X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Ffreedreno%2Fir3%2Fir3_shader.c;h=519fafec900e22ffd3d2317cb9ba87ea6455bf20;hb=9ed33383a5b1399315cd62f119234ee2332ee76f;hp=78ed751a3f25d0c0724b0e39b2b6dddc9b0b695b;hpb=65660622a1f7284c457388e553867b239343d314;p=mesa.git

diff --git a/src/freedreno/ir3/ir3_shader.c b/src/freedreno/ir3/ir3_shader.c
index 78ed751a3f2..519fafec900 100644
--- a/src/freedreno/ir3/ir3_shader.c
+++ b/src/freedreno/ir3/ir3_shader.c
@@ -26,6 +26,7 @@
 
 #include "util/u_atomic.h"
 #include "util/u_string.h"
+#include "util/u_math.h"
 #include "util/u_memory.h"
 #include "util/format/u_format.h"
 
@@ -35,24 +36,14 @@
 #include "ir3_compiler.h"
 #include "ir3_nir.h"
 
+#include "disasm.h"
+
 int
 ir3_glsl_type_size(const struct glsl_type *type, bool bindless)
 {
 	return glsl_count_attribute_slots(type, false);
 }
 
-static void
-delete_variant(struct ir3_shader_variant *v)
-{
-	if (v->ir)
-		ir3_destroy(v->ir);
-	assert(!v->bo);
-	if (v->binning)
-		delete_variant(v->binning);
-	free(v->bin);
-	free(v);
-}
-
 /* for vertex shader, the inputs are loaded into registers before the shader
  * is executed, so max_regs from the shader instructions might not properly
  * reflect the # of registers actually used, especially in case passthrough
@@ -133,25 +124,26 @@ fixup_regfootprint(struct ir3_shader_variant *v)
  */
 void * ir3_shader_assemble(struct ir3_shader_variant *v)
 {
-	unsigned gpu_id = v->shader->compiler->gpu_id;
+	const struct ir3_compiler *compiler = v->shader->compiler;
 	void *bin;
 
 	bin = ir3_assemble(v);
 	if (!bin)
 		return NULL;
 
-	if (gpu_id >= 400) {
-		v->instrlen = v->info.sizedwords / (2 * 16);
-	} else {
-		v->instrlen = v->info.sizedwords / (2 * 4);
-	}
-
 	/* NOTE: if relative addressing is used, we set constlen in
 	 * the compiler (to worst-case value) since we don't know in
 	 * the assembler what the max addr reg value can be:
 	 */
 	v->constlen = MAX2(v->constlen, v->info.max_const + 1);
 
+	/* On a4xx and newer, constlen must be a multiple of 16 dwords even though
+	 * uploads are in units of 4 dwords. Round it up here to make calculations
+	 * regarding the shared constlen simpler.
+	 */
+	if (compiler->gpu_id >= 400)
+		v->constlen = align(v->constlen, 4);
+
 	fixup_regfootprint(v);
 
 	return bin;
@@ -175,17 +167,42 @@ assemble_variant(struct ir3_shader_variant *v)
 	v->ir = NULL;
 }
 
+static bool
+compile_variant(struct ir3_shader_variant *v)
+{
+	int ret = ir3_compile_shader_nir(v->shader->compiler, v);
+	if (ret) {
+		_debug_printf("compile failed! (%s:%s)", v->shader->nir->info.name,
+				v->shader->nir->info.label);
+		return false;
+	}
+
+	assemble_variant(v);
+	if (!v->bin) {
+		_debug_printf("assemble failed! (%s:%s)", v->shader->nir->info.name,
+				v->shader->nir->info.label);
+		return false;
+	}
+
+	return true;
+}
+
 /*
  * For creating normal shader variants, 'nonbinning' is NULL.  For
  * creating binning pass shader, it is link to corresponding normal
  * (non-binning) variant.
  */
 static struct ir3_shader_variant *
-create_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
+alloc_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
 		struct ir3_shader_variant *nonbinning)
 {
-	struct ir3_shader_variant *v = CALLOC_STRUCT(ir3_shader_variant);
-	int ret;
+	void *mem_ctx = shader;
+	/* hang the binning variant off it's non-binning counterpart instead
+	 * of the shader, to simplify the error cleanup paths
+	 */
+	if (nonbinning)
+		mem_ctx = nonbinning;
+	struct ir3_shader_variant *v = rzalloc_size(mem_ctx, sizeof(*v));
 
 	if (!v)
 		return NULL;
@@ -196,67 +213,75 @@ create_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
 	v->nonbinning = nonbinning;
 	v->key = *key;
 	v->type = shader->type;
+	v->mergedregs = shader->compiler->gpu_id >= 600;
 
