_ptr_ptr;
643 LLVMValueRef depth_ptr;
644 LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
645 LLVMBasicBlockRef block;
646 LLVMBuilderRef builder;
649 struct lp_build_sampler_soa *sampler;
650 struct lp_build_interp_soa_context interp;
651 LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
652 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
653 LLVMValueRef blend_mask;
654 LLVMValueRef blend_in_color[NUM_CHANNELS];
655 LLVMValueRef function;
663 /* TODO: actually pick these based on the fs and color buffer
664 * characteristics. */
666 memset(&fs_type, 0, sizeof fs_type);
667 fs_type.floating = TRUE; /* floating point values */
668 fs_type.sign = TRUE; /* values are signed */
669 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
670 fs_type.width = 32; /* 32-bit float */
671 fs_type.length = 4; /* 4 elements per vector */
672 num_fs = 4; /* number of quads per block */
674 memset(&blend_type, 0, sizeof blend_type);
675 blend_type.floating = FALSE; /* values are integers */
676 blend_type.sign = FALSE; /* values are unsigned */
677 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
678 blend_type.width = 8; /* 8-bit ubyte values */
679 blend_type.length = 16; /* 16 elements per vector */
682 * Generate the function prototype. Any change here must be reflected in
683 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
686 fs_elem_type = lp_build_elem_type(fs_type);
687 fs_vec_type = lp_build_vec_type(fs_type);
688 fs_int_vec_type = lp_build_int_vec_type(fs_type);
690 blend_vec_type = lp_build_vec_type(blend_type);
691 blend_int_vec_type = lp_build_int_vec_type(blend_type);
693 arg_types[0] = screen->context_ptr_type; /* context */
694 arg_types[1] = LLVMInt32Type(); /* x */
695 arg_types[2] = LLVMInt32Type(); /* y */
696 arg_types[3] = LLVMFloatType(); /* facing */
697 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
698 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
699 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
700 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
701 arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
702 arg_types[9] = LLVMInt32Type(); /* c0 */
703 arg_types[10] = LLVMInt32Type(); /* c1 */
704 arg_types[11] = LLVMInt32Type(); /* c2 */
705 /* Note: the step arrays are built as int32[16] but we interpret
706 * them here as int32_vec4[4].
708 arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
709 arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
710 arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
712 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
714 function = LLVMAddFunction(screen->module, "shader", func_type);
715 LLVMSetFunctionCallConv(function, LLVMCCallConv);
717 variant->function[do_tri_test] = function;
720 /* XXX: need to propagate noalias down into color param now we are
721 * passing a pointer-to-pointer?
723 for(i = 0; i < Elements(arg_types); ++i)
724 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
725 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
727 context_ptr = LLVMGetParam(function, 0);
728 x = LLVMGetParam(function, 1);
729 y = LLVMGetParam(function, 2);
730 facing = LLVMGetParam(function, 3);
731 a0_ptr = LLVMGetParam(function, 4);
732 dadx_ptr = LLVMGetParam(function, 5);
733 dady_ptr = LLVMGetParam(function, 6);
734 color_ptr_ptr = LLVMGetParam(function, 7);
735 depth_ptr = LLVMGetParam(function, 8);
736 c0 = LLVMGetParam(function, 9);
737 c1 = LLVMGetParam(function, 10);
738 c2 = LLVMGetParam(function, 11);
739 step0_ptr = LLVMGetParam(function, 12);
740 step1_ptr = LLVMGetParam(function, 13);
741 step2_ptr = LLVMGetParam(function, 14);
743 lp_build_name(context_ptr, "context");
744 lp_build_name(x, "x");
745 lp_build_name(y, "y");
746 lp_build_name(a0_ptr, "a0");
747 lp_build_name(dadx_ptr, "dadx");
748 lp_build_name(dady_ptr, "dady");
749 lp_build_name(color_ptr_ptr, "color_ptr");
750 lp_build_name(depth_ptr, "depth");
751 lp_build_name(c0, "c0");
752 lp_build_name(c1, "c1");
753 lp_build_name(c2, "c2");
754 lp_build_name(step0_ptr, "step0");
755 lp_build_name(step1_ptr, "step1");
756 lp_build_name(step2_ptr, "step2");
762 block = LLVMAppendBasicBlock(function, "entry");
763 builder = LLVMCreateBuilder();
764 LLVMPositionBuilderAtEnd(builder, block);
766 generate_pos0(builder, x, y, &x0, &y0);
768 lp_build_interp_soa_init(&interp,
772 a0_ptr, dadx_ptr, dady_ptr,
775 /* code generated texture sampling */
776 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
778 /* loop over quads in the block */
779 for(i = 0; i < num_fs; ++i) {
780 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
781 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
782 LLVMValueRef depth_ptr_i;
786 lp_build_interp_soa_update(&interp, i);
788 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
790 generate_fs(lp, shader, key,
797 &fs_mask[i], /* output */
803 step0_ptr, step1_ptr, step2_ptr);
805 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
806 for(chan = 0; chan < NUM_CHANNELS; ++chan)
807 fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
810 sampler->destroy(sampler);
812 /* Loop over color outputs / color buffers to do blending.
