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

freedreno: Track the set of UBOs to be uploaded in UBO analysis.

Browse Source 
We were iterating over the entire 32-entry array each time, when we
can just use a bitset to know that we're only uploading from the first
entry normally.

Knocks ir3_emit_user_consts down from ~.5% of CPU to .1% on WebGL
fishtank.

Reviewed-by: Rob Clark <robdclark@chromium.org>
This commit is contained in:
Eric Anholt
2019-12-09 11:55:21 -08:00
parent 10da0a9d18
commit 0470a03769
3 changed files with 25 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
View File

@@ -246,6 +246,9 @@ ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader)
state->range[i].end = state->range[i].start + range_size;
}
offset += range_size;
if (state->range[i].start < state->range[i].end)
state->enabled |= 1 << i;
}
state->size = offset;

1
src/freedreno/ir3/ir3_shader.h
View File

@@ -615,6 +615,7 @@ struct ir3_ubo_range {
struct ir3_ubo_analysis_state {
struct ir3_ubo_range range[IR3_MAX_CONSTANT_BUFFERS];
uint32_t enabled;
uint32_t size;
uint32_t lower_count;
uint32_t cmdstream_size; /* for per-gen backend to stash required cmdstream size */

40
src/gallium/drivers/freedreno/ir3/ir3_gallium.c
View File

@@ -235,6 +235,11 @@ ir3_user_consts_size(struct ir3_ubo_analysis_state *state,
}
}
/**
* Uploads sub-ranges of UBOs to the hardware's constant buffer (UBO access
* outside of these ranges will be done using full UBO accesses in the
* shader).
*/
void
ir3_emit_user_consts(struct fd_screen *screen, const struct ir3_shader_variant *v,
struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
@@ -242,31 +247,28 @@ ir3_emit_user_consts(struct fd_screen *screen, const struct ir3_shader_variant *
struct ir3_ubo_analysis_state *state;
state = &v->shader->ubo_state;
for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
uint32_t i;
foreach_bit(i, state->enabled & constbuf->enabled_mask) {
struct pipe_constant_buffer *cb = &constbuf->cb[i];
if (state->range[i].start < state->range[i].end &&
constbuf->enabled_mask & (1 << i)) {
uint32_t size = state->range[i].end - state->range[i].start;
uint32_t offset = cb->buffer_offset + state->range[i].start;
uint32_t size = state->range[i].end - state->range[i].start;
uint32_t offset = cb->buffer_offset + state->range[i].start;
/* and even if the start of the const buffer is before
* first_immediate, the end may not be:
*/
size = MIN2(size, (16 * v->constlen) - state->range[i].offset);
/* and even if the start of the const buffer is before
* first_immediate, the end may not be:
*/
size = MIN2(size, (16 * v->constlen) - state->range[i].offset);
if (size == 0)
continue;
if (size == 0)
continue;
/* things should be aligned to vec4: */
debug_assert((state->range[i].offset % 16) == 0);
debug_assert((size % 16) == 0);
debug_assert((offset % 16) == 0);
/* things should be aligned to vec4: */
debug_assert((state->range[i].offset % 16) == 0);
debug_assert((size % 16) == 0);
debug_assert((offset % 16) == 0);
emit_const(screen, ring, v, state->range[i].offset / 4,
offset, size / 4, cb->user_buffer, cb->buffer);
}
emit_const(screen, ring, v, state->range[i].offset / 4,
offset, size / 4, cb->user_buffer, cb->buffer);
}
}