projects / mesa.git / blob
? search:


fcb6944b5ae598199d4f8e2e9c1babd23f921fa1
[mesa.git] / src / gallium / drivers / radeonsi / si_compute_prim_discard.c
1 /*
2 * Copyright 2019 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26 #include "si_pipe.h"
27 #include "si_shader_internal.h"
28 #include "sid.h"
29 #include "si_build_pm4.h"
30 #include "ac_llvm_cull.h"
31
32 #include "util/u_prim.h"
33 #include "util/u_suballoc.h"
34 #include "util/u_upload_mgr.h"
35 #include "util/fast_idiv_by_const.h"
36
37 /* Based on:
38 * https://frostbite-wp-prd.s3.amazonaws.com/wp-content/uploads/2016/03/29204330/GDC_2016_Compute.pdf
39 */
40
41 /* This file implements primitive culling using asynchronous compute.
42 * It's written to be GL conformant.
43 *
44 * It takes a monolithic VS in LLVM IR returning gl_Position and invokes it
45 * in a compute shader. The shader processes 1 primitive/thread by invoking
46 * the VS for each vertex to get the positions, decomposes strips and fans
47 * into triangles (if needed), eliminates primitive restart (if needed),
48 * does (W<0) culling, face culling, view XY culling, zero-area and
49 * small-primitive culling, and generates a new index buffer that doesn't
50 * contain culled primitives.
51 *
52 * The index buffer is generated using the Ordered Count feature of GDS,
53 * which is an atomic counter that is incremented in the wavefront launch
54 * order, so that the original primitive order is preserved.
55 *
56 * Another GDS ordered counter is used to eliminate primitive restart indices.
57 * If a restart index lands on an even thread ID, the compute shader has to flip
58 * the primitive orientation of the whole following triangle strip. The primitive
59 * orientation has to be correct after strip and fan decomposition for two-sided
60 * shading to behave correctly. The decomposition also needs to be aware of
61 * which vertex is the provoking vertex for flat shading to behave correctly.
62 *
63 * IB = a GPU command buffer
64 *
65 * Both the compute and gfx IBs run in parallel sort of like CE and DE.
66 * The gfx IB has a CP barrier (REWIND packet) before a draw packet. REWIND
67 * doesn't continue if its word isn't 0x80000000. Once compute shaders are
68 * finished culling, the last wave will write the final primitive count from
69 * GDS directly into the count word of the draw packet in the gfx IB, and
70 * a CS_DONE event will signal the REWIND packet to continue. It's really
71 * a direct draw with command buffer patching from the compute queue.
72 *
73 * The compute IB doesn't have to start when its corresponding gfx IB starts,
74 * but can start sooner. The compute IB is signaled to start after the last
75 * execution barrier in the *previous* gfx IB. This is handled as follows.
76 * The kernel GPU scheduler starts the compute IB after the previous gfx IB has
77 * started. The compute IB then waits (WAIT_REG_MEM) for a mid-IB fence that
78 * represents the barrier in the previous gfx IB.
79 *
80 * Features:
81 * - Triangle strips and fans are decomposed into an indexed triangle list.
82 * The decomposition differs based on the provoking vertex state.
83 * - Instanced draws are converted into non-instanced draws for 16-bit indices.
84 * (InstanceID is stored in the high bits of VertexID and unpacked by VS)
85 * - Primitive restart is fully supported with triangle strips, including
86 * correct primitive orientation across multiple waves. (restart indices
87 * reset primitive orientation)
88 * - W<0 culling (W<0 is behind the viewer, sort of like near Z culling).
89 * - Back face culling, incl. culling zero-area / degenerate primitives.
90 * - View XY culling.
91 * - View Z culling (disabled due to limited impact with perspective projection).
92 * - Small primitive culling for all MSAA modes and all quant modes.
93 *
94 * The following are not implemented:
95 * - ClipVertex/ClipDistance/CullDistance-based culling.
96 * - Scissor culling.
97 * - HiZ culling.
98 *
99 * Limitations (and unimplemented features that may be possible to implement):
100 * - Only triangles, triangle strips, and triangle fans are supported.
101 * - Primitive restart is only supported with triangle strips.
102 * - Instancing and primitive restart can't be used together.
103 * - Instancing is only supported with 16-bit indices and instance count <= 2^16.
104 * - The instance divisor buffer is unavailable, so all divisors must be
105 * either 0 or 1.
106 * - Multidraws where the vertex shader reads gl_DrawID are unsupported.
107 * - No support for tessellation and geometry shaders.
108 * (patch elimination where tess factors are 0 would be possible to implement)
109 * - The vertex shader must not contain memory stores.
110 * - All VS resources must not have a write usage in the command buffer.
111 * - Bindless textures and images must not occur in the vertex shader.
112 *
113 * User data SGPR layout:
114 * INDEX_BUFFERS: pointer to constants
115 * 0..3: input index buffer - typed buffer view
116 * 4..7: output index buffer - typed buffer view
117 * 8..11: viewport state - scale.xy, translate.xy
118 * VERTEX_COUNTER: counter address or first primitive ID
119 * - If unordered memory counter: address of "count" in the draw packet
120 * and is incremented atomically by the shader.
121 * - If unordered GDS counter: address of "count" in GDS starting from 0,
122 * must be initialized to 0 before the dispatch.
123 * - If ordered GDS counter: the primitive ID that should reset the vertex
124 * counter to 0 in GDS
125 * LAST_WAVE_PRIM_ID: the primitive ID that should write the final vertex
126 * count to memory if using GDS ordered append
127 * VERTEX_COUNT_ADDR: where the last wave should write the vertex count if
128 * using GDS ordered append
129 * VS.VERTEX_BUFFERS: same value as VS
130 * VS.CONST_AND_SHADER_BUFFERS: same value as VS
131 * VS.SAMPLERS_AND_IMAGES: same value as VS
132 * VS.BASE_VERTEX: same value as VS
133 * VS.START_INSTANCE: same value as VS
134 * NUM_PRIMS_UDIV_MULTIPLIER: For fast 31-bit division by the number of primitives
135 * per instance for instancing.
136 * NUM_PRIMS_UDIV_TERMS:
137 * - Bits [0:4]: "post_shift" for fast 31-bit division for instancing.
138 * - Bits [5:31]: The number of primitives per instance for computing the remainder.
139 * PRIMITIVE_RESTART_INDEX
140 * SMALL_PRIM_CULLING_PRECISION: Scale the primitive bounding box by this number.
141 *
142 *
143 * The code contains 3 codepaths:
144 * - Unordered memory counter (for debugging, random primitive order, no primitive restart)
145 * - Unordered GDS counter (for debugging, random primitive order, no primitive restart)
146 * - Ordered GDS counter (it preserves the primitive order)
147 *
148 * How to test primitive restart (the most complicated part because it needs
149 * to get the primitive orientation right):
150 * Set THREADGROUP_SIZE to 2 to exercise both intra-wave and inter-wave
151 * primitive orientation flips with small draw calls, which is what most tests use.
152 * You can also enable draw call splitting into draw calls with just 2 primitives.
153 */
154
155 /* At least 256 is needed for the fastest wave launch rate from compute queues
156 * due to hw constraints. Nothing in the code needs more than 1 wave/threadgroup. */
157 #define THREADGROUP_SIZE 256 /* high numbers limit available VGPRs */
158 #define THREADGROUPS_PER_CU 1 /* TGs to launch on 1 CU before going onto the next, max 8 */
159 #define MAX_WAVES_PER_SH 0 /* no limit */
160 #define INDEX_STORES_USE_SLC 1 /* don't cache indices if L2 is full */
161 /* Don't cull Z. We already do (W < 0) culling for primitives behind the viewer. */
162 #define CULL_Z 0
163 /* 0 = unordered memory counter, 1 = unordered GDS counter, 2 = ordered GDS counter */
164 #define VERTEX_COUNTER_GDS_MODE 2
165 #define GDS_SIZE_UNORDERED (4 * 1024) /* only for the unordered GDS counter */
166
167 /* Grouping compute dispatches for small draw calls: How many primitives from multiple
168 * draw calls to process by compute before signaling the gfx IB. This reduces the number
169 * of EOP events + REWIND packets, because they decrease performance. */
170 #define PRIMS_PER_BATCH (512 * 1024)
171 /* Draw call splitting at the packet level. This allows signaling the gfx IB
172 * for big draw calls sooner, but doesn't allow context flushes between packets.
173 * Primitive restart is supported. Only implemented for ordered append. */
174 #define SPLIT_PRIMS_PACKET_LEVEL_VALUE PRIMS_PER_BATCH
175 /* If there is not enough ring buffer space for the current IB, split draw calls into
176 * this number of primitives, so that we can flush the context and get free ring space. */
177 #define SPLIT_PRIMS_DRAW_LEVEL PRIMS_PER_BATCH
178
179 /* Derived values. */
180 #define WAVES_PER_TG DIV_ROUND_UP(THREADGROUP_SIZE, 64)
181 #define SPLIT_PRIMS_PACKET_LEVEL (VERTEX_COUNTER_GDS_MODE == 2 ? \
182 SPLIT_PRIMS_PACKET_LEVEL_VALUE : \
183 UINT_MAX & ~(THREADGROUP_SIZE - 1))
184
185 #define REWIND_SIGNAL_BIT 0x80000000
186 /* For emulating the rewind packet on CI. */
187 #define FORCE_REWIND_EMULATION 0
188
189 void si_initialize_prim_discard_tunables(struct si_screen *sscreen,
190 bool is_aux_context,
191 unsigned *prim_discard_vertex_count_threshold,
192 unsigned *index_ring_size_per_ib)
193 {
194 *prim_discard_vertex_count_threshold = UINT_MAX; /* disable */
195
196 if (sscreen->info.chip_class == GFX6 || /* SI support is not implemented */
197 !sscreen->info.has_gds_ordered_append ||
198 sscreen->debug_flags & DBG(NO_PD) ||
199 is_aux_context)
200 return;
201
202 /* TODO: enable this after the GDS kernel memory management is fixed */
203 bool enable_on_pro_graphics_by_default = false;
204
205 if (sscreen->debug_flags & DBG(ALWAYS_PD) ||
206 sscreen->debug_flags & DBG(PD) ||
207 (enable_on_pro_graphics_by_default &&
208 sscreen->info.is_pro_graphics &&
209 (sscreen->info.family == CHIP_BONAIRE ||
210 sscreen->info.family == CHIP_HAWAII ||
211 sscreen->info.family == CHIP_TONGA ||
212 sscreen->info.family == CHIP_FIJI ||
213 sscreen->info.family == CHIP_POLARIS10 ||
214 sscreen->info.family == CHIP_POLARIS11 ||
215 sscreen->info.family == CHIP_VEGA10 ||
216 sscreen->info.family == CHIP_VEGA20))) {
217 *prim_discard_vertex_count_threshold = 6000 * 3; /* 6K triangles */
218
219 if (sscreen->debug_flags & DBG(ALWAYS_PD))
220 *prim_discard_vertex_count_threshold = 0; /* always enable */
221
222 const uint32_t MB = 1024 * 1024;
223 const uint64_t GB = 1024 * 1024 * 1024;
224
225 /* The total size is double this per context.
