radeonsi: move some global shader cache flags to per-binary flags

Reviewed-by: Pierre-Eric Pelloux-Prayer <[email protected]>

4 files changed, 23 insertions, 20 deletions

diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index ac4a7aa3135..f534b5c2e5e 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -147,7 +147,7 @@ static void si_create_compute_state_async(void *job, int thread_index)
 	program->num_cs_user_data_dwords =
 		sel->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
 
-	void *ir_binary = si_get_ir_binary(sel, false);
+	void *ir_binary = si_get_ir_binary(sel, false, false);
 
 	/* Try to load the shader from the shader cache. */
 	mtx_lock(&sscreen->shader_cache_mutex);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index fef44836ea2..514ed9a0324 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -871,22 +871,9 @@ static void si_disk_cache_create(struct si_screen *sscreen)
 	disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
 
 	/* These flags affect shader compilation. */
-	#define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) |	\
-			   DBG(SI_SCHED) |			\
-			   DBG(GISEL) |				\
-			   DBG(W32_GE) |			\
-			   DBG(W32_PS) |			\
-			   DBG(W32_CS) |			\
-			   DBG(W64_GE) |			\
-			   DBG(W64_PS) |			\
-			   DBG(W64_CS))
+	#define ALL_FLAGS (DBG(SI_SCHED) | DBG(GISEL))
 	uint64_t shader_debug_flags = sscreen->debug_flags & ALL_FLAGS;
 
-	if (sscreen->options.enable_nir) {
-		STATIC_ASSERT((ALL_FLAGS & (1u << 31)) == 0);
-		shader_debug_flags |= 1u << 31;
-	}
-
 	/* Add the high bits of 32-bit addresses, which affects
 	 * how 32-bit addresses are expanded to 64 bits.
 	 */
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 323313764d8..11c7224b1d8 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -573,7 +573,7 @@ si_compute_fast_udiv_info32(uint32_t D, unsigned num_bits);
 void si_emit_dpbb_state(struct si_context *sctx);
 
 /* si_state_shaders.c */
-void *si_get_ir_binary(struct si_shader_selector *sel, bool as_ngg);
+void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es);
 bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
 				 struct si_shader *shader);
 bool si_shader_cache_insert_shader(struct si_screen *sscreen, void *ir_binary,
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 63014c9687c..c9800c00b25 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -45,7 +45,7 @@
  * Return the IR binary in a buffer. For TGSI the first 4 bytes contain its
  * size as integer.
  */
-void *si_get_ir_binary(struct si_shader_selector *sel, bool as_ngg)
+void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es)
 {
 	struct blob blob;
 	unsigned ir_size;
@@ -64,13 +64,27 @@ void *si_get_ir_binary(struct si_shader_selector *sel, bool as_ngg)
 		ir_size = blob.size;
 	}
 
+	/* These settings affect the compilation, but they are not derived
+	 * from the input shader IR.
+	 */
+	unsigned shader_variant_flags = 0;
+
+	if (ngg)
+		shader_variant_flags |= 1 << 0;
+	if (sel->nir)
+		shader_variant_flags |= 1 << 1;
+	if (si_get_wave_size(sel->screen, sel->type, ngg, es) == 32)
+		shader_variant_flags |= 1 << 2;
+	if (sel->force_correct_derivs_after_kill)
+		shader_variant_flags |= 1 << 3;
+
 	unsigned size = 4 + 4 + ir_size + sizeof(sel->so);
 	char *result = (char*)MALLOC(size);
 	if (!result)
 		return NULL;
 
 	((uint32_t*)result)[0] = size;
-	((uint32_t*)result)[1] = as_ngg;
+	((uint32_t*)result)[1] = shader_variant_flags;
 	memcpy(result + 8, ir_binary, ir_size);
 	memcpy(result + 8 + ir_size, &sel->so, sizeof(sel->so));
 
@@ -2462,8 +2476,10 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 		     sel->type == PIPE_SHADER_GEOMETRY))
 			shader->key.as_ngg = 1;
 
-		if (sel->tokens || sel->nir)
-			ir_binary = si_get_ir_binary(sel, shader->key.as_ngg);
+		if (sel->tokens || sel->nir) {
+			ir_binary = si_get_ir_binary(sel, shader->key.as_ngg,
+						     shader->key.as_es);
+		}
 
 		/* Try to load the shader from the shader cache. */
 		mtx_lock(&sscreen->shader_cache_mutex);