2 * Copyright © 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
21 * DEALINGS IN THE SOFTWARE.
24 #include "util/ralloc.h"
25 #include "brw_context.h"
29 #include "brw_shader.h"
30 #include "intel_mipmap_tree.h"
31 #include "intel_batchbuffer.h"
32 #include "brw_state.h"
35 get_cs_thread_count(const struct brw_cs_prog_data
*cs_prog_data
)
37 const unsigned simd_size
= cs_prog_data
->simd_size
;
38 unsigned group_size
= cs_prog_data
->local_size
[0] *
39 cs_prog_data
->local_size
[1] * cs_prog_data
->local_size
[2];
41 return (group_size
+ simd_size
- 1) / simd_size
;
46 brw_upload_cs_state(struct brw_context
*brw
)
48 if (!brw
->cs
.prog_data
)
52 uint32_t *desc
= (uint32_t*) brw_state_batch(brw
, AUB_TRACE_SURFACE_STATE
,
54 struct gl_program
*prog
= (struct gl_program
*) brw
->compute_program
;
55 struct brw_stage_state
*stage_state
= &brw
->cs
.base
;
56 struct brw_cs_prog_data
*cs_prog_data
= brw
->cs
.prog_data
;
57 struct brw_stage_prog_data
*prog_data
= &cs_prog_data
->base
;
59 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
) {
60 brw
->vtbl
.emit_buffer_surface_state(
61 brw
, &stage_state
->surf_offset
[
62 prog_data
->binding_table
.shader_time_start
],
63 brw
->shader_time
.bo
, 0, BRW_SURFACEFORMAT_RAW
,
64 brw
->shader_time
.bo
->size
, 1, true);
67 uint32_t *bind
= (uint32_t*) brw_state_batch(brw
, AUB_TRACE_BINDING_TABLE
,
68 prog_data
->binding_table
.size_bytes
,
69 32, &stage_state
->bind_bo_offset
);
71 unsigned local_id_dwords
= 0;
73 if (prog
->SystemValuesRead
& SYSTEM_BIT_LOCAL_INVOCATION_ID
)
74 local_id_dwords
= cs_prog_data
->local_invocation_id_regs
* 8;
76 unsigned push_constant_data_size
=
77 (prog_data
->nr_params
+ local_id_dwords
) * sizeof(gl_constant_value
);
78 unsigned reg_aligned_constant_size
= ALIGN(push_constant_data_size
, 32);
79 unsigned push_constant_regs
= reg_aligned_constant_size
/ 32;
80 unsigned threads
= get_cs_thread_count(cs_prog_data
);
82 uint32_t dwords
= brw
->gen
< 8 ? 8 : 9;
84 OUT_BATCH(MEDIA_VFE_STATE
<< 16 | (dwords
- 2));
86 if (prog_data
->total_scratch
) {
88 OUT_RELOC64(stage_state
->scratch_bo
,
89 I915_GEM_DOMAIN_RENDER
, I915_GEM_DOMAIN_RENDER
,
90 ffs(prog_data
->total_scratch
) - 11);
92 OUT_RELOC(stage_state
->scratch_bo
,
93 I915_GEM_DOMAIN_RENDER
, I915_GEM_DOMAIN_RENDER
,
94 ffs(prog_data
->total_scratch
) - 11);
101 const uint32_t vfe_num_urb_entries
= brw
->gen
>= 8 ? 2 : 0;
102 const uint32_t vfe_gpgpu_mode
=
103 brw
->gen
== 7 ? SET_FIELD(1, GEN7_MEDIA_VFE_STATE_GPGPU_MODE
) : 0;
104 OUT_BATCH(SET_FIELD(brw
->max_cs_threads
- 1, MEDIA_VFE_STATE_MAX_THREADS
) |
105 SET_FIELD(vfe_num_urb_entries
, MEDIA_VFE_STATE_URB_ENTRIES
) |
106 SET_FIELD(1, MEDIA_VFE_STATE_RESET_GTW_TIMER
) |
107 SET_FIELD(1, MEDIA_VFE_STATE_BYPASS_GTW
) |
111 const uint32_t vfe_urb_allocation
= brw
->gen
>= 8 ? 2 : 0;
113 /* We are uploading duplicated copies of push constant uniforms for each
114 * thread. Although the local id data needs to vary per thread, it won't
115 * change for other uniform data. Unfortunately this duplication is
116 * required for gen7. As of Haswell, this duplication can be avoided, but
117 * this older mechanism with duplicated data continues to work.
119 * FINISHME: As of Haswell, we could make use of the
120 * INTERFACE_DESCRIPTOR_DATA "Cross-Thread Constant Data Read Length" field
121 * to only store one copy of uniform data.
123 * FINISHME: Broadwell adds a new alternative "Indirect Payload Storage"
124 * which is described in the GPGPU_WALKER command and in the Broadwell PRM
125 * Volume 7: 3D Media GPGPU, under Media GPGPU Pipeline => Mode of
126 * Operations => GPGPU Mode => Indirect Payload Storage.
128 * Note: The constant data is built in brw_upload_cs_push_constants below.