226 * Greater numbers allow bigger gfx IBs.
227 */
228 if (sscreen->info.vram_size <= 2 * GB)
229 *index_ring_size_per_ib = 64 * MB;
230 else if (sscreen->info.vram_size <= 4 * GB)
231 *index_ring_size_per_ib = 128 * MB;
232 else
233 *index_ring_size_per_ib = 256 * MB;
234 }
235 }
236
237 /* Opcode can be "add" or "swap". */
238 static LLVMValueRef
239 si_build_ds_ordered_op(struct si_shader_context *ctx, const char *opcode,
240 LLVMValueRef m0, LLVMValueRef value, unsigned ordered_count_index,
241 bool release, bool done)
242 {
243 LLVMValueRef args[] = {
244 LLVMBuildIntToPtr(ctx->ac.builder, m0,
245 LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS), ""),
246 value,
247 LLVMConstInt(ctx->ac.i32, LLVMAtomicOrderingMonotonic, 0), /* ordering */
248 ctx->ac.i32_0, /* scope */
249 ctx->ac.i1false, /* volatile */
250 LLVMConstInt(ctx->ac.i32, ordered_count_index, 0),
251 LLVMConstInt(ctx->ac.i1, release, 0),
252 LLVMConstInt(ctx->ac.i1, done, 0),
253 };
254
255 char intrinsic[64];
256 snprintf(intrinsic, sizeof(intrinsic), "llvm.amdgcn.ds.ordered.%s", opcode);
257 return ac_build_intrinsic(&ctx->ac, intrinsic, ctx->ac.i32, args, ARRAY_SIZE(args), 0);
258 }
259
260 static LLVMValueRef si_expand_32bit_pointer(struct si_shader_context *ctx, LLVMValueRef ptr)
261 {
262 uint64_t hi = (uint64_t)ctx->screen->info.address32_hi << 32;
263 ptr = LLVMBuildZExt(ctx->ac.builder, ptr, ctx->ac.i64, "");
264 ptr = LLVMBuildOr(ctx->ac.builder, ptr, LLVMConstInt(ctx->ac.i64, hi, 0), "");
265 return LLVMBuildIntToPtr(ctx->ac.builder, ptr,
266 LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GLOBAL), "");
267 }
268
269 struct si_thread0_section {
270 struct si_shader_context *ctx;
271 LLVMValueRef vgpr_result; /* a VGPR for the value on thread 0. */
272 LLVMValueRef saved_exec;
273 };
274
275 /* Enter a section that only executes on thread 0. */
276 static void si_enter_thread0_section(struct si_shader_context *ctx,
277 struct si_thread0_section *section,
278 LLVMValueRef thread_id)
279 {
280 section->ctx = ctx;
281 section->vgpr_result = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "result0");
282
283 /* This IF has 4 instructions:
284 * v_and_b32_e32 v, 63, v ; get the thread ID
285 * v_cmp_eq_u32_e32 vcc, 0, v ; thread ID == 0
286 * s_and_saveexec_b64 s, vcc
287 * s_cbranch_execz BB0_4
288 *
289 * It could just be s_and_saveexec_b64 s, 1.
290 */
291 ac_build_ifcc(&ctx->ac,
292 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, thread_id,
293 ctx->ac.i32_0, ""), 12601);
294 }
295
296 /* Exit a section that only executes on thread 0 and broadcast the result
297 * to all threads. */
298 static void si_exit_thread0_section(struct si_thread0_section *section,
299 LLVMValueRef *result)
300 {
301 struct si_shader_context *ctx = section->ctx;
302
303 LLVMBuildStore(ctx->ac.builder, *result, section->vgpr_result);
304
305 ac_build_endif(&ctx->ac, 12601);
306
307 /* Broadcast the result from thread 0 to all threads. */
308 *result = ac_build_readlane(&ctx->ac,
309 LLVMBuildLoad(ctx->ac.builder, section->vgpr_result, ""), NULL);
310 }
311
312 void si_build_prim_discard_compute_shader(struct si_shader_context *ctx)
313 {
314 struct si_shader_key *key = &ctx->shader->key;
315 LLVMBuilderRef builder = ctx->ac.builder;
316 LLVMValueRef vs = ctx->main_fn;
317
318 /* Always inline the VS function. */
319 ac_add_function_attr(ctx->ac.context, vs, -1, AC_FUNC_ATTR_ALWAYSINLINE);
320 LLVMSetLinkage(vs, LLVMPrivateLinkage);
321
322 enum ac_arg_type const_desc_type;
323 if (ctx->shader->selector->info.const_buffers_declared == 1 &&
324 ctx->shader->selector->info.shader_buffers_declared == 0)
325 const_desc_type = AC_ARG_CONST_FLOAT_PTR;
326 else
327 const_desc_type = AC_ARG_CONST_DESC_PTR;
328
329 memset(&ctx->args, 0, sizeof(ctx->args));
330
331 struct ac_arg param_index_buffers_and_constants, param_vertex_counter;
332 struct ac_arg param_vb_desc, param_const_desc;
333 struct ac_arg param_base_vertex, param_start_instance;
334 struct ac_arg param_block_id, param_local_id, param_ordered_wave_id;
335 struct ac_arg param_restart_index, param_smallprim_precision;
336 struct ac_arg param_num_prims_udiv_multiplier, param_num_prims_udiv_terms;
337 struct ac_arg param_sampler_desc, param_last_wave_prim_id, param_vertex_count_addr;
338
339 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
340 &param_index_buffers_and_constants);
341 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_vertex_counter);
342 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_last_wave_prim_id);
343 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_vertex_count_addr);
344 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
345 &param_vb_desc);
346 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, const_desc_type,
347 &param_const_desc);
348 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
349 &param_sampler_desc);
350 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_base_vertex);
351 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_start_instance);
352 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_num_prims_udiv_multiplier);
353 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_num_prims_udiv_terms);
354 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_restart_index);
355 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, &param_smallprim_precision);
356
357 /* Block ID and thread ID inputs. */
358 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_block_id);
359 if (VERTEX_COUNTER_GDS_MODE == 2)
360 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &param_ordered_wave_id);
361 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &param_local_id);
362
363 /* Create the compute shader function. */
364 unsigned old_type = ctx->type;
365 ctx->type = PIPE_SHADER_COMPUTE;
366 si_llvm_create_func(ctx, "prim_discard_cs", NULL, 0, THREADGROUP_SIZE);
367 ctx->type = old_type;
368
369 if (VERTEX_COUNTER_GDS_MODE == 1) {
370 ac_llvm_add_target_dep_function_attr(ctx->main_fn, "amdgpu-gds-size",
371 GDS_SIZE_UNORDERED);
372 }
373
374 /* Assemble parameters for VS. */
375 LLVMValueRef vs_params[16];
376 unsigned num_vs_params = 0;
377 unsigned param_vertex_id, param_instance_id;
378
379 vs_params[num_vs_params++] = LLVMGetUndef(LLVMTypeOf(LLVMGetParam(vs, 0))); /* RW_BUFFERS */
380 vs_params[num_vs_params++] = LLVMGetUndef(LLVMTypeOf(LLVMGetParam(vs, 1))); /* BINDLESS */
381 vs_params[num_vs_params++] = ac_get_arg(&ctx->ac, param_const_desc);
382 vs_params[num_vs_params++] = ac_get_arg(&ctx->ac, param_sampler_desc);
383 vs_params[num_vs_params++] = LLVMConstInt(ctx->ac.i32,
384 S_VS_STATE_INDEXED(key->opt.cs_indexed), 0);
385 vs_params[num_vs_params++] = ac_get_arg(&ctx->ac, param_base_vertex);
386 vs_params[num_vs_params++] = ac_get_arg(&ctx->ac, param_start_instance);
387 vs_params[num_vs_params++] = ctx->ac.i32_0; /* DrawID */
388 vs_params[num_vs_params++] = ac_get_arg(&ctx->ac, param_vb_desc);
389
390 vs_params[(param_vertex_id = num_vs_params++)] = NULL; /* VertexID */
391 vs_params[(param_instance_id = num_vs_params++)] = NULL; /* InstanceID */
392 vs_params[num_vs_params++] = ctx->ac.i32_0; /* unused (PrimID) */
393 vs_params[num_vs_params++] = ctx->ac.i32_0; /* unused */
394
395 assert(num_vs_params <= ARRAY_SIZE(vs_params));
396 assert(num_vs_params == LLVMCountParamTypes(LLVMGetElementType(LLVMTypeOf(vs))));
397
398 /* Load descriptors. (load 8 dwords at once) */
