2 * Copyright (c) 2015 Intel Corporation
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
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
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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "common/gen_l3_config.h"
26 #include "brw_context.h"
27 #include "brw_defines.h"
28 #include "brw_state.h"
29 #include "intel_batchbuffer.h"
32 * Calculate the desired L3 partitioning based on the current state of the
33 * pipeline. For now this simply returns the conservative defaults calculated
34 * by get_default_l3_weights(), but we could probably do better by gathering
35 * more statistics from the pipeline state (e.g. guess of expected URB usage
36 * and bound surfaces), or by using feed-back from performance counters.
38 static struct gen_l3_weights
39 get_pipeline_state_l3_weights(const struct brw_context
*brw
)
41 const struct brw_stage_state
*stage_states
[] = {
42 [MESA_SHADER_VERTEX
] = &brw
->vs
.base
,
43 [MESA_SHADER_TESS_CTRL
] = &brw
->tcs
.base
,
44 [MESA_SHADER_TESS_EVAL
] = &brw
->tes
.base
,
45 [MESA_SHADER_GEOMETRY
] = &brw
->gs
.base
,
46 [MESA_SHADER_FRAGMENT
] = &brw
->wm
.base
,
47 [MESA_SHADER_COMPUTE
] = &brw
->cs
.base
49 bool needs_dc
= false, needs_slm
= false;
51 for (unsigned i
= 0; i
< ARRAY_SIZE(stage_states
); i
++) {
52 const struct gl_shader_program
*prog
=
53 brw
->ctx
._Shader
->CurrentProgram
[stage_states
[i
]->stage
];
54 const struct brw_stage_prog_data
*prog_data
= stage_states
[i
]->prog_data
;
56 needs_dc
|= (prog
&& prog
->NumAtomicBuffers
) ||
57 (prog_data
&& (prog_data
->total_scratch
|| prog_data
->nr_image_params
));
58 needs_slm
|= prog_data
&& prog_data
->total_shared
;
61 return gen_get_default_l3_weights(&brw
->screen
->devinfo
,
66 * Program the hardware to use the specified L3 configuration.
69 setup_l3_config(struct brw_context
*brw
, const struct gen_l3_config
*cfg
)
71 const bool has_dc
= cfg
->n
[GEN_L3P_DC
] || cfg
->n
[GEN_L3P_ALL
];
72 const bool has_is
= cfg
->n
[GEN_L3P_IS
] || cfg
->n
[GEN_L3P_RO
] ||
74 const bool has_c
= cfg
->n
[GEN_L3P_C
] || cfg
->n
[GEN_L3P_RO
] ||
76 const bool has_t
= cfg
->n
[GEN_L3P_T
] || cfg
->n
[GEN_L3P_RO
] ||
78 const bool has_slm
= cfg
->n
[GEN_L3P_SLM
];
80 /* According to the hardware docs, the L3 partitioning can only be changed
81 * while the pipeline is completely drained and the caches are flushed,
82 * which involves a first PIPE_CONTROL flush which stalls the pipeline...
84 brw_emit_pipe_control_flush(brw
,
85 PIPE_CONTROL_DATA_CACHE_FLUSH
|
86 PIPE_CONTROL_NO_WRITE
|
87 PIPE_CONTROL_CS_STALL
);
89 /* ...followed by a second pipelined PIPE_CONTROL that initiates
90 * invalidation of the relevant caches. Note that because RO invalidation
91 * happens at the top of the pipeline (i.e. right away as the PIPE_CONTROL
92 * command is processed by the CS) we cannot combine it with the previous
93 * stalling flush as the hardware documentation suggests, because that
94 * would cause the CS to stall on previous rendering *after* RO
95 * invalidation and wouldn't prevent the RO caches from being polluted by
96 * concurrent rendering before the stall completes. This intentionally
97 * doesn't implement the SKL+ hardware workaround suggesting to enable CS
98 * stall on PIPE_CONTROLs with the texture cache invalidation bit set for
99 * GPGPU workloads because the previous and subsequent PIPE_CONTROLs
100 * already guarantee that there is no concurrent GPGPU kernel execution
101 * (see SKL HSD 2132585).
103 brw_emit_pipe_control_flush(brw
,
104 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
105 PIPE_CONTROL_CONST_CACHE_INVALIDATE
|
106 PIPE_CONTROL_INSTRUCTION_INVALIDATE
|
107 PIPE_CONTROL_STATE_CACHE_INVALIDATE
|
108 PIPE_CONTROL_NO_WRITE
);
110 /* Now send a third stalling flush to make sure that invalidation is
111 * complete when the L3 configuration registers are modified.