-	if (shader->compiler->gpu_id >= 600) {
-		switch (v->type) {
-		case MESA_SHADER_TESS_CTRL:
-		case MESA_SHADER_TESS_EVAL:
-			v->mergedregs = false;
-			break;
-		case MESA_SHADER_VERTEX:
-		case MESA_SHADER_GEOMETRY:
-			/* For VS/GS, normally do mergedregs, but if there is tess
-			 * we need to not used MERGEDREGS
-			 */
-			v->mergedregs = !key->tessellation;
-			break;
-		default:
-			v->mergedregs = true;
+	if (!v->binning_pass)
+		v->const_state = rzalloc_size(v, sizeof(*v->const_state));
+
+	return v;
+}
+
+static bool
+needs_binning_variant(struct ir3_shader_variant *v)
+{
+	if ((v->type == MESA_SHADER_VERTEX) && ir3_has_binning_vs(&v->key))
+		return true;
+	return false;
+}
+
+static struct ir3_shader_variant *
+create_variant(struct ir3_shader *shader, const struct ir3_shader_key *key)
+{
+	struct ir3_shader_variant *v = alloc_variant(shader, key, NULL);
+
+	if (!v)
+		goto fail;
+
+	if (needs_binning_variant(v)) {
+		v->binning = alloc_variant(shader, key, v);
+		if (!v->binning)
+			goto fail;
+	}
+
+	if (ir3_disk_cache_retrieve(shader->compiler, v))
+		return v;
+
+	if (!shader->nir_finalized) {
+		ir3_nir_post_finalize(shader->compiler, shader->nir);
+
+		if (ir3_shader_debug & IR3_DBG_DISASM) {
+			printf("dump nir%d: type=%d", shader->id, shader->type);
+			nir_print_shader(shader->nir, stdout);
 		}
-	} else {
-		v->mergedregs = false;
+
+		shader->nir_finalized = true;
 	}
 
-	ret = ir3_compile_shader_nir(shader->compiler, v);
-	if (ret) {
-		debug_error("compile failed!");
+	if (!compile_variant(v))
 		goto fail;
-	}
 
-	assemble_variant(v);
-	if (!v->bin) {
-		debug_error("assemble failed!");
+	if (needs_binning_variant(v) && !compile_variant(v->binning))
 		goto fail;
-	}
+
+	ir3_disk_cache_store(shader->compiler, v);
 
 	return v;
 
 fail:
-	delete_variant(v);
+	ralloc_free(v);
 	return NULL;
 }
 
 static inline struct ir3_shader_variant *
-shader_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
-		bool *created)
+shader_variant(struct ir3_shader *shader, const struct ir3_shader_key *key)
 {
 	struct ir3_shader_variant *v;
 
-	*created = false;
-
 	for (v = shader->variants; v; v = v->next)
 		if (ir3_shader_key_equal(key, &v->key))
 			return v;
 
-	/* compile new variant if it doesn't exist already: */
-	v = create_variant(shader, key, NULL);
-	if (v) {
-		v->next = shader->variants;
-		shader->variants = v;
-		*created = true;
-	}
-
-	return v;
+	return NULL;
 }
 
 struct ir3_shader_variant *
@@ -264,17 +289,23 @@ ir3_shader_get_variant(struct ir3_shader *shader, const struct ir3_shader_key *k
 		bool binning_pass, bool *created)
 {
 	mtx_lock(&shader->variants_lock);
-	struct ir3_shader_variant *v =
-			shader_variant(shader, key, created);
-
-	if (v && binning_pass) {
-		if (!v->binning) {
-			v->binning = create_variant(shader, key, v);
+	struct ir3_shader_variant *v = shader_variant(shader, key);
+
+	if (!v) {
+		/* compile new variant if it doesn't exist already: */
+		v = create_variant(shader, key);
+		if (v) {
+			v->next = shader->variants;
+			shader->variants = v;
 			*created = true;
 		}
-		mtx_unlock(&shader->variants_lock);
-		return v->binning;
 	}
+
+	if (v && binning_pass) {
+		v = v->binning;
+		assert(v);
+	}
+
 	mtx_unlock(&shader->variants_lock);
 
