1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
|
From 9bb867075fc4b0b7efc1640dc4cdd5b70b482ff1 Mon Sep 17 00:00:00 2001
From: Jonas Pfeil <pfeiljonas@gmx.de>
Date: Tue, 8 Nov 2016 00:18:39 +0100
Subject: [PATCH] drm/vc4: Add fragment shader threading support
FS threading brings performance improvements of 0-20% in glmark2.
The validation code checks for thread switch signals and ensures that
the registers of the other thread are not touched, and that our clamps
are not live across thread switches. It also checks that the
threading and branching instructions do not interfere.
(Original patch by Jonas, changes by anholt for style cleanup,
removing validation the kernel doesn't need to do, and adding the flag
for userspace).
v2: Minor style fixes from checkpatch.
Signed-off-by: Jonas Pfeil <pfeiljonas@gmx.de>
Signed-off-by: Eric Anholt <eric@anholt.net>
(cherry picked from commit c778cc5df944291dcdb1ca7a6bb781fbc22550c5)
---
drivers/gpu/drm/vc4/vc4_drv.c | 1 +
drivers/gpu/drm/vc4/vc4_drv.h | 2 +
drivers/gpu/drm/vc4/vc4_validate.c | 17 +++++---
drivers/gpu/drm/vc4/vc4_validate_shaders.c | 63 ++++++++++++++++++++++++++++++
include/uapi/drm/vc4_drm.h | 1 +
5 files changed, 79 insertions(+), 5 deletions(-)
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -107,6 +107,7 @@ static int vc4_get_param_ioctl(struct dr
break;
case DRM_VC4_PARAM_SUPPORTS_BRANCHES:
case DRM_VC4_PARAM_SUPPORTS_ETC1:
+ case DRM_VC4_PARAM_SUPPORTS_THREADED_FS:
args->value = true;
break;
default:
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -395,6 +395,8 @@ struct vc4_validated_shader_info {
uint32_t num_uniform_addr_offsets;
uint32_t *uniform_addr_offsets;
+
+ bool is_threaded;
};
/**
--- a/drivers/gpu/drm/vc4/vc4_validate.c
+++ b/drivers/gpu/drm/vc4/vc4_validate.c
@@ -789,11 +789,6 @@ validate_gl_shader_rec(struct drm_device
exec->shader_rec_v += roundup(packet_size, 16);
exec->shader_rec_size -= packet_size;
- if (!(*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD)) {
- DRM_ERROR("Multi-threaded fragment shaders not supported.\n");
- return -EINVAL;
- }
-
for (i = 0; i < shader_reloc_count; i++) {
if (src_handles[i] > exec->bo_count) {
DRM_ERROR("Shader handle %d too big\n", src_handles[i]);
@@ -810,6 +805,18 @@ validate_gl_shader_rec(struct drm_device
return -EINVAL;
}
+ if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) !=
+ to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) {
+ DRM_ERROR("Thread mode of CL and FS do not match\n");
+ return -EINVAL;
+ }
+
+ if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded ||
+ to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) {
+ DRM_ERROR("cs and vs cannot be threaded\n");
+ return -EINVAL;
+ }
+
for (i = 0; i < shader_reloc_count; i++) {
struct vc4_validated_shader_info *validated_shader;
uint32_t o = shader_reloc_offsets[i];
--- a/drivers/gpu/drm/vc4/vc4_validate_shaders.c
+++ b/drivers/gpu/drm/vc4/vc4_validate_shaders.c
@@ -83,6 +83,13 @@ struct vc4_shader_validation_state {
* basic blocks.
*/
bool needs_uniform_address_for_loop;
+
+ /* Set when we find an instruction writing the top half of the
+ * register files. If we allowed writing the unusable regs in
+ * a threaded shader, then the other shader running on our
+ * QPU's clamp validation would be invalid.
