radeonsi: do not do two full flushes on every compute dispatch
[mesa.git] / src / gallium / drivers / radeonsi / si_hw_context.c
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jerome Glisse
25 */
26
27 #include "si_pipe.h"
28
29 static unsigned si_descriptor_list_cs_space(unsigned count, unsigned element_size)
30 {
31 /* Ensure we have enough space to start a new range in a hole */
32 assert(element_size >= 3);
33
34 /* 5 dwords for possible load to reinitialize when we have no preamble
35 * IB + 5 dwords for write to L2 + 3 bytes for every range written to
36 * CE RAM.
37 */
38 return 5 + 5 + 3 + count * element_size;
39 }
40
41 static unsigned si_ce_needed_cs_space(void)
42 {
43 unsigned space = 0;
44
45 space += si_descriptor_list_cs_space(SI_NUM_CONST_BUFFERS, 4);
46 space += si_descriptor_list_cs_space(SI_NUM_RW_BUFFERS, 4);
47 space += si_descriptor_list_cs_space(SI_NUM_SHADER_BUFFERS, 4);
48 space += si_descriptor_list_cs_space(SI_NUM_SAMPLERS, 16);
49 space += si_descriptor_list_cs_space(SI_NUM_IMAGES, 8);
50
51 space *= SI_NUM_SHADERS;
52
53 /* Increment CE counter packet */
54 space += 2;
55
56 return space;
57 }
58
59 /* initialize */
60 void si_need_cs_space(struct si_context *ctx)
61 {
62 struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
63 struct radeon_winsys_cs *ce_ib = ctx->ce_ib;
64 struct radeon_winsys_cs *dma = ctx->b.dma.cs;
65
66 /* Flush the DMA IB if it's not empty. */
67 if (dma && dma->cdw)
68 ctx->b.dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
69
70 /* There are two memory usage counters in the winsys for all buffers
71 * that have been added (cs_add_buffer) and two counters in the pipe
72 * driver for those that haven't been added yet.
73 */
74 if (unlikely(!ctx->b.ws->cs_memory_below_limit(ctx->b.gfx.cs,
75 ctx->b.vram, ctx->b.gtt))) {
76 ctx->b.gtt = 0;
77 ctx->b.vram = 0;
78 ctx->b.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
79 return;
80 }
81 ctx->b.gtt = 0;
82 ctx->b.vram = 0;
83
84 /* If the CS is sufficiently large, don't count the space needed
85 * and just flush if there is not enough space left.
86 */
87 if (unlikely(cs->cdw > cs->max_dw - 2048 ||
88 (ce_ib && ce_ib->max_dw - ce_ib->cdw <
89 si_ce_needed_cs_space())))
90 ctx->b.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
91 }
92
93 void si_context_gfx_flush(void *context, unsigned flags,
94 struct pipe_fence_handle **fence)
95 {
96 struct si_context *ctx = context;
97 struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
98 struct radeon_winsys *ws = ctx->b.ws;
99
100 if (ctx->gfx_flush_in_progress)
101 return;
102
103 ctx->gfx_flush_in_progress = true;
104
105 if (cs->cdw == ctx->b.initial_gfx_cs_size &&
106 (!fence || ctx->last_gfx_fence)) {
107 if (fence)
108 ws->fence_reference(fence, ctx->last_gfx_fence);
109 if (!(flags & RADEON_FLUSH_ASYNC))
110 ws->cs_sync_flush(cs);
111 ctx->gfx_flush_in_progress = false;
112 return;
113 }
114
115 r600_preflush_suspend_features(&ctx->b);
116
117 ctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER |
118 SI_CONTEXT_INV_VMEM_L1 |
119 SI_CONTEXT_INV_GLOBAL_L2 |
120 SI_CONTEXT_CS_PARTIAL_FLUSH |
121 /* this is probably not needed anymore */
122 SI_CONTEXT_PS_PARTIAL_FLUSH;
123 si_emit_cache_flush(ctx, NULL);
124
125 /* force to keep tiling flags */
126 flags |= RADEON_FLUSH_KEEP_TILING_FLAGS;
127
128 if (ctx->trace_buf)
129 si_trace_emit(ctx);
130
131 if (ctx->is_debug) {
132 unsigned i;
133
134 /* Save the IB for debug contexts. */
135 free(ctx->last_ib);
136 ctx->last_ib_dw_size = cs->cdw;
137 ctx->last_ib = malloc(cs->cdw * 4);
138 memcpy(ctx->last_ib, cs->buf, cs->cdw * 4);
139 r600_resource_reference(&ctx->last_trace_buf, ctx->trace_buf);
140 r600_resource_reference(&ctx->trace_buf, NULL);