399 LLVMValueRef input_indexbuf, output_indexbuf, tmp, desc[8];
400
401 LLVMValueRef index_buffers_and_constants = ac_get_arg(&ctx->ac, param_index_buffers_and_constants);
402 tmp = LLVMBuildPointerCast(builder, index_buffers_and_constants,
403 ac_array_in_const32_addr_space(ctx->ac.v8i32), "");
404 tmp = ac_build_load_to_sgpr(&ctx->ac, tmp, ctx->ac.i32_0);
405
406 for (unsigned i = 0; i < 8; i++)
407 desc[i] = ac_llvm_extract_elem(&ctx->ac, tmp, i);
408
409 input_indexbuf = ac_build_gather_values(&ctx->ac, desc, 4);
410 output_indexbuf = ac_build_gather_values(&ctx->ac, desc + 4, 4);
411
412 /* Compute PrimID and InstanceID. */
413 LLVMValueRef global_thread_id =
414 ac_build_imad(&ctx->ac, ac_get_arg(&ctx->ac, param_block_id),
415 LLVMConstInt(ctx->ac.i32, THREADGROUP_SIZE, 0),
416 ac_get_arg(&ctx->ac, param_local_id));
417 LLVMValueRef prim_id = global_thread_id; /* PrimID within an instance */
418 LLVMValueRef instance_id = ctx->ac.i32_0;
419
420 if (key->opt.cs_instancing) {
421 LLVMValueRef num_prims_udiv_terms =
422 ac_get_arg(&ctx->ac, param_num_prims_udiv_terms);
423 LLVMValueRef num_prims_udiv_multiplier =
424 ac_get_arg(&ctx->ac, param_num_prims_udiv_multiplier);
425 /* Unpack num_prims_udiv_terms. */
426 LLVMValueRef post_shift = LLVMBuildAnd(builder, num_prims_udiv_terms,
427 LLVMConstInt(ctx->ac.i32, 0x1f, 0), "");
428 LLVMValueRef prims_per_instance = LLVMBuildLShr(builder, num_prims_udiv_terms,
429 LLVMConstInt(ctx->ac.i32, 5, 0), "");
430 /* Divide the total prim_id by the number of prims per instance. */
431 instance_id = ac_build_fast_udiv_u31_d_not_one(&ctx->ac, prim_id,
432 num_prims_udiv_multiplier,
433 post_shift);
434 /* Compute the remainder. */
435 prim_id = LLVMBuildSub(builder, prim_id,
436 LLVMBuildMul(builder, instance_id,
437 prims_per_instance, ""), "");
438 }
439
440 /* Generate indices (like a non-indexed draw call). */
441 LLVMValueRef index[4] = {NULL, NULL, NULL, LLVMGetUndef(ctx->ac.i32)};
442 unsigned vertices_per_prim = 3;
443
444 switch (key->opt.cs_prim_type) {
445 case PIPE_PRIM_TRIANGLES:
446 for (unsigned i = 0; i < 3; i++) {
447 index[i] = ac_build_imad(&ctx->ac, prim_id,
448 LLVMConstInt(ctx->ac.i32, 3, 0),
449 LLVMConstInt(ctx->ac.i32, i, 0));
450 }
451 break;
452 case PIPE_PRIM_TRIANGLE_STRIP:
453 for (unsigned i = 0; i < 3; i++) {
454 index[i] = LLVMBuildAdd(builder, prim_id,
455 LLVMConstInt(ctx->ac.i32, i, 0), "");
456 }
457 break;
458 case PIPE_PRIM_TRIANGLE_FAN:
459 /* Vertex 1 is first and vertex 2 is last. This will go to the hw clipper
460 * and rasterizer as a normal triangle, so we need to put the provoking
461 * vertex into the correct index variable and preserve orientation at the same time.
462 * gl_VertexID is preserved, because it's equal to the index.
463 */
464 if (key->opt.cs_provoking_vertex_first) {
465 index[0] = LLVMBuildAdd(builder, prim_id, LLVMConstInt(ctx->ac.i32, 1, 0), "");
466 index[1] = LLVMBuildAdd(builder, prim_id, LLVMConstInt(ctx->ac.i32, 2, 0), "");
467 index[2] = ctx->ac.i32_0;
468 } else {
469 index[0] = ctx->ac.i32_0;
470 index[1] = LLVMBuildAdd(builder, prim_id, LLVMConstInt(ctx->ac.i32, 1, 0), "");
471 index[2] = LLVMBuildAdd(builder, prim_id, LLVMConstInt(ctx->ac.i32, 2, 0), "");
472 }
473 break;
474 default:
475 unreachable("unexpected primitive type");
476 }
477
478 /* Fetch indices. */
479 if (key->opt.cs_indexed) {
480 for (unsigned i = 0; i < 3; i++) {
481 index[i] = ac_build_buffer_load_format(&ctx->ac, input_indexbuf,
482 index[i], ctx->ac.i32_0, 1,
483 0, true);
484 index[i] = ac_to_integer(&ctx->ac, index[i]);
485 }
486 }
487
488 LLVMValueRef ordered_wave_id = ac_get_arg(&ctx->ac, param_ordered_wave_id);
489
490 /* Extract the ordered wave ID. */
491 if (VERTEX_COUNTER_GDS_MODE == 2) {
492 ordered_wave_id = LLVMBuildLShr(builder, ordered_wave_id,
493 LLVMConstInt(ctx->ac.i32, 6, 0), "");
494 ordered_wave_id = LLVMBuildAnd(builder, ordered_wave_id,
495 LLVMConstInt(ctx->ac.i32, 0xfff, 0), "");
496 }
497 LLVMValueRef thread_id =
498 LLVMBuildAnd(builder, ac_get_arg(&ctx->ac, param_local_id),
499 LLVMConstInt(ctx->ac.i32, 63, 0), "");
500
501 /* Every other triangle in a strip has a reversed vertex order, so we
502 * need to swap vertices of odd primitives to get the correct primitive
503 * orientation when converting triangle strips to triangles. Primitive
504 * restart complicates it, because a strip can start anywhere.
505 */
506 LLVMValueRef prim_restart_accepted = ctx->ac.i1true;
507 LLVMValueRef vertex_counter = ac_get_arg(&ctx->ac, param_vertex_counter);
508
509 if (key->opt.cs_prim_type == PIPE_PRIM_TRIANGLE_STRIP) {
510 /* Without primitive restart, odd primitives have reversed orientation.
511 * Only primitive restart can flip it with respect to the first vertex
512 * of the draw call.
513 */
514 LLVMValueRef first_is_odd = ctx->ac.i1false;
515
516 /* Handle primitive restart. */
517 if (key->opt.cs_primitive_restart) {
518 /* Get the GDS primitive restart continue flag and clear
519 * the flag in vertex_counter. This flag is used when the draw
520 * call was split and we need to load the primitive orientation
521 * flag from GDS for the first wave too.
522 */
523 LLVMValueRef gds_prim_restart_continue =
524 LLVMBuildLShr(builder, vertex_counter,
525 LLVMConstInt(ctx->ac.i32, 31, 0), "");
526 gds_prim_restart_continue =
527 LLVMBuildTrunc(builder, gds_prim_restart_continue, ctx->ac.i1, "");
528 vertex_counter = LLVMBuildAnd(builder, vertex_counter,
529 LLVMConstInt(ctx->ac.i32, 0x7fffffff, 0), "");
530
531 LLVMValueRef index0_is_reset;
532
533 for (unsigned i = 0; i < 3; i++) {
534 LLVMValueRef not_reset = LLVMBuildICmp(builder, LLVMIntNE, index[i],
535 ac_get_arg(&ctx->ac, param_restart_index),
536 "");
537 if (i == 0)
538 index0_is_reset = LLVMBuildNot(builder, not_reset, "");
539 prim_restart_accepted = LLVMBuildAnd(builder, prim_restart_accepted,
540 not_reset, "");
541 }
542
543 /* If the previous waves flip the primitive orientation
544 * of the current triangle strip, it will be stored in GDS.
545 *
546 * Sometimes the correct orientation is not needed, in which case
547 * we don't need to execute this.
548 */
549 if (key->opt.cs_need_correct_orientation && VERTEX_COUNTER_GDS_MODE == 2) {
550 /* If there are reset indices in this wave, get the thread index
551 * where the most recent strip starts relative to each thread.
552 */
553 LLVMValueRef preceding_threads_mask =
554 LLVMBuildSub(builder,
555 LLVMBuildShl(builder, ctx->ac.i64_1,
556 LLVMBuildZExt(builder, thread_id, ctx->ac.i64, ""), ""),
557 ctx->ac.i64_1, "");
558
559 LLVMValueRef reset_threadmask = ac_get_i1_sgpr_mask(&ctx->ac, index0_is_reset);
560 LLVMValueRef preceding_reset_threadmask =
561 LLVMBuildAnd(builder, reset_threadmask, preceding_threads_mask, "");
562 LLVMValueRef strip_start =
563 ac_build_umsb(&ctx->ac, preceding_reset_threadmask, NULL);
564 strip_start = LLVMBuildAdd(builder, strip_start, ctx->ac.i32_1, "");
565
566 /* This flips the orientatino based on reset indices within this wave only. */
567 first_is_odd = LLVMBuildTrunc(builder, strip_start, ctx->ac.i1, "");
568
569 LLVMValueRef last_strip_start, prev_wave_state, ret, tmp;
570 LLVMValueRef is_first_wave, current_wave_resets_index;
571
572 /* Get the thread index where the last strip starts in this wave.
573 *
574 * If the last strip doesn't start in this wave, the thread index
575 * will be 0.
576 *
577 * If the last strip starts in the next wave, the thread index will
578 * be 64.