113 brw_emit_pipe_control_flush(brw
,
114 PIPE_CONTROL_DATA_CACHE_FLUSH
|
115 PIPE_CONTROL_NO_WRITE
|
116 PIPE_CONTROL_CS_STALL
);
119 assert(!cfg
->n
[GEN_L3P_IS
] && !cfg
->n
[GEN_L3P_C
] && !cfg
->n
[GEN_L3P_T
]);
122 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (3 - 2));
124 /* Set up the L3 partitioning. */
125 OUT_BATCH(GEN8_L3CNTLREG
);
126 OUT_BATCH((has_slm
? GEN8_L3CNTLREG_SLM_ENABLE
: 0) |
127 SET_FIELD(cfg
->n
[GEN_L3P_URB
], GEN8_L3CNTLREG_URB_ALLOC
) |
128 SET_FIELD(cfg
->n
[GEN_L3P_RO
], GEN8_L3CNTLREG_RO_ALLOC
) |
129 SET_FIELD(cfg
->n
[GEN_L3P_DC
], GEN8_L3CNTLREG_DC_ALLOC
) |
130 SET_FIELD(cfg
->n
[GEN_L3P_ALL
], GEN8_L3CNTLREG_ALL_ALLOC
));
135 assert(!cfg
->n
[GEN_L3P_ALL
]);
137 /* When enabled SLM only uses a portion of the L3 on half of the banks,
138 * the matching space on the remaining banks has to be allocated to a
139 * client (URB for all validated configurations) set to the
140 * lower-bandwidth 2-bank address hashing mode.
142 const bool urb_low_bw
= has_slm
&& !brw
->is_baytrail
;
143 assert(!urb_low_bw
|| cfg
->n
[GEN_L3P_URB
] == cfg
->n
[GEN_L3P_SLM
]);
145 /* Minimum number of ways that can be allocated to the URB. */
146 const unsigned n0_urb
= (brw
->is_baytrail
? 32 : 0);
147 assert(cfg
->n
[GEN_L3P_URB
] >= n0_urb
);
150 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (7 - 2));
152 /* Demote any clients with no ways assigned to LLC. */
153 OUT_BATCH(GEN7_L3SQCREG1
);
154 OUT_BATCH((brw
->is_haswell
? HSW_L3SQCREG1_SQGHPCI_DEFAULT
:
155 brw
->is_baytrail
? VLV_L3SQCREG1_SQGHPCI_DEFAULT
:
156 IVB_L3SQCREG1_SQGHPCI_DEFAULT
) |
157 (has_dc
? 0 : GEN7_L3SQCREG1_CONV_DC_UC
) |
158 (has_is
? 0 : GEN7_L3SQCREG1_CONV_IS_UC
) |
159 (has_c
? 0 : GEN7_L3SQCREG1_CONV_C_UC
) |
160 (has_t
? 0 : GEN7_L3SQCREG1_CONV_T_UC
));
162 /* Set up the L3 partitioning. */
163 OUT_BATCH(GEN7_L3CNTLREG2
);
164 OUT_BATCH((has_slm
? GEN7_L3CNTLREG2_SLM_ENABLE
: 0) |
165 SET_FIELD(cfg
->n
[GEN_L3P_URB
] - n0_urb
, GEN7_L3CNTLREG2_URB_ALLOC
) |
166 (urb_low_bw
? GEN7_L3CNTLREG2_URB_LOW_BW
: 0) |
167 SET_FIELD(cfg
->n
[GEN_L3P_ALL
], GEN7_L3CNTLREG2_ALL_ALLOC
) |
168 SET_FIELD(cfg
->n
[GEN_L3P_RO
], GEN7_L3CNTLREG2_RO_ALLOC
) |
169 SET_FIELD(cfg
->n
[GEN_L3P_DC
], GEN7_L3CNTLREG2_DC_ALLOC
));
170 OUT_BATCH(GEN7_L3CNTLREG3
);
171 OUT_BATCH(SET_FIELD(cfg
->n
[GEN_L3P_IS
], GEN7_L3CNTLREG3_IS_ALLOC
) |
172 SET_FIELD(cfg
->n
[GEN_L3P_C
], GEN7_L3CNTLREG3_C_ALLOC
) |
173 SET_FIELD(cfg
->n
[GEN_L3P_T
], GEN7_L3CNTLREG3_T_ALLOC
));
177 if (brw
->is_haswell
&& brw
->screen
->cmd_parser_version
>= 4) {
178 /* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
179 * them disabled to avoid crashing the system hard.