141
142 /* Save the buffer list. */
143 if (ctx->last_bo_list) {
144 for (i = 0; i < ctx->last_bo_count; i++)
145 pb_reference(&ctx->last_bo_list[i].buf, NULL);
146 free(ctx->last_bo_list);
147 }
148 ctx->last_bo_count = ws->cs_get_buffer_list(cs, NULL);
149 ctx->last_bo_list = calloc(ctx->last_bo_count,
150 sizeof(ctx->last_bo_list[0]));
151 ws->cs_get_buffer_list(cs, ctx->last_bo_list);
152 }
153
154 /* Flush the CS. */
155 ws->cs_flush(cs, flags, &ctx->last_gfx_fence);
156
157 if (fence)
158 ws->fence_reference(fence, ctx->last_gfx_fence);
159
160 /* Check VM faults if needed. */
161 if (ctx->screen->b.debug_flags & DBG_CHECK_VM)
162 si_check_vm_faults(ctx);
163
164 si_begin_new_cs(ctx);
165 ctx->gfx_flush_in_progress = false;
166 }
167
168 void si_begin_new_cs(struct si_context *ctx)
169 {
170 if (ctx->is_debug) {
171 uint32_t zero = 0;
172
173 /* Create a buffer used for writing trace IDs and initialize it to 0. */
174 assert(!ctx->trace_buf);
175 ctx->trace_buf = (struct r600_resource*)
176 pipe_buffer_create(ctx->b.b.screen, PIPE_BIND_CUSTOM,
177 PIPE_USAGE_STAGING, 4);
178 if (ctx->trace_buf)
179 pipe_buffer_write_nooverlap(&ctx->b.b, &ctx->trace_buf->b.b,
180 0, sizeof(zero), &zero);
181 ctx->trace_id = 0;
182 }
183
184 if (ctx->trace_buf)
185 si_trace_emit(ctx);
186
187 /* Flush read caches at the beginning of CS. */
188 ctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER |
189 SI_CONTEXT_INV_VMEM_L1 |
190 SI_CONTEXT_INV_GLOBAL_L2 |
191 SI_CONTEXT_INV_SMEM_L1 |
192 SI_CONTEXT_INV_ICACHE |
193 R600_CONTEXT_START_PIPELINE_STATS;
194
195 /* set all valid group as dirty so they get reemited on
196 * next draw command
197 */
198 si_pm4_reset_emitted(ctx);
199
200 /* The CS initialization should be emitted before everything else. */
201 si_pm4_emit(ctx, ctx->init_config);
202 if (ctx->init_config_gs_rings)
203 si_pm4_emit(ctx, ctx->init_config_gs_rings);
204
205 ctx->framebuffer.dirty_cbufs = (1 << 8) - 1;
206 ctx->framebuffer.dirty_zsbuf = true;
207 si_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
208
209 si_mark_atom_dirty(ctx, &ctx->clip_regs);
210 si_mark_atom_dirty(ctx, &ctx->clip_state.atom);
211 si_mark_atom_dirty(ctx, &ctx->msaa_sample_locs);
212 si_mark_atom_dirty(ctx, &ctx->msaa_config);
213 si_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
214 si_mark_atom_dirty(ctx, &ctx->cb_render_state);
215 si_mark_atom_dirty(ctx, &ctx->blend_color.atom);
216 si_mark_atom_dirty(ctx, &ctx->db_render_state);
217 si_mark_atom_dirty(ctx, &ctx->stencil_ref.atom);
218 si_mark_atom_dirty(ctx, &ctx->spi_map);
219 si_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
220 si_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
221 si_all_descriptors_begin_new_cs(ctx);
222
223 ctx->b.scissors.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
224 ctx->b.viewports.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
225 si_mark_atom_dirty(ctx, &ctx->b.scissors.atom);
226 si_mark_atom_dirty(ctx, &ctx->b.viewports.atom);
227
228 r600_postflush_resume_features(&ctx->b);
229
230 ctx->b.initial_gfx_cs_size = ctx->b.gfx.cs->cdw;
231
232 /* Invalidate various draw states so that they are emitted before
233 * the first draw call. */
234 si_invalidate_draw_sh_constants(ctx);
235 ctx->last_primitive_restart_en = -1;
236 ctx->last_restart_index = SI_RESTART_INDEX_UNKNOWN;
237 ctx->last_gs_out_prim = -1;
238 ctx->last_prim = -1;
239 ctx->last_multi_vgt_param = -1;
240 ctx->last_ls_hs_config = -1;
241 ctx->last_rast_prim = -1;
242 ctx->last_sc_line_stipple = ~0;
243 ctx->emit_scratch_reloc = true;
244 ctx->last_ls = NULL;
245 ctx->last_tcs = NULL;
246 ctx->last_tes_sh_base = -1;
247 ctx->last_num_tcs_input_cp = -1;
248
249 ctx->cs_shader_state.initialized = false;
250 }