579 */
580 last_strip_start = ac_build_umsb(&ctx->ac, reset_threadmask, NULL);
581 last_strip_start = LLVMBuildAdd(builder, last_strip_start, ctx->ac.i32_1, "");
582
583 struct si_thread0_section section;
584 si_enter_thread0_section(ctx, &section, thread_id);
585
586 /* This must be done in the thread 0 section, because
587 * we expect PrimID to be 0 for the whole first wave
588 * in this expression.
589 *
590 * NOTE: This will need to be different if we wanna support
591 * instancing with primitive restart.
592 */
593 is_first_wave = LLVMBuildICmp(builder, LLVMIntEQ, prim_id, ctx->ac.i32_0, "");
594 is_first_wave = LLVMBuildAnd(builder, is_first_wave,
595 LLVMBuildNot(builder,
596 gds_prim_restart_continue, ""), "");
597 current_wave_resets_index = LLVMBuildICmp(builder, LLVMIntNE,
598 last_strip_start, ctx->ac.i32_0, "");
599
600 ret = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "prev_state");
601
602 /* Save the last strip start primitive index in GDS and read
603 * the value that previous waves stored.
604 *
605 * if (is_first_wave || current_wave_resets_strip)
606 * // Read the value that previous waves stored and store a new one.
607 * first_is_odd = ds.ordered.swap(last_strip_start);
608 * else
609 * // Just read the value that previous waves stored.
610 * first_is_odd = ds.ordered.add(0);
611 */
612 ac_build_ifcc(&ctx->ac,
613 LLVMBuildOr(builder, is_first_wave,
614 current_wave_resets_index, ""), 12602);
615 {
616 /* The GDS address is always 0 with ordered append. */
617 tmp = si_build_ds_ordered_op(ctx, "swap",
618 ordered_wave_id, last_strip_start,
619 1, true, false);
620 LLVMBuildStore(builder, tmp, ret);
621 }
622 ac_build_else(&ctx->ac, 12603);
623 {
624 /* Just read the value from GDS. */
625 tmp = si_build_ds_ordered_op(ctx, "add",
626 ordered_wave_id, ctx->ac.i32_0,
627 1, true, false);
628 LLVMBuildStore(builder, tmp, ret);
629 }
630 ac_build_endif(&ctx->ac, 12602);
631
632 prev_wave_state = LLVMBuildLoad(builder, ret, "");
633 /* Ignore the return value if this is the first wave. */
634 prev_wave_state = LLVMBuildSelect(builder, is_first_wave,
635 ctx->ac.i32_0, prev_wave_state, "");
636 si_exit_thread0_section(&section, &prev_wave_state);
637 prev_wave_state = LLVMBuildTrunc(builder, prev_wave_state, ctx->ac.i1, "");
638
639 /* If the strip start appears to be on thread 0 for the current primitive
640 * (meaning the reset index is not present in this wave and might have
641 * appeared in previous waves), use the value from GDS to determine
642 * primitive orientation.
643 *
644 * If the strip start is in this wave for the current primitive, use
645 * the value from the current wave to determine primitive orientation.
646 */
647 LLVMValueRef strip_start_is0 = LLVMBuildICmp(builder, LLVMIntEQ,
648 strip_start, ctx->ac.i32_0, "");
649 first_is_odd = LLVMBuildSelect(builder, strip_start_is0, prev_wave_state,
650 first_is_odd, "");
651 }
652 }
653 /* prim_is_odd = (first_is_odd + current_is_odd) % 2. */
654 LLVMValueRef prim_is_odd =
655 LLVMBuildXor(builder, first_is_odd,
656 LLVMBuildTrunc(builder, thread_id, ctx->ac.i1, ""), "");
657
658 /* Convert triangle strip indices to triangle indices. */
659 ac_build_triangle_strip_indices_to_triangle(&ctx->ac, prim_is_odd,
660 LLVMConstInt(ctx->ac.i1, key->opt.cs_provoking_vertex_first, 0),
661 index);
662 }
663
664 /* Execute the vertex shader for each vertex to get vertex positions. */
665 LLVMValueRef pos[3][4];
666 for (unsigned i = 0; i < vertices_per_prim; i++) {
667 vs_params[param_vertex_id] = index[i];
668 vs_params[param_instance_id] = instance_id;
669
670 LLVMValueRef ret = ac_build_call(&ctx->ac, vs, vs_params, num_vs_params);
671 for (unsigned chan = 0; chan < 4; chan++)
672 pos[i][chan] = LLVMBuildExtractValue(builder, ret, chan, "");
673 }
674
675 /* Divide XYZ by W. */
676 for (unsigned i = 0; i < vertices_per_prim; i++) {
677 for (unsigned chan = 0; chan < 3; chan++)
678 pos[i][chan] = ac_build_fdiv(&ctx->ac, pos[i][chan], pos[i][3]);
679 }
680
681 /* Load the viewport state. */
682 LLVMValueRef vp = ac_build_load_invariant(&ctx->ac, index_buffers_and_constants,
683 LLVMConstInt(ctx->ac.i32, 2, 0));
684 vp = LLVMBuildBitCast(builder, vp, ctx->ac.v4f32, "");
685 LLVMValueRef vp_scale[2], vp_translate[2];
686 vp_scale[0] = ac_llvm_extract_elem(&ctx->ac, vp, 0);
687 vp_scale[1] = ac_llvm_extract_elem(&ctx->ac, vp, 1);
688 vp_translate[0] = ac_llvm_extract_elem(&ctx->ac, vp, 2);
689 vp_translate[1] = ac_llvm_extract_elem(&ctx->ac, vp, 3);
690
691 /* Do culling. */
692 struct ac_cull_options options = {};
693 options.cull_front = key->opt.cs_cull_front;
694 options.cull_back = key->opt.cs_cull_back;
695 options.cull_view_xy = true;
696 options.cull_view_near_z = CULL_Z && key->opt.cs_cull_z;
697 options.cull_view_far_z = CULL_Z && key->opt.cs_cull_z;
698 options.cull_small_prims = true;
699 options.cull_zero_area = true;
700 options.cull_w = true;
701 options.use_halfz_clip_space = key->opt.cs_halfz_clip_space;
702
703 LLVMValueRef accepted =
704 ac_cull_triangle(&ctx->ac, pos, prim_restart_accepted,
705 vp_scale, vp_translate,
706 ac_get_arg(&ctx->ac, param_smallprim_precision),
707 &options);
708
709 ac_build_optimization_barrier(&ctx->ac, &accepted);
710 LLVMValueRef accepted_threadmask = ac_get_i1_sgpr_mask(&ctx->ac, accepted);
711
712 /* Count the number of active threads by doing bitcount(accepted). */
713 LLVMValueRef num_prims_accepted =
714 ac_build_intrinsic(&ctx->ac, "llvm.ctpop.i64", ctx->ac.i64,
715 &accepted_threadmask, 1, AC_FUNC_ATTR_READNONE);
716 num_prims_accepted = LLVMBuildTrunc(builder, num_prims_accepted, ctx->ac.i32, "");
717
718 LLVMValueRef start;
719
720 /* Execute atomic_add on the vertex count. */
721 struct si_thread0_section section;
722 si_enter_thread0_section(ctx, &section, thread_id);
723 {
724 if (VERTEX_COUNTER_GDS_MODE == 0) {
725 LLVMValueRef num_indices = LLVMBuildMul(builder, num_prims_accepted,
726 LLVMConstInt(ctx->ac.i32, vertices_per_prim, 0), "");
727 vertex_counter = si_expand_32bit_pointer(ctx, vertex_counter);
728 start = LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
729 vertex_counter, num_indices,
730 LLVMAtomicOrderingMonotonic, false);
731 } else if (VERTEX_COUNTER_GDS_MODE == 1) {
732 LLVMValueRef num_indices = LLVMBuildMul(builder, num_prims_accepted,
733 LLVMConstInt(ctx->ac.i32, vertices_per_prim, 0), "");
734 vertex_counter = LLVMBuildIntToPtr(builder, vertex_counter,
735 LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS), "");
736 start = LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
737 vertex_counter, num_indices,
738 LLVMAtomicOrderingMonotonic, false);
739 } else if (VERTEX_COUNTER_GDS_MODE == 2) {
740 LLVMValueRef tmp_store = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "");
741
742 /* If the draw call was split into multiple subdraws, each using
743 * a separate draw packet, we need to start counting from 0 for
744 * the first compute wave of the subdraw.
745 *
746 * vertex_counter contains the primitive ID of the first thread
747 * in the first wave.
748 *
749 * This is only correct with VERTEX_COUNTER_GDS_MODE == 2:
750 */
751 LLVMValueRef is_first_wave =
752 LLVMBuildICmp(builder, LLVMIntEQ, global_thread_id,
753 vertex_counter, "");
754
755 /* Store the primitive count for ordered append, not vertex count.
756 * The idea is to avoid GDS initialization via CP DMA. The shader
757 * effectively stores the first count using "swap".
758 *
759 * if (first_wave) {
760 * ds.ordered.swap(num_prims_accepted); // store the first primitive count
761 * previous = 0;
762 * } else {
763 * previous = ds.ordered.add(num_prims_accepted) // add the primitive count
764 * }
765 */
766 ac_build_ifcc(&ctx->ac, is_first_wave, 12604);
767 {
768 /* The GDS address is always 0 with ordered append. */
769 si_build_ds_ordered_op(ctx, "swap", ordered_wave_id,
770 num_prims_accepted, 0, true, true);
771 LLVMBuildStore(builder, ctx->ac.i32_0, tmp_store);
772 }
773 ac_build_else(&ctx->ac, 12605);
774 {
775 LLVMBuildStore(builder,
776 si_build_ds_ordered_op(ctx, "add", ordered_wave_id,
777 num_prims_accepted, 0,
778 true, true),
779 tmp_store);
780 }
781 ac_build_endif(&ctx->ac, 12604);
782
783 start = LLVMBuildLoad(builder, tmp_store, "");
784 }
785 }
786 si_exit_thread0_section(&section, &start);
787
788 /* Write the final vertex count to memory. An EOS/EOP event could do this,
789 * but those events are super slow and should be avoided if performance
790 * is a concern. Thanks to GDS ordered append, we can emulate a CS_DONE
791 * event like this.
