summaryrefslogtreecommitdiffstats
path: root/src/gallium
diff options
context:
space:
mode:
authorRoland Scheidegger <[email protected]>2015-04-01 00:56:12 +0200
committerRoland Scheidegger <[email protected]>2015-04-01 00:56:12 +0200
commit611bd80f3b4972622a9f2c155c95d3241668e4d9 (patch)
tree039c81641c0fd302ec82eb5b6feb793e3188d19a /src/gallium
parent47c4b3854076adfe5a27b537f36262ac4ec4530d (diff)
gallivm: do some hack heuristic to disable texture functions
We've seen some cases where performance can hurt quite a bit. Technically, the more simple the function the more overhead there is for using a function for this (and the less benefits this provides). Hence don't do this if we expect the generated code to be simple. There's an even more important reason why this hurts performance, which is shaders reusing the same unit with some of the same inputs, as llvm cannot figure out the calculations are the same if they are performned in the function (even just reusing the same unit without any input being the same provides such optimization opportunities though not very much). This is something which would need to be handled by IPO passes however.
Diffstat (limited to 'src/gallium')
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c40
1 files changed, 40 insertions, 0 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index ff508e20b42..378c562bb9f 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -3297,7 +3297,47 @@ lp_build_sample_soa(const struct lp_static_texture_state *static_texture_state,
struct gallivm_state *gallivm,
const struct lp_sampler_params *params)
{
+ boolean use_tex_func = FALSE;
+
+ /*
+ * Do not use a function call if the sampling is "simple enough".
+ * We define this by
+ * a) format
+ * b) no mips (either one level only or no mip filter)
+ * No mips will definitely make the code smaller, though
+ * the format requirement is a bit iffy - there's some (SoA) formats
+ * which definitely generate less code. This does happen to catch
+ * some important cases though which are hurt quite a bit by using
+ * a call (though not really because of the call overhead but because
+ * they are reusing the same texture unit with some of the same
+ * parameters).
+ * Ideally we'd let llvm recognize this stuff by doing IPO passes.
+ */
+
if (USE_TEX_FUNC_CALL) {
+ const struct util_format_description *format_desc;
+ boolean simple_format;
+ boolean simple_tex;
+ enum lp_sampler_op_type op_type;
+ format_desc = util_format_description(static_texture_state->format);
+ simple_format = !format_desc ||
+ (util_format_is_rgba8_variant(format_desc) &&
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB);
+
+ op_type = (params->sample_key & LP_SAMPLER_OP_TYPE_MASK) >>
+ LP_SAMPLER_OP_TYPE_SHIFT;
+ simple_tex =
+ op_type != LP_SAMPLER_OP_TEXTURE ||
+ ((static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE ||
+ static_texture_state->level_zero_only == TRUE) &&
+ static_sampler_state->min_img_filter == static_sampler_state->mag_img_filter &&
+ (static_sampler_state->min_img_filter == PIPE_TEX_FILTER_NEAREST ||
+ static_sampler_state->min_img_filter == PIPE_TEX_FILTER_NEAREST));
+
+ use_tex_func = format_desc && !(simple_format && simple_tex);
+ }
+
+ if (use_tex_func) {
lp_build_sample_soa_func(gallivm,
static_texture_state,
static_sampler_state,