Mesa (master): freedreno/ir3: limit # of tex prefetch by shader size

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: f351e1d137603213b5daacece5ff67ad0786d982
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=f351e1d137603213b5daacece5ff67ad0786d982

Author: Rob Clark <robdclark at chromium.org>
Date:   Mon May  4 15:13:20 2020 -0700

freedreno/ir3: limit # of tex prefetch by shader size

It seems for short frag shaders, too much prefetch can be detrimental.
I think what we *really* want to do is decide after pre-RA sched, when
we also know about nop's and what the actual ir3 instruction count is.
But that will require re-working how prefetch lowering works.  For now
this is a super crude heuristic to attempt to approximate a good
solution.

Signed-off-by: Rob Clark <robdclark at chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>

---

 src/freedreno/ir3/ir3_compiler_nir.c |  2 +-
 src/freedreno/ir3/ir3_context.c      | 37 ++++++++++++++++++++++++++++++++++++
 src/freedreno/ir3/ir3_context.h      |  2 ++
 3 files changed, 40 insertions(+), 1 deletion(-)

diff --git a/src/freedreno/ir3/ir3_compiler_nir.c b/src/freedreno/ir3/ir3_compiler_nir.c
index dc2492e242a..76f2f7525bf 100644
--- a/src/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/freedreno/ir3/ir3_compiler_nir.c
@@ -2261,7 +2261,7 @@ emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
 		compile_assert(ctx, nir_tex_instr_src_index(tex, nir_tex_src_texture_offset) < 0);
 		compile_assert(ctx, nir_tex_instr_src_index(tex, nir_tex_src_sampler_offset) < 0);
 
-		if (ctx->so->num_sampler_prefetch < IR3_MAX_SAMPLER_PREFETCH) {
+		if (ctx->so->num_sampler_prefetch < ctx->prefetch_limit) {
 			opc = OPC_META_TEX_PREFETCH;
 			ctx->so->num_sampler_prefetch++;
 			break;
diff --git a/src/freedreno/ir3/ir3_context.c b/src/freedreno/ir3/ir3_context.c
index 38f870321a7..6b6f485af05 100644
--- a/src/freedreno/ir3/ir3_context.c
+++ b/src/freedreno/ir3/ir3_context.c
@@ -114,6 +114,43 @@ ir3_context_init(struct ir3_compiler *compiler,
 
 	NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);
 
+	/* Super crude heuristic to limit # of tex prefetch in small
+	 * shaders.  This completely ignores loops.. but that's really
+	 * not the worst of it's problems.  (A frag shader that has
+	 * loops is probably going to be big enough to not trigger a
+	 * lower threshold.)
+	 *
+	 *   1) probably want to do this in terms of ir3 instructions
+	 *   2) probably really want to decide this after scheduling
+	 *      (or at least pre-RA sched) so we have a rough idea about
+	 *      nops, and don't count things that get cp'd away
+	 *   3) blob seems to use higher thresholds with a mix of more
+	 *      SFU instructions.  Which partly makes sense, more SFU
+	 *      instructions probably means you want to get the real
+	 *      shader started sooner, but that considers where in the
+	 *      shader the SFU instructions are, which blob doesn't seem
+	 *      to do.
+	 *
+	 * This uses more conservative thresholds assuming a more alu
+	 * than sfu heavy instruction mix.
+	 */
+	if (so->type == MESA_SHADER_FRAGMENT) {
+		nir_function_impl *fxn = nir_shader_get_entrypoint(ctx->s);
+
+		unsigned instruction_count = 0;
+		nir_foreach_block (block, fxn) {
+			instruction_count += exec_list_length(&block->instr_list);
+		}
+
+		if (instruction_count < 50) {
+			ctx->prefetch_limit = 2;
+		} else if (instruction_count < 70) {
+			ctx->prefetch_limit = 3;
+		} else {
+			ctx->prefetch_limit = IR3_MAX_SAMPLER_PREFETCH;
+		}
+	}
+
 	if (shader_debug_enabled(so->type)) {
 		fprintf(stdout, "NIR (final form) for %s shader %s:\n",
 			ir3_shader_stage(so), so->shader->nir->info.name);
diff --git a/src/freedreno/ir3/ir3_context.h b/src/freedreno/ir3/ir3_context.h
index d1de1a67a26..659303de9e1 100644
--- a/src/freedreno/ir3/ir3_context.h
+++ b/src/freedreno/ir3/ir3_context.h
@@ -146,6 +146,8 @@ struct ir3_context {
 
 	unsigned max_texture_index;
 
+	unsigned prefetch_limit;
+
 	/* set if we encounter something we can't handle yet, so we
 	 * can bail cleanly and fallback to TGSI compiler f/e
 	 */