aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/freedreno
diff options
context:
space:
mode:
authorEric Anholt <[email protected]>2019-12-09 11:55:21 -0800
committerEric Anholt <[email protected]>2019-12-09 14:13:50 -0800
commit0470a03769715a3258c04e6b18848cef321bd213 (patch)
tree3ea1ea958e76193e9efaebcf75a2b8469fef4137 /src/gallium/drivers/freedreno
parent10da0a9d18a15044ca92553ed17adb7c445394c8 (diff)
freedreno: Track the set of UBOs to be uploaded in UBO analysis.
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 <[email protected]>
Diffstat (limited to 'src/gallium/drivers/freedreno')
-rw-r--r--src/gallium/drivers/freedreno/ir3/ir3_gallium.c40
1 files changed, 21 insertions, 19 deletions
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_gallium.c b/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
index 6ba01c62a53..e5c44521af6 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
@@ -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);
}
}