814 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
815 LLVMValueRef color_ptr;
816 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
819 * Convert the fs's output color and mask to fit to the blending type.
821 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
822 lp_build_conv(builder, fs_type, blend_type,
823 fs_out_color[cbuf][chan], num_fs,
824 &blend_in_color[chan], 1);
825 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
828 lp_build_conv_mask(builder, fs_type, blend_type,
832 color_ptr = LLVMBuildLoad(builder,
833 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
835 lp_build_name(color_ptr, "color_ptr%d", cbuf);
840 generate_blend(&key->blend,
849 LLVMBuildRetVoid(builder);
851 LLVMDisposeBuilder(builder);
854 /* Verify the LLVM IR. If invalid, dump and abort */
856 if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
858 LLVMDumpValue(function);
863 /* Apply optimizations to LLVM IR */
865 LLVMRunFunctionPassManager(screen->pass, function);
867 if (LP_DEBUG & DEBUG_JIT) {
868 /* Print the LLVM IR to stderr */
869 LLVMDumpValue(function);
874 * Translate the LLVM IR into machine code.
877 void *f = LLVMGetPointerToGlobal(screen->engine, function);
879 variant->jit_function[do_tri_test] = cast_voidptr_to_lp_jit_frag_func(f);
881 if (LP_DEBUG & DEBUG_ASM)
887 static struct lp_fragment_shader_variant *
888 generate_variant(struct llvmpipe_context *lp,
889 struct lp_fragment_shader *shader,
890 const struct lp_fragment_shader_variant_key *key)
892 struct lp_fragment_shader_variant *variant;
894 if (LP_DEBUG & DEBUG_JIT) {
897 tgsi_dump(shader->base.tokens, 0);
898 if(key->depth.enabled) {
899 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
900 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
901 debug_printf("depth.writemask = %u\n", key->depth.writemask);
903 if(key->alpha.enabled) {
904 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
905 debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
907 if(key->blend.logicop_enable) {
908 debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
910 else if(key->blend.rt[0].blend_enable) {
911 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
912 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
913 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
914 debug_printf("alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
915 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
916 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
918 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
919 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
920 if(key->sampler[i].format) {
921 debug_printf("sampler[%u] = \n", i);
922 debug_printf(" .format = %s\n",
923 util_format_name(key->sampler[i].format));
924 debug_printf(" .target = %s\n",
925 util_dump_tex_target(key->sampler[i].target, TRUE));
926 debug_printf(" .pot = %u %u %u\n",
927 key->sampler[i].pot_width,
928 key->sampler[i].pot_height,
929 key->sampler[i].pot_depth);
930 debug_printf(" .wrap = %s %s %s\n",
931 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
932 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
933 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
934 debug_printf(" .min_img_filter = %s\n",
935 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
936 debug_printf(" .min_mip_filter = %s\n",
937 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
938 debug_printf(" .mag_img_filter = %s\n",
939 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
940 if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
941 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
942 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
947 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
951 variant->shader = shader;
952 memcpy(&variant->key, key, sizeof *key);
954 generate_fragment(lp, shader, variant, 0);
955 generate_fragment(lp, shader, variant, 1);
957 /* insert new variant into linked list */
958 variant->next = shader->variants;
959 shader->variants = variant;
966 llvmpipe_create_fs_state(struct pipe_context *pipe,
967 const struct pipe_shader_state *templ)
969 struct lp_fragment_shader *shader;
971 shader = CALLOC_STRUCT(lp_fragment_shader);
975 /* get/save the summary info for this shader */
976 tgsi_scan_shader(templ->tokens, &shader->info);
978 /* we need to keep a local copy of the tokens */
979 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
986 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
988 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
990 if (llvmpipe->fs == fs)
993 draw_flush(llvmpipe->draw);
997 llvmpipe->dirty |= LP_NEW_FS;
1002 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1004 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1005 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