130 const uint32_t vfe_curbe_allocation
= push_constant_regs
* threads
;
131 OUT_BATCH(SET_FIELD(vfe_urb_allocation
, MEDIA_VFE_STATE_URB_ALLOC
) |
132 SET_FIELD(vfe_curbe_allocation
, MEDIA_VFE_STATE_CURBE_ALLOC
));
138 if (reg_aligned_constant_size
> 0) {
140 OUT_BATCH(MEDIA_CURBE_LOAD
<< 16 | (4 - 2));
142 OUT_BATCH(reg_aligned_constant_size
* threads
);
143 OUT_BATCH(stage_state
->push_const_offset
);
147 /* BRW_NEW_SURFACES and BRW_NEW_*_CONSTBUF */
148 memcpy(bind
, stage_state
->surf_offset
,
149 prog_data
->binding_table
.size_bytes
);
151 memset(desc
, 0, 8 * 4);
154 desc
[dw
++] = brw
->cs
.base
.prog_offset
;
156 desc
[dw
++] = 0; /* Kernel Start Pointer High */
158 desc
[dw
++] = stage_state
->sampler_offset
|
159 ((stage_state
->sampler_count
+ 3) / 4);
160 desc
[dw
++] = stage_state
->bind_bo_offset
;
161 desc
[dw
++] = SET_FIELD(push_constant_regs
, MEDIA_CURBE_READ_LENGTH
);
162 const uint32_t media_threads
=
164 SET_FIELD(threads
, GEN8_MEDIA_GPGPU_THREAD_COUNT
) :
165 SET_FIELD(threads
, MEDIA_GPGPU_THREAD_COUNT
);
166 assert(threads
<= brw
->max_cs_threads
);
168 SET_FIELD(cs_prog_data
->uses_barrier
, MEDIA_BARRIER_ENABLE
) |
172 OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD
<< 16 | (4 - 2));
179 const struct brw_tracked_state brw_cs_state
= {
181 .mesa
= _NEW_PROGRAM_CONSTANTS
,
182 .brw
= BRW_NEW_BATCH
|
183 BRW_NEW_CS_PROG_DATA
|
184 BRW_NEW_PUSH_CONSTANT_ALLOCATION
|
187 .emit
= brw_upload_cs_state
192 * Creates a region containing the push constants for the CS on gen7+.
194 * Push constants are constant values (such as GLSL uniforms) that are
195 * pre-loaded into a shader stage's register space at thread spawn time.
197 * For other stages, see brw_curbe.c:brw_upload_constant_buffer for the
198 * equivalent gen4/5 code and gen6_vs_state.c:gen6_upload_push_constants for
202 brw_upload_cs_push_constants(struct brw_context
*brw
,
203 const struct gl_program
*prog
,
204 const struct brw_cs_prog_data
*cs_prog_data
,
205 struct brw_stage_state
*stage_state
,
206 enum aub_state_struct_type type
)
208 struct gl_context
*ctx
= &brw
->ctx
;
209 const struct brw_stage_prog_data
*prog_data
=
210 (struct brw_stage_prog_data
*) cs_prog_data
;
211 unsigned local_id_dwords
= 0;
213 if (prog
->SystemValuesRead
& SYSTEM_BIT_LOCAL_INVOCATION_ID
)
214 local_id_dwords
= cs_prog_data
->local_invocation_id_regs
* 8;
216 /* Updates the ParamaterValues[i] pointers for all parameters of the
217 * basic type of PROGRAM_STATE_VAR.
219 /* XXX: Should this happen somewhere before to get our state flag set? */
220 _mesa_load_state_parameters(ctx
, prog
->Parameters
);
222 if (prog_data
->nr_params
== 0 && local_id_dwords
== 0) {
223 stage_state
->push_const_size
= 0;
225 gl_constant_value
*param
;
228 const unsigned push_constant_data_size
=
229 (local_id_dwords
+ prog_data
->nr_params
) * sizeof(gl_constant_value
);
230 const unsigned reg_aligned_constant_size
= ALIGN(push_constant_data_size
, 32);
231 const unsigned param_aligned_count
=
232 reg_aligned_constant_size
/ sizeof(*param
);
234 unsigned threads
= get_cs_thread_count(cs_prog_data
);
236 param
= (gl_constant_value
*)
237 brw_state_batch(brw
, type
,
238 reg_aligned_constant_size
* threads
,
239 32, &stage_state
->push_const_offset
);
242 STATIC_ASSERT(sizeof(gl_constant_value
) == sizeof(float));
244 brw_cs_fill_local_id_payload(cs_prog_data
, param
, threads
,
245 reg_aligned_constant_size
);
247 /* _NEW_PROGRAM_CONSTANTS */
248 for (t
= 0; t
< threads
; t
++) {
249 gl_constant_value
*next_param
=
250 ¶m
[t
* param_aligned_count
+ local_id_dwords
];
251 for (i
= 0; i
< prog_data
->nr_params
; i
++) {
252 next_param
[i
] = *prog_data
->param
[i
];
256 stage_state
->push_const_size
= ALIGN(prog_data
->nr_params
, 8) / 8;
262 gen7_upload_cs_push_constants(struct brw_context
*brw
)
264 struct brw_stage_state
*stage_state
= &brw
->cs
.base
;
266 /* BRW_NEW_COMPUTE_PROGRAM */
267 const struct brw_compute_program
*cp
=
268 (struct brw_compute_program
*) brw
->compute_program
;
271 /* CACHE_NEW_CS_PROG */
272 struct brw_cs_prog_data
*cs_prog_data
= brw
->cs
.prog_data
;
274 brw_upload_cs_push_constants(brw
, &cp
->program
.Base
, cs_prog_data
,
275 stage_state
, AUB_TRACE_WM_CONSTANTS
);
279 const struct brw_tracked_state gen7_cs_push_constants
= {
281 .mesa
= _NEW_PROGRAM_CONSTANTS
,
282 .brw
= BRW_NEW_BATCH
|
283 BRW_NEW_COMPUTE_PROGRAM
|
284 BRW_NEW_PUSH_CONSTANT_ALLOCATION
,
286 .emit
= gen7_upload_cs_push_constants
,