792 */
793 if (VERTEX_COUNTER_GDS_MODE == 2) {
794 ac_build_ifcc(&ctx->ac,
795 LLVMBuildICmp(builder, LLVMIntEQ, global_thread_id,
796 ac_get_arg(&ctx->ac, param_last_wave_prim_id), ""),
797 12606);
798 LLVMValueRef count = LLVMBuildAdd(builder, start, num_prims_accepted, "");
799 count = LLVMBuildMul(builder, count,
800 LLVMConstInt(ctx->ac.i32, vertices_per_prim, 0), "");
801
802 /* GFX8 needs to disable caching, so that the CP can see the stored value.
803 * MTYPE=3 bypasses TC L2.
804 */
805 if (ctx->screen->info.chip_class <= GFX8) {
806 LLVMValueRef desc[] = {
807 ac_get_arg(&ctx->ac, param_vertex_count_addr),
808 LLVMConstInt(ctx->ac.i32,
809 S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0),
810 LLVMConstInt(ctx->ac.i32, 4, 0),
811 LLVMConstInt(ctx->ac.i32, S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
812 S_008F0C_MTYPE(3 /* uncached */), 0),
813 };
814 LLVMValueRef rsrc = ac_build_gather_values(&ctx->ac, desc, 4);
815 ac_build_buffer_store_dword(&ctx->ac, rsrc, count, 1, ctx->ac.i32_0,
816 ctx->ac.i32_0, 0, ac_glc | ac_slc);
817 } else {
818 LLVMBuildStore(builder, count,
819 si_expand_32bit_pointer(ctx,
820 ac_get_arg(&ctx->ac,
821 param_vertex_count_addr)));
822 }
823 ac_build_endif(&ctx->ac, 12606);
824 } else {
825 /* For unordered modes that increment a vertex count instead of
826 * primitive count, convert it into the primitive index.
827 */
828 start = LLVMBuildUDiv(builder, start,
829 LLVMConstInt(ctx->ac.i32, vertices_per_prim, 0), "");
830 }
831
832 /* Now we need to store the indices of accepted primitives into
833 * the output index buffer.
834 */
835 ac_build_ifcc(&ctx->ac, accepted, 16607);
836 {
837 /* Get the number of bits set before the index of this thread. */
838 LLVMValueRef prim_index = ac_build_mbcnt(&ctx->ac, accepted_threadmask);
839
840 /* We have lowered instancing. Pack the instance ID into vertex ID. */
841 if (key->opt.cs_instancing) {
842 instance_id = LLVMBuildShl(builder, instance_id,
843 LLVMConstInt(ctx->ac.i32, 16, 0), "");
844
845 for (unsigned i = 0; i < vertices_per_prim; i++)
846 index[i] = LLVMBuildOr(builder, index[i], instance_id, "");
847 }
848
849 if (VERTEX_COUNTER_GDS_MODE == 2) {
850 /* vertex_counter contains the first primitive ID
851 * for this dispatch. If the draw call was split into
852 * multiple subdraws, the first primitive ID is > 0
853 * for subsequent subdraws. Each subdraw uses a different
854 * portion of the output index buffer. Offset the store
855 * vindex by the first primitive ID to get the correct
856 * store address for the subdraw.
857 */
858 start = LLVMBuildAdd(builder, start, vertex_counter, "");
859 }
860
861 /* Write indices for accepted primitives. */
862 LLVMValueRef vindex = LLVMBuildAdd(builder, start, prim_index, "");
863 LLVMValueRef vdata = ac_build_gather_values(&ctx->ac, index, 3);
864
865 if (!ac_has_vec3_support(ctx->ac.chip_class, true))
866 vdata = ac_build_expand_to_vec4(&ctx->ac, vdata, 3);
867
868 ac_build_buffer_store_format(&ctx->ac, output_indexbuf, vdata,
869 vindex, ctx->ac.i32_0, 3,
870 ac_glc | (INDEX_STORES_USE_SLC ? ac_slc : 0));
871 }
872 ac_build_endif(&ctx->ac, 16607);
873
874 LLVMBuildRetVoid(builder);
875 }
876
877 /* Return false if the shader isn't ready. */
878 static bool si_shader_select_prim_discard_cs(struct si_context *sctx,
879 const struct pipe_draw_info *info,
880 bool primitive_restart)
881 {
882 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
883 struct si_shader_key key;
884
885 /* Primitive restart needs ordered counters. */
886 assert(!primitive_restart || VERTEX_COUNTER_GDS_MODE == 2);
887 assert(!primitive_restart || info->instance_count == 1);
888
889 memset(&key, 0, sizeof(key));
890 si_shader_selector_key_vs(sctx, sctx->vs_shader.cso, &key, &key.part.vs.prolog);
891 assert(!key.part.vs.prolog.instance_divisor_is_fetched);
892
893 key.part.vs.prolog.unpack_instance_id_from_vertex_id = 0;
894 key.opt.vs_as_prim_discard_cs = 1;
895 key.opt.cs_prim_type = info->mode;
896 key.opt.cs_indexed = info->index_size != 0;
897 key.opt.cs_instancing = info->instance_count > 1;
898 key.opt.cs_primitive_restart = primitive_restart;
899 key.opt.cs_provoking_vertex_first = rs->provoking_vertex_first;
900
901 /* Primitive restart with triangle strips needs to preserve primitive
902 * orientation for cases where front and back primitive orientation matters.
903 */
904 if (primitive_restart) {
905 struct si_shader_selector *ps = sctx->ps_shader.cso;
906
907 key.opt.cs_need_correct_orientation =
908 rs->cull_front != rs->cull_back ||
909 ps->info.uses_frontface ||
910 (rs->two_side && ps->info.colors_read);
911 }
912
913 if (rs->rasterizer_discard) {
914 /* Just for performance testing and analysis of trivial bottlenecks.
915 * This should result in a very short compute shader. */
916 key.opt.cs_cull_front = 1;
917 key.opt.cs_cull_back = 1;
918 } else {
919 key.opt.cs_cull_front =
920 sctx->viewports.y_inverted ? rs->cull_back : rs->cull_front;
921 key.opt.cs_cull_back =
922 sctx->viewports.y_inverted ? rs->cull_front : rs->cull_back;
923 }
924
925 if (!rs->depth_clamp_any && CULL_Z) {
926 key.opt.cs_cull_z = 1;
927 key.opt.cs_halfz_clip_space = rs->clip_halfz;
928 }
929
930 sctx->cs_prim_discard_state.cso = sctx->vs_shader.cso;
931 sctx->cs_prim_discard_state.current = NULL;
932
933 if (!sctx->compiler.passes)
934 si_init_compiler(sctx->screen, &sctx->compiler);
935
936 struct si_compiler_ctx_state compiler_state;
937 compiler_state.compiler = &sctx->compiler;
938 compiler_state.debug = sctx->debug;
939 compiler_state.is_debug_context = sctx->is_debug;
940
941 return si_shader_select_with_key(sctx->screen, &sctx->cs_prim_discard_state,
942 &compiler_state, &key, -1, true) == 0 &&
943 /* Disallow compute shaders using the scratch buffer. */
944 sctx->cs_prim_discard_state.current->config.scratch_bytes_per_wave == 0;
945 }
946
947 static bool si_initialize_prim_discard_cmdbuf(struct si_context *sctx)
948 {
949 if (sctx->index_ring)
950 return true;
951
952 if (!sctx->prim_discard_compute_cs) {
953 struct radeon_winsys *ws = sctx->ws;
954 unsigned gds_size = VERTEX_COUNTER_GDS_MODE == 1 ? GDS_SIZE_UNORDERED :
955 VERTEX_COUNTER_GDS_MODE == 2 ? 8 : 0;
956 unsigned num_oa_counters = VERTEX_COUNTER_GDS_MODE == 2 ? 2 : 0;
957
958 if (gds_size) {
959 sctx->gds = ws->buffer_create(ws, gds_size, 4,
960 RADEON_DOMAIN_GDS, 0);
961 if (!sctx->gds)
962 return false;
963
964 ws->cs_add_buffer(sctx->gfx_cs, sctx->gds,
965 RADEON_USAGE_READWRITE, 0, 0);
966 }
967 if (num_oa_counters) {
968 assert(gds_size);
969 sctx->gds_oa = ws->buffer_create(ws, num_oa_counters,
970 1, RADEON_DOMAIN_OA, 0);
971 if (!sctx->gds_oa)
972 return false;
973
974 ws->cs_add_buffer(sctx->gfx_cs, sctx->gds_oa,
975 RADEON_USAGE_READWRITE, 0, 0);
976 }
977
978 sctx->prim_discard_compute_cs =
979 ws->cs_add_parallel_compute_ib(sctx->gfx_cs,
980 num_oa_counters > 0);
981 if (!sctx->prim_discard_compute_cs)
982 return false;
983 }
984
985 if (!sctx->index_ring) {
986 sctx->index_ring =
987 si_aligned_buffer_create(sctx->b.screen,
988 SI_RESOURCE_FLAG_UNMAPPABLE,
989 PIPE_USAGE_DEFAULT,
990 sctx->index_ring_size_per_ib * 2,
991 sctx->screen->info.pte_fragment_size);
992 if (!sctx->index_ring)
993 return false;
994 }
995 return true;
996 }
997
998 static bool si_check_ring_space(struct si_context *sctx, unsigned out_indexbuf_size)
999 {
1000 return sctx->index_ring_offset +
1001 align(out_indexbuf_size, sctx->screen->info.tcc_cache_line_size) <=
1002 sctx->index_ring_size_per_ib;
1003 }
1004
1005 enum si_prim_discard_outcome
1006 si_prepare_prim_discard_or_split_draw(struct si_context *sctx,
1007 const struct pipe_draw_info *info,
1008 bool primitive_restart)
1009 {
1010 /* If the compute shader compilation isn't finished, this returns false. */
1011 if (!si_shader_select_prim_discard_cs(sctx, info, primitive_restart))
1012 return SI_PRIM_DISCARD_DISABLED;
1013
1014 if (!si_initialize_prim_discard_cmdbuf(sctx))
1015 return SI_PRIM_DISCARD_DISABLED;
1016
1017 struct radeon_cmdbuf *gfx_cs = sctx->gfx_cs;
1018 unsigned prim = info->mode;
1019 unsigned count = info->count;
1020 unsigned instance_count = info->instance_count;
1021 unsigned num_prims_per_instance = u_decomposed_prims_for_vertices(prim, count);
1022 unsigned num_prims = num_prims_per_instance * instance_count;
1023 unsigned out_indexbuf_size = num_prims * 12;
1024 bool ring_full = !si_check_ring_space(sctx, out_indexbuf_size);
1025 const unsigned split_prims_draw_level = SPLIT_PRIMS_DRAW_LEVEL;
1026
1027 /* Split draws at the draw call level if the ring is full. This makes
1028 * better use of the ring space.