182 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (5 - 2));
183 OUT_BATCH(HSW_SCRATCH1
);
184 OUT_BATCH(has_dc
? 0 : HSW_SCRATCH1_L3_ATOMIC_DISABLE
);
185 OUT_BATCH(HSW_ROW_CHICKEN3
);
186 OUT_BATCH(REG_MASK(HSW_ROW_CHICKEN3_L3_ATOMIC_DISABLE
) |
187 (has_dc
? 0 : HSW_ROW_CHICKEN3_L3_ATOMIC_DISABLE
));
194 * Update the URB size in the context state for the specified L3
198 update_urb_size(struct brw_context
*brw
, const struct gen_l3_config
*cfg
)
200 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
201 const unsigned sz
= gen_get_l3_config_urb_size(devinfo
, cfg
);
203 if (brw
->urb
.size
!= sz
) {
205 brw
->ctx
.NewDriverState
|= BRW_NEW_URB_SIZE
;
210 emit_l3_state(struct brw_context
*brw
)
212 const struct gen_l3_weights w
= get_pipeline_state_l3_weights(brw
);
213 const float dw
= gen_diff_l3_weights(w
, gen_get_l3_config_weights(brw
->l3
.config
));
214 /* The distance between any two compatible weight vectors cannot exceed two
215 * due to the triangle inequality.
217 const float large_dw_threshold
= 2.0;
218 /* Somewhat arbitrary, simply makes sure that there will be no repeated
219 * transitions to the same L3 configuration, could probably do better here.
221 const float small_dw_threshold
= 0.5;
222 /* If we're emitting a new batch the caches should already be clean and the
223 * transition should be relatively cheap, so it shouldn't hurt much to use
224 * the smaller threshold. Otherwise use the larger threshold so that we
225 * only reprogram the L3 mid-batch if the most recently programmed
226 * configuration is incompatible with the current pipeline state.
228 const float dw_threshold
= (brw
->ctx
.NewDriverState
& BRW_NEW_BATCH
?
229 small_dw_threshold
: large_dw_threshold
);
231 if (dw
> dw_threshold
&& brw
->can_do_pipelined_register_writes
) {
232 const struct gen_l3_config
*const cfg
=
233 gen_get_l3_config(&brw
->screen
->devinfo
, w
);
235 setup_l3_config(brw
, cfg
);
236 update_urb_size(brw
, cfg
);
237 brw
->l3
.config
= cfg
;
239 if (unlikely(INTEL_DEBUG
& DEBUG_L3
)) {
240 fprintf(stderr
, "L3 config transition (%f > %f): ", dw
, dw_threshold
);
241 gen_dump_l3_config(cfg
, stderr
);
246 const struct brw_tracked_state gen7_l3_state
= {
249 .brw
= BRW_NEW_BATCH
|
251 BRW_NEW_CS_PROG_DATA
|
252 BRW_NEW_FS_PROG_DATA
|
253 BRW_NEW_GS_PROG_DATA
|
254 BRW_NEW_VS_PROG_DATA
,
256 .emit
= emit_l3_state
260 * Hack to restore the default L3 configuration.
262 * This will be called at the end of every batch in order to reset the L3
263 * configuration to the default values for the time being until the kernel is
264 * fixed. Until kernel commit 6702cf16e0ba8b0129f5aa1b6609d4e9c70bc13b
265 * (included in v4.1) we would set the MI_RESTORE_INHIBIT bit when submitting
266 * batch buffers for the default context used by the DDX, which meant that any
267 * context state changed by the GL would leak into the DDX, the assumption
268 * being that the DDX would initialize any state it cares about manually. The
269 * DDX is however not careful enough to program an L3 configuration
270 * explicitly, and it makes assumptions about it (URB size) which won't hold
271 * and cause it to misrender if we let our L3 set-up to leak into the DDX.
273 * Since v4.1 of the Linux kernel the default context is saved and restored
274 * normally, so it's far less likely for our L3 programming to interfere with
275 * other contexts -- In fact restoring the default L3 configuration at the end
276 * of the batch will be redundant most of the time. A kind of state leak is
277 * still possible though if the context making assumptions about L3 state is
278 * created immediately after our context was active (e.g. without the DDX
279 * default context being scheduled in between) because at present the DRM
280 * doesn't fully initialize the contents of newly created contexts and instead
281 * sets the MI_RESTORE_INHIBIT flag causing it to inherit the state from the
282 * last active context.
284 * It's possible to realize such a scenario if, say, an X server (or a GL
285 * application using an outdated non-L3-aware Mesa version) is started while
286 * another GL application is running and happens to have modified the L3
287 * configuration, or if no X server is running at all and a GL application
288 * using a non-L3-aware Mesa version is started after another GL application
289 * ran and modified the L3 configuration -- The latter situation can actually
290 * be reproduced easily on IVB in our CI system.
293 gen7_restore_default_l3_config(struct brw_context
*brw
)
295 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
296 const struct gen_l3_config
*const cfg
= gen_get_default_l3_config(devinfo
);
298 if (cfg
!= brw
->l3
.config
&& brw
->can_do_pipelined_register_writes
) {
299 setup_l3_config(brw
, cfg
);
300 update_urb_size(brw
, cfg
);
301 brw
->l3
.config
= cfg
;