+ */
+ bool all_registers_used;
};
static uint32_t
@@ -119,6 +126,13 @@ raddr_add_a_to_live_reg_index(uint64_t i
}
static bool
+live_reg_is_upper_half(uint32_t lri)
+{
+ return (lri >= 16 && lri < 32) ||
+ (lri >= 32 + 16 && lri < 32 + 32);
+}
+
+static bool
is_tmu_submit(uint32_t waddr)
{
return (waddr == QPU_W_TMU0_S ||
@@ -390,6 +404,9 @@ check_reg_write(struct vc4_validated_sha
} else {
validation_state->live_immediates[lri] = ~0;
}
+
+ if (live_reg_is_upper_half(lri))
+ validation_state->all_registers_used = true;
}
switch (waddr) {
@@ -598,6 +615,11 @@ check_instruction_reads(struct vc4_valid
}
}
+ if ((raddr_a >= 16 && raddr_a < 32) ||
+ (raddr_b >= 16 && raddr_b < 32 && sig != QPU_SIG_SMALL_IMM)) {
+ validation_state->all_registers_used = true;
+ }
+
return true;
}
@@ -753,6 +775,7 @@ vc4_validate_shader(struct drm_gem_cma_o
{
bool found_shader_end = false;
int shader_end_ip = 0;
+ uint32_t last_thread_switch_ip = -3;
uint32_t ip;
struct vc4_validated_shader_info *validated_shader = NULL;
struct vc4_shader_validation_state validation_state;
@@ -785,6 +808,17 @@ vc4_validate_shader(struct drm_gem_cma_o
if (!vc4_handle_branch_target(&validation_state))
goto fail;
+ if (ip == last_thread_switch_ip + 3) {
+ /* Reset r0-r3 live clamp data */
+ int i;
+
+ for (i = 64; i < LIVE_REG_COUNT; i++) {
+ validation_state.live_min_clamp_offsets[i] = ~0;
+ validation_state.live_max_clamp_regs[i] = false;
+ validation_state.live_immediates[i] = ~0;
+ }
+ }
+
switch (sig) {
case QPU_SIG_NONE:
case QPU_SIG_WAIT_FOR_SCOREBOARD:
@@ -794,6 +828,8 @@ vc4_validate_shader(struct drm_gem_cma_o
case QPU_SIG_LOAD_TMU1:
case QPU_SIG_PROG_END:
case QPU_SIG_SMALL_IMM:
+ case QPU_SIG_THREAD_SWITCH:
+ case QPU_SIG_LAST_THREAD_SWITCH:
if (!check_instruction_writes(validated_shader,
&validation_state)) {
DRM_ERROR("Bad write at ip %d\n", ip);
@@ -809,6 +845,18 @@ vc4_validate_shader(struct drm_gem_cma_o
shader_end_ip = ip;
}
+ if (sig == QPU_SIG_THREAD_SWITCH ||
+ sig == QPU_SIG_LAST_THREAD_SWITCH) {
+ validated_shader->is_threaded = true;
+
+ if (ip < last_thread_switch_ip + 3) {
+ DRM_ERROR("Thread switch too soon after "
+ "last switch at ip %d\n", ip);
+ goto fail;
+ }
+ last_thread_switch_ip = ip;
+ }
+
break;
case QPU_SIG_LOAD_IMM:
@@ -823,6 +871,13 @@ vc4_validate_shader(struct drm_gem_cma_o
if (!check_branch(inst, validated_shader,
&validation_state, ip))
goto fail;
+
+ if (ip < last_thread_switch_ip + 3) {
+ DRM_ERROR("Branch in thread switch at ip %d",
+ ip);
+ goto fail;
+ }
+
break;
default:
DRM_ERROR("Unsupported QPU signal %d at "
@@ -844,6 +899,14 @@ vc4_validate_shader(struct drm_gem_cma_o
goto fail;
}
+ /* Might corrupt other thread */
+ if (validated_shader->is_threaded &&
+ validation_state.all_registers_used) {
+ DRM_ERROR("Shader uses threading, but uses the upper "
+ "half of the registers, too\n");
+ goto fail;
+ }
+
/* If we did a backwards branch and we haven't emitted a uniforms
* reset since then, we still need the uniforms stream to have the
* uniforms address available so that the backwards branch can do its
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -287,6 +287,7 @@ struct drm_vc4_get_hang_state {
#define DRM_VC4_PARAM_V3D_IDENT2 2
#define DRM_VC4_PARAM_SUPPORTS_BRANCHES 3
#define DRM_VC4_PARAM_SUPPORTS_ETC1 4
+#define DRM_VC4_PARAM_SUPPORTS_THREADED_FS 5
struct drm_vc4_get_param {
__u32 param;
|