1029 */
1030 if (ring_full &&
1031 num_prims > split_prims_draw_level &&
1032 instance_count == 1 && /* TODO: support splitting instanced draws */
1033 (1 << prim) & ((1 << PIPE_PRIM_TRIANGLES) |
1034 (1 << PIPE_PRIM_TRIANGLE_STRIP))) {
1035 /* Split draws. */
1036 struct pipe_draw_info split_draw = *info;
1037 split_draw.primitive_restart = primitive_restart;
1038
1039 unsigned base_start = split_draw.start;
1040
1041 if (prim == PIPE_PRIM_TRIANGLES) {
1042 unsigned vert_count_per_subdraw = split_prims_draw_level * 3;
1043 assert(vert_count_per_subdraw < count);
1044
1045 for (unsigned start = 0; start < count; start += vert_count_per_subdraw) {
1046 split_draw.start = base_start + start;
1047 split_draw.count = MIN2(count - start, vert_count_per_subdraw);
1048
1049 sctx->b.draw_vbo(&sctx->b, &split_draw);
1050 }
1051 } else if (prim == PIPE_PRIM_TRIANGLE_STRIP) {
1052 /* No primitive pair can be split, because strips reverse orientation
1053 * for odd primitives. */
1054 STATIC_ASSERT(split_prims_draw_level % 2 == 0);
1055
1056 unsigned vert_count_per_subdraw = split_prims_draw_level;
1057
1058 for (unsigned start = 0; start < count - 2; start += vert_count_per_subdraw) {
1059 split_draw.start = base_start + start;
1060 split_draw.count = MIN2(count - start, vert_count_per_subdraw + 2);
1061
1062 sctx->b.draw_vbo(&sctx->b, &split_draw);
1063
1064 if (start == 0 &&
1065 primitive_restart &&
1066 sctx->cs_prim_discard_state.current->key.opt.cs_need_correct_orientation)
1067 sctx->preserve_prim_restart_gds_at_flush = true;
1068 }
1069 sctx->preserve_prim_restart_gds_at_flush = false;
1070 } else {
1071 assert(0);
1072 }
1073
1074 return SI_PRIM_DISCARD_DRAW_SPLIT;
1075 }
1076
1077 /* Just quit if the draw call doesn't fit into the ring and can't be split. */
1078 if (out_indexbuf_size > sctx->index_ring_size_per_ib) {
1079 if (SI_PRIM_DISCARD_DEBUG)
1080 puts("PD failed: draw call too big, can't be split");
1081 return SI_PRIM_DISCARD_DISABLED;
1082 }
1083
1084 unsigned num_subdraws = DIV_ROUND_UP(num_prims, SPLIT_PRIMS_PACKET_LEVEL);
1085 unsigned need_compute_dw = 11 /* shader */ + 34 /* first draw */ +
1086 24 * (num_subdraws - 1) + /* subdraws */
1087 20; /* leave some space at the end */
1088 unsigned need_gfx_dw = si_get_minimum_num_gfx_cs_dwords(sctx);
1089
1090 if (sctx->chip_class <= GFX7 || FORCE_REWIND_EMULATION)
1091 need_gfx_dw += 9; /* NOP(2) + WAIT_REG_MEM(7), then chain */
1092 else
1093 need_gfx_dw += num_subdraws * 8; /* use REWIND(2) + DRAW(6) */
1094
1095 if (ring_full ||
1096 (VERTEX_COUNTER_GDS_MODE == 1 && sctx->compute_gds_offset + 8 > GDS_SIZE_UNORDERED) ||
1097 !sctx->ws->cs_check_space(gfx_cs, need_gfx_dw, false)) {
1098 /* If the current IB is empty but the size is too small, add a NOP
1099 * packet to force a flush and get a bigger IB.
1100 */
1101 if (!radeon_emitted(gfx_cs, sctx->initial_gfx_cs_size) &&
1102 gfx_cs->current.cdw + need_gfx_dw > gfx_cs->current.max_dw) {
1103 radeon_emit(gfx_cs, PKT3(PKT3_NOP, 0, 0));
1104 radeon_emit(gfx_cs, 0);
1105 }
1106
1107 si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
1108 }
1109
1110 /* The compute IB is always chained, but we need to call cs_check_space to add more space. */
1111 struct radeon_cmdbuf *cs = sctx->prim_discard_compute_cs;
1112 ASSERTED bool compute_has_space = sctx->ws->cs_check_space(cs, need_compute_dw, false);
1113 assert(compute_has_space);
1114 assert(si_check_ring_space(sctx, out_indexbuf_size));
1115 return SI_PRIM_DISCARD_ENABLED;
1116 }
1117
1118 void si_compute_signal_gfx(struct si_context *sctx)
1119 {
1120 struct radeon_cmdbuf *cs = sctx->prim_discard_compute_cs;
1121 unsigned writeback_L2_flags = 0;
1122
1123 /* The writeback L2 flags vary with each chip generation. */
1124 /* CI needs to flush vertex indices to memory. */
1125 if (sctx->chip_class <= GFX7)
1126 writeback_L2_flags = EVENT_TC_WB_ACTION_ENA;
1127 else if (sctx->chip_class == GFX8 && VERTEX_COUNTER_GDS_MODE == 0)
1128 writeback_L2_flags = EVENT_TC_WB_ACTION_ENA | EVENT_TC_NC_ACTION_ENA;
1129
1130 if (!sctx->compute_num_prims_in_batch)
1131 return;
1132
1133 assert(sctx->compute_rewind_va);
1134
1135 /* After the queued dispatches are done and vertex counts are written to
1136 * the gfx IB, signal the gfx IB to continue. CP doesn't wait for
1137 * the dispatches to finish, it only adds the CS_DONE event into the event
1138 * queue.
1139 */
1140 si_cp_release_mem(sctx, cs, V_028A90_CS_DONE, writeback_L2_flags,
1141 sctx->chip_class <= GFX8 ? EOP_DST_SEL_MEM : EOP_DST_SEL_TC_L2,
1142 writeback_L2_flags ? EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM :
1143 EOP_INT_SEL_NONE,
1144 EOP_DATA_SEL_VALUE_32BIT,
1145 NULL,
1146 sctx->compute_rewind_va |
1147 ((uint64_t)sctx->screen->info.address32_hi << 32),
1148 REWIND_SIGNAL_BIT, /* signaling value for the REWIND packet */
1149 SI_NOT_QUERY);
1150
1151 sctx->compute_rewind_va = 0;
1152 sctx->compute_num_prims_in_batch = 0;
1153 }
1154
1155 /* Dispatch a primitive discard compute shader. */
1156 void si_dispatch_prim_discard_cs_and_draw(struct si_context *sctx,
1157 const struct pipe_draw_info *info,
1158 unsigned index_size,
1159 unsigned base_vertex,
1160 uint64_t input_indexbuf_va,
1161 unsigned input_indexbuf_num_elements)
1162 {
1163 struct radeon_cmdbuf *gfx_cs = sctx->gfx_cs;
1164 struct radeon_cmdbuf *cs = sctx->prim_discard_compute_cs;
1165 unsigned num_prims_per_instance = u_decomposed_prims_for_vertices(info->mode, info->count);
1166 if (!num_prims_per_instance)
1167 return;
1168
1169 unsigned num_prims = num_prims_per_instance * info->instance_count;
1170 unsigned vertices_per_prim, output_indexbuf_format;
1171
1172 switch (info->mode) {
1173 case PIPE_PRIM_TRIANGLES:
1174 case PIPE_PRIM_TRIANGLE_STRIP:
1175 case PIPE_PRIM_TRIANGLE_FAN:
1176 vertices_per_prim = 3;
1177 output_indexbuf_format = V_008F0C_BUF_DATA_FORMAT_32_32_32;
1178 break;
1179 default:
1180 unreachable("unsupported primitive type");
1181 return;
1182 }
1183
1184 unsigned out_indexbuf_offset;
1185 uint64_t output_indexbuf_size = num_prims * vertices_per_prim * 4;
1186 bool first_dispatch = !sctx->prim_discard_compute_ib_initialized;
1187
1188 /* Initialize the compute IB if it's empty. */
1189 if (!sctx->prim_discard_compute_ib_initialized) {
1190 /* 1) State initialization. */
1191 sctx->compute_gds_offset = 0;
1192 sctx->compute_ib_last_shader = NULL;
1193
1194 if (sctx->last_ib_barrier_fence) {
1195 assert(!sctx->last_ib_barrier_buf);
1196 sctx->ws->cs_add_fence_dependency(gfx_cs,
1197 sctx->last_ib_barrier_fence,
1198 RADEON_DEPENDENCY_PARALLEL_COMPUTE_ONLY);
1199 }
1200
1201 /* 2) IB initialization. */
1202
1203 /* This needs to be done at the beginning of IBs due to possible
1204 * TTM buffer moves in the kernel.
