OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

zink: cap driver inlining using ssa allocation limit

Browse Source 
usually inlining is optimal for cpu drivers since the majority of
time is spent in the shaders, and any amount of reduction to shader code
will be optimal

if, however, the shaders are still really big after inlining, this improvement
will be negated by the insane amount of time spent doing stupid llvm optimizer
passes, so check post-inline size to see whether it exceeds a size threshold

lavapipe release build - 1700% improvement
* spec@arb_tessellation_shader@execution@variable-indexing@tcs-output-array-vec4-index-rd-after-barrier

before: 142.15s user 0.42s system 99% cpu 2:23.14 total

after: 8.60s user 0.07s system 99% cpu 8.677 total

fixes #6647

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16977>
This commit is contained in:
Mike Blumenkrantz
2022-06-10 12:43:45 -04:00
committed by Marge Bot
parent 44223e5f28
commit 11e55bce49
4 changed files with 13 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/gallium/drivers/zink/ci/zink-lvp-flakes.txt
View File

@@ -20,8 +20,3 @@ glx@glx-shader-sharing
spec@arb_fragment_program@no-newline
# glx-destroycontext-1: ../../src/xcb_conn.c:215: write_vec: Assertion `!c->out.queue_len' failed.
glx@glx-destroycontext-1
# #6647
spec@arb_tessellation_shader@execution@variable-indexing@tcs-output-array-vec4-index-rd-after-barrier
spec@arb_tessellation_shader@execution@variable-indexing@tcs-output-array-vec4-index-wr
spec@arb_tessellation_shader@execution@variable-indexing@tcs-output-array-vec4-index-wr-before-barrier

6
src/gallium/drivers/zink/zink_compiler.c
View File

@@ -2010,6 +2010,10 @@ zink_shader_compile(struct zink_screen *screen, struct zink_shader *zs, nir_shad
/* This must be done again. */
NIR_PASS_V(nir, nir_io_add_const_offset_to_base, nir_var_shader_in |
nir_var_shader_out);
nir_function_impl *impl = nir_shader_get_entrypoint(nir);
if (impl->ssa_alloc > ZINK_ALWAYS_INLINE_LIMIT)
zs->can_inline = false;
} else if (need_optimize)
optimize_nir(nir);
prune_io(nir);
@@ -2835,6 +2839,8 @@ zink_shader_create(struct zink_screen *screen, struct nir_shader *nir,
}
}
ret->can_inline = true;
return ret;
}

6
src/gallium/drivers/zink/zink_compiler.h
View File

@@ -38,6 +38,11 @@
#define ZINK_WORKGROUP_SIZE_Y 2
#define ZINK_WORKGROUP_SIZE_Z 3
/* stop inlining shaders if they have >limit ssa vals after inlining:
* recompile time isn't worth the inline
*/
#define ZINK_ALWAYS_INLINE_LIMIT 1500
struct pipe_screen;
struct zink_context;
struct zink_screen;
@@ -90,6 +95,7 @@ struct zink_shader {
uint32_t ubos_used; // bitfield of which ubo indices are used
uint32_t ssbos_used; // bitfield of which ssbo indices are used
bool bindless;
bool can_inline;
struct spirv_shader *spirv;
simple_mtx_t lock;

2
src/gallium/drivers/zink/zink_program.c
View File

@@ -106,7 +106,7 @@ get_shader_module_for_stage(struct zink_context *ctx, struct zink_screen *screen
}
if (ctx && zs->nir->info.num_inlinable_uniforms &&
ctx->inlinable_uniforms_valid_mask & BITFIELD64_BIT(pstage)) {
if (screen->is_cpu || prog->inlined_variant_count[pstage] < ZINK_MAX_INLINED_VARIANTS)
if (zs->can_inline && (screen->is_cpu || prog->inlined_variant_count[pstage] < ZINK_MAX_INLINED_VARIANTS))
inline_size = zs->nir->info.num_inlinable_uniforms;
else
key->inline_uniforms = false;