1006 struct lp_fragment_shader *shader = fs;
1007 struct lp_fragment_shader_variant *variant;
1009 assert(fs != llvmpipe->fs);
1013 * XXX: we need to flush the context until we have some sort of reference
1014 * counting in fragment shaders as they may still be binned
1016 draw_flush(llvmpipe->draw);
1017 lp_setup_flush(llvmpipe->setup, 0);
1019 variant = shader->variants;
1021 struct lp_fragment_shader_variant *next = variant->next;
1024 for (i = 0; i < Elements(variant->function); i++) {
1025 if (variant->function[i]) {
1026 if (variant->jit_function[i])
1027 LLVMFreeMachineCodeForFunction(screen->engine,
1028 variant->function[i]);
1029 LLVMDeleteFunction(variant->function[i]);
1038 FREE((void *) shader->base.tokens);
1045 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1046 uint shader, uint index,
1047 struct pipe_buffer *constants)
1049 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1050 unsigned size = constants ? constants->size : 0;
1051 const void *data = constants ? llvmpipe_buffer(constants)->data : NULL;
1053 assert(shader < PIPE_SHADER_TYPES);
1056 if(llvmpipe->constants[shader] == constants)
1059 draw_flush(llvmpipe->draw);
1061 /* note: reference counting */
1062 pipe_buffer_reference(&llvmpipe->constants[shader], constants);
1064 if(shader == PIPE_SHADER_VERTEX) {
1065 draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, 0,
1069 llvmpipe->dirty |= LP_NEW_CONSTANTS;
1074 * We need to generate several variants of the fragment pipeline to match
1075 * all the combinations of the contributing state atoms.
1077 * TODO: there is actually no reason to tie this to context state -- the
1078 * generated code could be cached globally in the screen.
1081 make_variant_key(struct llvmpipe_context *lp,
1082 struct lp_fragment_shader *shader,
1083 struct lp_fragment_shader_variant_key *key)
1087 memset(key, 0, sizeof *key);
1089 if (lp->framebuffer.zsbuf) {
1090 if (lp->depth_stencil->depth.enabled) {
1091 key->zsbuf_format = lp->framebuffer.zsbuf->format;
1092 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1094 if (lp->depth_stencil->stencil[0].enabled) {
1095 key->zsbuf_format = lp->framebuffer.zsbuf->format;
1096 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1100 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1101 if(key->alpha.enabled)
1102 key->alpha.func = lp->depth_stencil->alpha.func;
1103 /* alpha.ref_value is passed in jit_context */
1105 key->flatshade = lp->rasterizer->flatshade;
1106 key->scissor = lp->rasterizer->scissor;
1108 if (lp->framebuffer.nr_cbufs) {
1109 memcpy(&key->blend, lp->blend, sizeof key->blend);
1112 key->nr_cbufs = lp->framebuffer.nr_cbufs;
1113 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1114 const struct util_format_description *format_desc;
1117 format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1118 assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
1119 format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
1121 /* mask out color channels not present in the color buffer.
1122 * Should be simple to incorporate per-cbuf writemasks:
1124 for(chan = 0; chan < 4; ++chan) {
1125 enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1127 if(swizzle <= UTIL_FORMAT_SWIZZLE_W)
1128 key->blend.rt[0].colormask |= (1 << chan);
1132 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1133 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1134 lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i]->texture, lp->sampler[i]);
1139 * Update fragment state. This is called just prior to drawing
1140 * something when some fragment-related state has changed.
1143 llvmpipe_update_fs(struct llvmpipe_context *lp)
1145 struct lp_fragment_shader *shader = lp->fs;
1146 struct lp_fragment_shader_variant_key key;
1147 struct lp_fragment_shader_variant *variant;
1150 make_variant_key(lp, shader, &key);
1152 variant = shader->variants;
1154 if(memcmp(&variant->key, &key, sizeof key) == 0)
1157 variant = variant->next;
1165 variant = generate_variant(lp, shader, &key);
1169 LP_COUNT_ADD(llvm_compile_time, dt);
1170 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
1173 shader->current = variant;
1175 /* TODO: put this in the variant */
1176 /* TODO: most of these can be relaxed, in particular the colormask */
1177 opaque = !key.blend.logicop_enable &&
1178 !key.blend.rt[0].blend_enable &&
1179 key.blend.rt[0].colormask == 0xf &&
1180 !key.stencil[0].enabled &&
1181 !key.alpha.enabled &&
1182 !key.depth.enabled &&
1184 !shader->info.uses_kill
1187 lp_setup_set_fs_functions(lp->setup,
1188 shader->current->jit_function[RAST_WHOLE],
1189 shader->current->jit_function[RAST_EDGE_TEST],