1205 *
1206 * TODO: update for GFX10
1207 */
1208 si_emit_surface_sync(sctx, cs,
1209 S_0085F0_TC_ACTION_ENA(1) |
1210 S_0085F0_TCL1_ACTION_ENA(1) |
1211 S_0301F0_TC_WB_ACTION_ENA(sctx->chip_class >= GFX8) |
1212 S_0085F0_SH_ICACHE_ACTION_ENA(1) |
1213 S_0085F0_SH_KCACHE_ACTION_ENA(1));
1214
1215 /* Restore the GDS prim restart counter if needed. */
1216 if (sctx->preserve_prim_restart_gds_at_flush) {
1217 si_cp_copy_data(sctx, cs,
1218 COPY_DATA_GDS, NULL, 4,
1219 COPY_DATA_SRC_MEM, sctx->wait_mem_scratch, 4);
1220 }
1221
1222 si_emit_initial_compute_regs(sctx, cs);
1223
1224 radeon_set_sh_reg(cs, R_00B860_COMPUTE_TMPRING_SIZE,
1225 S_00B860_WAVES(sctx->scratch_waves) |
1226 S_00B860_WAVESIZE(0)); /* no scratch */
1227
1228 /* Only 1D grids are launched. */
1229 radeon_set_sh_reg_seq(cs, R_00B820_COMPUTE_NUM_THREAD_Y, 2);
1230 radeon_emit(cs, S_00B820_NUM_THREAD_FULL(1) |
1231 S_00B820_NUM_THREAD_PARTIAL(1));
1232 radeon_emit(cs, S_00B824_NUM_THREAD_FULL(1) |
1233 S_00B824_NUM_THREAD_PARTIAL(1));
1234
1235 radeon_set_sh_reg_seq(cs, R_00B814_COMPUTE_START_Y, 2);
1236 radeon_emit(cs, 0);
1237 radeon_emit(cs, 0);
1238
1239 /* Disable ordered alloc for OA resources. */
1240 for (unsigned i = 0; i < 2; i++) {
1241 radeon_set_uconfig_reg_seq(cs, R_031074_GDS_OA_CNTL, 3);
1242 radeon_emit(cs, S_031074_INDEX(i));
1243 radeon_emit(cs, 0);
1244 radeon_emit(cs, S_03107C_ENABLE(0));
1245 }
1246
1247 if (sctx->last_ib_barrier_buf) {
1248 assert(!sctx->last_ib_barrier_fence);
1249 radeon_add_to_buffer_list(sctx, gfx_cs, sctx->last_ib_barrier_buf,
1250 RADEON_USAGE_READ, RADEON_PRIO_FENCE);
1251 si_cp_wait_mem(sctx, cs,
1252 sctx->last_ib_barrier_buf->gpu_address +
1253 sctx->last_ib_barrier_buf_offset, 1, 1,
1254 WAIT_REG_MEM_EQUAL);
1255 }
1256
1257 sctx->prim_discard_compute_ib_initialized = true;
1258 }
1259
1260 /* Allocate the output index buffer. */
1261 output_indexbuf_size = align(output_indexbuf_size,
1262 sctx->screen->info.tcc_cache_line_size);
1263 assert(sctx->index_ring_offset + output_indexbuf_size <= sctx->index_ring_size_per_ib);
1264 out_indexbuf_offset = sctx->index_ring_base + sctx->index_ring_offset;
1265 sctx->index_ring_offset += output_indexbuf_size;
1266
1267 radeon_add_to_buffer_list(sctx, gfx_cs, sctx->index_ring, RADEON_USAGE_READWRITE,
1268 RADEON_PRIO_SHADER_RW_BUFFER);
1269 uint64_t out_indexbuf_va = sctx->index_ring->gpu_address + out_indexbuf_offset;
1270
1271 /* Prepare index buffer descriptors. */
1272 struct si_resource *indexbuf_desc = NULL;
1273 unsigned indexbuf_desc_offset;
1274 unsigned desc_size = 12 * 4;
1275 uint32_t *desc;
1276
1277 u_upload_alloc(sctx->b.const_uploader, 0, desc_size,
1278 si_optimal_tcc_alignment(sctx, desc_size),
1279 &indexbuf_desc_offset, (struct pipe_resource**)&indexbuf_desc,
1280 (void**)&desc);
1281 radeon_add_to_buffer_list(sctx, gfx_cs, indexbuf_desc, RADEON_USAGE_READ,
1282 RADEON_PRIO_DESCRIPTORS);
1283
1284 /* Input index buffer. */
1285 desc[0] = input_indexbuf_va;
1286 desc[1] = S_008F04_BASE_ADDRESS_HI(input_indexbuf_va >> 32) |
1287 S_008F04_STRIDE(index_size);
1288 desc[2] = input_indexbuf_num_elements * (sctx->chip_class == GFX8 ? index_size : 1);
1289 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
1290 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_UINT) |
1291 S_008F0C_DATA_FORMAT(index_size == 1 ? V_008F0C_BUF_DATA_FORMAT_8 :
1292 index_size == 2 ? V_008F0C_BUF_DATA_FORMAT_16 :
1293 V_008F0C_BUF_DATA_FORMAT_32);
1294
1295 /* Output index buffer. */
1296 desc[4] = out_indexbuf_va;
1297 desc[5] = S_008F04_BASE_ADDRESS_HI(out_indexbuf_va >> 32) |
1298 S_008F04_STRIDE(vertices_per_prim * 4);
1299 desc[6] = num_prims * (sctx->chip_class == GFX8 ? vertices_per_prim * 4 : 1);
1300 desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
1301 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
1302 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
1303 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_0) |
1304 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_UINT) |
1305 S_008F0C_DATA_FORMAT(output_indexbuf_format);
1306
1307 /* Viewport state. */
1308 struct si_small_prim_cull_info cull_info;
1309 si_get_small_prim_cull_info(sctx, &cull_info);
1310
1311 desc[8] = fui(cull_info.scale[0]);
1312 desc[9] = fui(cull_info.scale[1]);
1313 desc[10] = fui(cull_info.translate[0]);
1314 desc[11] = fui(cull_info.translate[1]);
1315
1316 /* Better subpixel precision increases the efficiency of small
1317 * primitive culling. */
1318 unsigned num_samples = sctx->framebuffer.nr_samples;
1319 unsigned quant_mode = sctx->viewports.as_scissor[0].quant_mode;
1320 float small_prim_cull_precision;
1321
1322 if (quant_mode == SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH)
1323 small_prim_cull_precision = num_samples / 4096.0;
1324 else if (quant_mode == SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH)
1325 small_prim_cull_precision = num_samples / 1024.0;
1326 else
1327 small_prim_cull_precision = num_samples / 256.0;
1328
1329 /* Set user data SGPRs. */
1330 /* This can't be greater than 14 if we want the fastest launch rate. */
1331 unsigned user_sgprs = 13;
1332
1333 uint64_t index_buffers_va = indexbuf_desc->gpu_address + indexbuf_desc_offset;
1334 unsigned vs_const_desc = si_const_and_shader_buffer_descriptors_idx(PIPE_SHADER_VERTEX);
1335 unsigned vs_sampler_desc = si_sampler_and_image_descriptors_idx(PIPE_SHADER_VERTEX);
1336 uint64_t vs_const_desc_va = sctx->descriptors[vs_const_desc].gpu_address;
1337 uint64_t vs_sampler_desc_va = sctx->descriptors[vs_sampler_desc].gpu_address;
1338 uint64_t vb_desc_va = sctx->vb_descriptors_buffer ?
1339 sctx->vb_descriptors_buffer->gpu_address +
1340 sctx->vb_descriptors_offset : 0;
1341 unsigned gds_offset, gds_size;
1342 struct si_fast_udiv_info32 num_prims_udiv = {};
1343
1344 if (info->instance_count > 1)
1345 num_prims_udiv = si_compute_fast_udiv_info32(num_prims_per_instance, 31);
1346
1347 /* Limitations on how these two are packed in the user SGPR. */
1348 assert(num_prims_udiv.post_shift < 32);
1349 assert(num_prims_per_instance < 1 << 27);
1350
1351 si_resource_reference(&indexbuf_desc, NULL);
1352
1353 bool primitive_restart = sctx->cs_prim_discard_state.current->key.opt.cs_primitive_restart;
1354
1355 if (VERTEX_COUNTER_GDS_MODE == 1) {
1356 gds_offset = sctx->compute_gds_offset;
1357 gds_size = primitive_restart ? 8 : 4;
1358 sctx->compute_gds_offset += gds_size;
1359
1360 /* Reset the counters in GDS for the first dispatch using WRITE_DATA.
1361 * The remainder of the GDS will be cleared after the dispatch packet
1362 * in parallel with compute shaders.