 	return v;
@@ -283,16 +314,9 @@ ir3_shader_get_variant(struct ir3_shader *shader, const struct ir3_shader_key *k
 void
 ir3_shader_destroy(struct ir3_shader *shader)
 {
-	struct ir3_shader_variant *v, *t;
-	for (v = shader->variants; v; ) {
-		t = v;
-		v = v->next;
-		delete_variant(t);
-	}
-	free(shader->const_state.immediates);
 	ralloc_free(shader->nir);
 	mtx_destroy(&shader->variants_lock);
-	free(shader);
+	ralloc_free(shader);
 }
 
 /**
@@ -312,6 +336,10 @@ ir3_setup_used_key(struct ir3_shader *shader)
 	 */
 	key->has_per_samp = true;
 
+	key->safe_constlen = true;
+
+	key->ucp_enables = 0xff;
+
 	if (info->stage == MESA_SHADER_FRAGMENT) {
 		key->fsaturate_s = ~0;
 		key->fsaturate_t = ~0;
@@ -324,6 +352,10 @@ ir3_setup_used_key(struct ir3_shader *shader)
 			key->color_two_side = true;
 		}
 
+		if (info->inputs_read & VARYING_BIT_LAYER) {
+			key->layer_zero = true;
+		}
+
 		if ((info->outputs_written & ~(FRAG_RESULT_DEPTH |
 								FRAG_RESULT_STENCIL |
 								FRAG_RESULT_SAMPLE_MASK)) != 0) {
@@ -351,11 +383,80 @@ ir3_setup_used_key(struct ir3_shader *shader)
 	}
 }
 
+
+/* Given an array of constlen's, decrease some of them so that the sum stays
+ * within "combined_limit" while trying to fairly share the reduction. Returns
+ * a bitfield of which stages should be trimmed.
+ */
+static uint32_t
+trim_constlens(unsigned *constlens,
+			   unsigned first_stage, unsigned last_stage,
+			   unsigned combined_limit, unsigned safe_limit)
+{
+   unsigned cur_total = 0;
+   for (unsigned i = first_stage; i <= last_stage; i++) {
+      cur_total += constlens[i];
+   }
+
+   unsigned max_stage = 0;
+   unsigned max_const = 0;
+   uint32_t trimmed = 0;
+
+   while (cur_total > combined_limit) {
+	   for (unsigned i = first_stage; i <= last_stage; i++) {
+		   if (constlens[i] >= max_const) {
+			   max_stage = i;
+			   max_const = constlens[i];
+		   }
+	   }
+
+	   assert(max_const > safe_limit);
+	   trimmed |= 1 << max_stage;
+	   cur_total = cur_total - max_const + safe_limit;
+	   constlens[max_stage] = safe_limit;
+   }
+
+   return trimmed;
+}
+
+/* Figures out which stages in the pipeline to use the "safe" constlen for, in
+ * order to satisfy all shared constlen limits.
+ */
+uint32_t
+ir3_trim_constlen(struct ir3_shader_variant **variants,
+				  const struct ir3_compiler *compiler)
+{
+	unsigned constlens[MESA_SHADER_STAGES] = {};
+
+	for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+		if (variants[i])
+			constlens[i] = variants[i]->constlen;
+	}
+
+	uint32_t trimmed = 0;
+	STATIC_ASSERT(MESA_SHADER_STAGES <= 8 * sizeof(trimmed));
+
+	/* There are two shared limits to take into account, the geometry limit on
+	 * a6xx and the total limit. The frag limit on a6xx only matters for a
+	 * single stage, so it's always satisfied with the first variant.
+	 */
+	if (compiler->gpu_id >= 600) {
+		trimmed |=
+			trim_constlens(constlens, MESA_SHADER_VERTEX, MESA_SHADER_GEOMETRY,
+						   compiler->max_const_geom, compiler->max_const_safe);
+	}
+	trimmed |=
+		trim_constlens(constlens, MESA_SHADER_VERTEX, MESA_SHADER_FRAGMENT,
+					   compiler->max_const_pipeline, compiler->max_const_safe);
+
+	return trimmed;
+}
+
 struct ir3_shader *
 ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir,
 		unsigned reserved_user_consts, struct ir3_stream_output_info *stream_output)
 {
-	struct ir3_shader *shader = CALLOC_STRUCT(ir3_shader);
+	struct ir3_shader *shader = rzalloc_size(NULL, sizeof(*shader));
 