1363 */
1364 if (first_dispatch) {
1365 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 2 + gds_size/4, 0));
1366 radeon_emit(cs, S_370_DST_SEL(V_370_GDS) | S_370_WR_CONFIRM(1));
1367 radeon_emit(cs, gds_offset);
1368 radeon_emit(cs, 0);
1369 radeon_emit(cs, 0); /* value to write */
1370 if (gds_size == 8)
1371 radeon_emit(cs, 0);
1372 }
1373 }
1374
1375 /* Set shader registers. */
1376 struct si_shader *shader = sctx->cs_prim_discard_state.current;
1377
1378 if (shader != sctx->compute_ib_last_shader) {
1379 radeon_add_to_buffer_list(sctx, gfx_cs, shader->bo, RADEON_USAGE_READ,
1380 RADEON_PRIO_SHADER_BINARY);
1381 uint64_t shader_va = shader->bo->gpu_address;
1382
1383 assert(shader->config.scratch_bytes_per_wave == 0);
1384 assert(shader->config.num_vgprs * WAVES_PER_TG <= 256 * 4);
1385
1386 radeon_set_sh_reg_seq(cs, R_00B830_COMPUTE_PGM_LO, 2);
1387 radeon_emit(cs, shader_va >> 8);
1388 radeon_emit(cs, S_00B834_DATA(shader_va >> 40));
1389
1390 radeon_set_sh_reg_seq(cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
1391 radeon_emit(cs, S_00B848_VGPRS((shader->config.num_vgprs - 1) / 4) |
1392 S_00B848_SGPRS((shader->config.num_sgprs - 1) / 8) |
1393 S_00B848_FLOAT_MODE(shader->config.float_mode) |
1394 S_00B848_DX10_CLAMP(1));
1395 radeon_emit(cs, S_00B84C_SCRATCH_EN(0 /* no scratch */) |
1396 S_00B84C_USER_SGPR(user_sgprs) |
1397 S_00B84C_TGID_X_EN(1 /* only blockID.x is used */) |
1398 S_00B84C_TG_SIZE_EN(VERTEX_COUNTER_GDS_MODE == 2 /* need the wave ID */) |
1399 S_00B84C_TIDIG_COMP_CNT(0 /* only threadID.x is used */) |
1400 S_00B84C_LDS_SIZE(shader->config.lds_size));
1401
1402 radeon_set_sh_reg(cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
1403 ac_get_compute_resource_limits(&sctx->screen->info,
1404 WAVES_PER_TG,
1405 MAX_WAVES_PER_SH,
1406 THREADGROUPS_PER_CU));
1407 sctx->compute_ib_last_shader = shader;
1408 }
1409
1410 STATIC_ASSERT(SPLIT_PRIMS_PACKET_LEVEL % THREADGROUP_SIZE == 0);
1411
1412 /* Big draw calls are split into smaller dispatches and draw packets. */
1413 for (unsigned start_prim = 0; start_prim < num_prims; start_prim += SPLIT_PRIMS_PACKET_LEVEL) {
1414 unsigned num_subdraw_prims;
1415
1416 if (start_prim + SPLIT_PRIMS_PACKET_LEVEL < num_prims)
1417 num_subdraw_prims = SPLIT_PRIMS_PACKET_LEVEL;
1418 else
1419 num_subdraw_prims = num_prims - start_prim;
1420
1421 /* Small dispatches are executed back to back until a specific primitive
1422 * count is reached. Then, a CS_DONE is inserted to signal the gfx IB
1423 * to start drawing the batch. This batching adds latency to the gfx IB,
1424 * but CS_DONE and REWIND are too slow.
1425 */
1426 if (sctx->compute_num_prims_in_batch + num_subdraw_prims > PRIMS_PER_BATCH)
1427 si_compute_signal_gfx(sctx);
1428
1429 if (sctx->compute_num_prims_in_batch == 0) {
1430 assert((gfx_cs->gpu_address >> 32) == sctx->screen->info.address32_hi);
1431 sctx->compute_rewind_va = gfx_cs->gpu_address + (gfx_cs->current.cdw + 1) * 4;
1432
1433 if (sctx->chip_class <= GFX7 || FORCE_REWIND_EMULATION) {
1434 radeon_emit(gfx_cs, PKT3(PKT3_NOP, 0, 0));
1435 radeon_emit(gfx_cs, 0);
1436
1437 si_cp_wait_mem(sctx, gfx_cs,
1438 sctx->compute_rewind_va |
1439 (uint64_t)sctx->screen->info.address32_hi << 32,
1440 REWIND_SIGNAL_BIT, REWIND_SIGNAL_BIT,
1441 WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_PFP);
1442
1443 /* Use INDIRECT_BUFFER to chain to a different buffer
1444 * to discard the CP prefetch cache.
1445 */
1446 sctx->ws->cs_check_space(gfx_cs, 0, true);
1447 } else {
1448 radeon_emit(gfx_cs, PKT3(PKT3_REWIND, 0, 0));
1449 radeon_emit(gfx_cs, 0);
1450 }
1451 }
1452
1453 sctx->compute_num_prims_in_batch += num_subdraw_prims;
1454
1455 uint32_t count_va = gfx_cs->gpu_address + (gfx_cs->current.cdw + 4) * 4;
1456 uint64_t index_va = out_indexbuf_va + start_prim * 12;
1457
1458 /* Emit the draw packet into the gfx IB. */
1459 radeon_emit(gfx_cs, PKT3(PKT3_DRAW_INDEX_2, 4, 0));
1460 radeon_emit(gfx_cs, num_prims * vertices_per_prim);
1461 radeon_emit(gfx_cs, index_va);
1462 radeon_emit(gfx_cs, index_va >> 32);
1463 radeon_emit(gfx_cs, 0);
1464 radeon_emit(gfx_cs, V_0287F0_DI_SRC_SEL_DMA);
1465
1466 /* Continue with the compute IB. */
1467 if (start_prim == 0) {
1468 uint32_t gds_prim_restart_continue_bit = 0;
1469
1470 if (sctx->preserve_prim_restart_gds_at_flush) {
1471 assert(primitive_restart &&
1472 info->mode == PIPE_PRIM_TRIANGLE_STRIP);
1473 assert(start_prim < 1 << 31);
1474 gds_prim_restart_continue_bit = 1 << 31;
1475 }
1476
1477 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0, user_sgprs);
1478 radeon_emit(cs, index_buffers_va);
1479 radeon_emit(cs,
1480 VERTEX_COUNTER_GDS_MODE == 0 ? count_va :
1481 VERTEX_COUNTER_GDS_MODE == 1 ? gds_offset :
1482 start_prim |
1483 gds_prim_restart_continue_bit);
1484 radeon_emit(cs, start_prim + num_subdraw_prims - 1);
1485 radeon_emit(cs, count_va);
1486 radeon_emit(cs, vb_desc_va);
1487 radeon_emit(cs, vs_const_desc_va);
1488 radeon_emit(cs, vs_sampler_desc_va);
1489 radeon_emit(cs, base_vertex);
1490 radeon_emit(cs, info->start_instance);
1491 radeon_emit(cs, num_prims_udiv.multiplier);
1492 radeon_emit(cs, num_prims_udiv.post_shift |
1493 (num_prims_per_instance << 5));
1494 radeon_emit(cs, info->restart_index);
1495 /* small-prim culling precision (same as rasterizer precision = QUANT_MODE) */
1496 radeon_emit(cs, fui(small_prim_cull_precision));
1497 } else {
1498 assert(VERTEX_COUNTER_GDS_MODE == 2);
1499 /* Only update the SGPRs that changed. */
1500 radeon_set_sh_reg_seq(cs, R_00B904_COMPUTE_USER_DATA_1, 3);
1501 radeon_emit(cs, start_prim);
1502 radeon_emit(cs, start_prim + num_subdraw_prims - 1);
1503 radeon_emit(cs, count_va);
1504 }
1505
1506 /* Set grid dimensions. */
1507 unsigned start_block = start_prim / THREADGROUP_SIZE;
1508 unsigned num_full_blocks = num_subdraw_prims / THREADGROUP_SIZE;
1509 unsigned partial_block_size = num_subdraw_prims % THREADGROUP_SIZE;
1510
1511 radeon_set_sh_reg(cs, R_00B810_COMPUTE_START_X, start_block);
1512 radeon_set_sh_reg(cs, R_00B81C_COMPUTE_NUM_THREAD_X,
1513 S_00B81C_NUM_THREAD_FULL(THREADGROUP_SIZE) |
1514 S_00B81C_NUM_THREAD_PARTIAL(partial_block_size));
1515
1516 radeon_emit(cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
1517 PKT3_SHADER_TYPE_S(1));
1518 radeon_emit(cs, start_block + num_full_blocks + !!partial_block_size);
1519 radeon_emit(cs, 1);
1520 radeon_emit(cs, 1);
1521 radeon_emit(cs, S_00B800_COMPUTE_SHADER_EN(1) |
1522 S_00B800_PARTIAL_TG_EN(!!partial_block_size) |
1523 S_00B800_ORDERED_APPEND_ENBL(VERTEX_COUNTER_GDS_MODE == 2) |
1524 S_00B800_ORDER_MODE(0 /* launch in order */));
1525
1526 /* This is only for unordered append. Ordered append writes this from
1527 * the shader.
1528 *
1529 * Note that EOP and EOS events are super slow, so emulating the event
1530 * in a shader is an important optimization.
1531 */
1532 if (VERTEX_COUNTER_GDS_MODE == 1) {
1533 si_cp_release_mem(sctx, cs, V_028A90_CS_DONE, 0,
1534 sctx->chip_class <= GFX8 ? EOP_DST_SEL_MEM : EOP_DST_SEL_TC_L2,
1535 EOP_INT_SEL_NONE,
1536 EOP_DATA_SEL_GDS,
1537 NULL,
1538 count_va | ((uint64_t)sctx->screen->info.address32_hi << 32),
1539 EOP_DATA_GDS(gds_offset / 4, 1),
1540 SI_NOT_QUERY);
1541
1542 /* Now that compute shaders are running, clear the remainder of GDS. */
1543 if (first_dispatch) {
1544 unsigned offset = gds_offset + gds_size;
1545 si_cp_dma_clear_buffer(sctx, cs, NULL, offset,
1546 GDS_SIZE_UNORDERED - offset,
1547 0,
1548 SI_CPDMA_SKIP_CHECK_CS_SPACE |
1549 SI_CPDMA_SKIP_GFX_SYNC |
1550 SI_CPDMA_SKIP_SYNC_BEFORE,
1551 SI_COHERENCY_NONE, L2_BYPASS);
1552 }
1553 }
1554 first_dispatch = false;
1555
1556 assert(cs->current.cdw <= cs->current.max_dw);
1557 assert(gfx_cs->current.cdw <= gfx_cs->current.max_dw);
1558 }
1559 }