 	mtx_init(&shader->variants_lock, mtx_plain);
 	shader->compiler = compiler;
@@ -363,41 +464,10 @@ ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir,
 	shader->type = nir->info.stage;
 	if (stream_output)
 		memcpy(&shader->stream_output, stream_output, sizeof(shader->stream_output));
-	shader->const_state.num_reserved_user_consts = reserved_user_consts;
-
-	if (nir->info.stage == MESA_SHADER_GEOMETRY)
-		NIR_PASS_V(nir, ir3_nir_lower_gs);
-
-	NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size,
-			   (nir_lower_io_options)0);
-
-	if (compiler->gpu_id >= 600 &&
-			nir->info.stage == MESA_SHADER_FRAGMENT &&
-			!(ir3_shader_debug & IR3_DBG_NOFP16))
-		NIR_PASS_V(nir, nir_lower_mediump_outputs);
-
-	if (nir->info.stage == MESA_SHADER_FRAGMENT) {
-		/* NOTE: lower load_barycentric_at_sample first, since it
-		 * produces load_barycentric_at_offset:
-		 */
-		NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_sample);
-		NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_offset);
-
-		NIR_PASS_V(nir, ir3_nir_move_varying_inputs);
-	}
-
-	NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, false);
-
-	NIR_PASS_V(nir, nir_lower_amul, ir3_glsl_type_size);
-
-	/* do first pass optimization, ignoring the key: */
-	ir3_optimize_nir(shader, nir);
-
+	shader->num_reserved_user_consts = reserved_user_consts;
 	shader->nir = nir;
-	if (ir3_shader_debug & IR3_DBG_DISASM) {
-		printf("dump nir%d: type=%d", shader->id, shader->type);
-		nir_print_shader(shader->nir, stdout);
-	}
+
+	ir3_disk_cache_init_shader_key(compiler, shader);
 
 	ir3_setup_used_key(shader);
 
@@ -495,13 +565,13 @@ ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin, FILE *out)
 	}
 
 	const struct ir3_const_state *const_state = ir3_const_state(so);
-	for (i = 0; i < const_state->immediates_count; i++) {
+	for (i = 0; i < DIV_ROUND_UP(const_state->immediates_count, 4); i++) {
 		fprintf(out, "@const(c%d.x)\t", const_state->offsets.immediate + i);
 		fprintf(out, "0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
-				const_state->immediates[i].val[0],
-				const_state->immediates[i].val[1],
-				const_state->immediates[i].val[2],
-				const_state->immediates[i].val[3]);
+				const_state->immediates[i * 4 + 0],
+				const_state->immediates[i * 4 + 1],
+				const_state->immediates[i * 4 + 2],
+				const_state->immediates[i * 4 + 3]);
 	}
 
 	disasm_a3xx(bin, so->info.sizedwords, 0, out, ir->compiler->gpu_id);
@@ -545,6 +615,17 @@ ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin, FILE *out)
 			so->info.max_reg + 1,
 			so->constlen);
 
+	fprintf(out, "; %s prog %d/%d: %u cat0, %u cat1, %u cat2, %u cat3, %u cat4, %u cat5, %u cat6, %u cat7, \n",
+			type, so->shader->id, so->id,
+			so->info.instrs_per_cat[0],
+			so->info.instrs_per_cat[1],
+			so->info.instrs_per_cat[2],
+			so->info.instrs_per_cat[3],
+			so->info.instrs_per_cat[4],
+			so->info.instrs_per_cat[5],
+			so->info.instrs_per_cat[6],
+			so->info.instrs_per_cat[7]);
+
 	fprintf(out, "; %s prog %d/%d: %u sstall, %u (ss), %u (sy), %d max_sun, %d loops\n",
 			type, so->shader->id, so->id,
 			so->info.sstall,