Attachment 121137 Details for Bug 93292 – Additional debugging for the r600/sb shader compiler

[patch] Additional debugging for the r600/sb shader compiler

mesa-r600-ext-dbg.diff (text/plain), 8.17 KB, created by Jan-Marek Glogowski on 2016-01-19 17:37:41 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Jan-Marek Glogowski

Created: 2016-01-19 17:37:41 UTC

Size: 8.17 KB

patch

obsolete

>diff --git a/src/gallium/drivers/r600/sb/sb_bc.h b/src/gallium/drivers/r600/sb/sb_bc.h
>index 6961600..c94aa36 100644
>--- a/src/gallium/drivers/r600/sb/sb_bc.h
>+++ b/src/gallium/drivers/r600/sb/sb_bc.h
>@@ -606,6 +606,31 @@ struct shader_stats {
> 	void dump_diff(shader_stats &s);
> };
> 
>+/* Use static sb_context::dskip_* member variables to skip shaders.
>+ * Default mode is skipping shaders (use shaders from the default
>+ * backend. Aternative mode is dry-running shaders.
>+ * range: [dskip_start, dskip_end], e.g. [5,6] means shaders 5 & 6
>+ * dskip_mode is a bitfield controlled by the SB_SKIP_* macros.
>+ */
>+
>+#define SB_SKIP_ENABLED      (1 << 0) // enable shader selection
>+#define SB_SKIP_IN_RANGE     (1 << 1) // shader id inside range else outside range
>+#define SB_SKIP_DRY_RUN      (1 << 2) // run in skip or dry-run mode
>+#define SB_SKIP_DUMP_IR      (1 << 3) // dump IR code
>+#define SB_SKIP_DUMP_STAT    (1 << 4) // dump stats
>+#define SB_SKIP_DUMP_BC      (1 << 5) // dump bytecode (disasm)
>+
>+#define SB_SKIP_DO(mode) \
>+	((sb_context::dskip_mode & (SB_SKIP_ENABLED | SB_SKIP_##mode)) \
>+		== (SB_SKIP_ENABLED | SB_SKIP_##mode))
>+#define SB_SKIP_DO_RANGE(id,mode) \
>+	(SB_SKIP_DO(mode) && sb_context::is_in_range(id))
>+
>+#define SB_DUMP_STAT(id,code) do { if (sb_context::dump_stat || \
>+	(SB_SKIP_DO_RANGE(id,DUMP_STAT))) { code } } while (0)
>+#define SB_DUMP_PASS(id,code) do { if (sb_context::dump_pass || \
>+	(SB_SKIP_DO_RANGE(id,DUMP_IR))) { code } } while (0)
>+
> class sb_context {
> 
> public:
>@@ -648,6 +673,12 @@ public:
> 
> 	int init(r600_isa *isa, sb_hw_chip chip, sb_hw_class cclass);
> 
>+	static bool is_in_range(unsigned shader_id)
>+	{
>+		return (dskip_start <= shader_id && shader_id <= dskip_end) ==
>+				((dskip_mode & SB_SKIP_IN_RANGE) > 0);
>+	}
>+
> 	bool is_r600() {return hw_class == HW_CLASS_R600;}
> 	bool is_r700() {return hw_class == HW_CLASS_R700;}
> 	bool is_evergreen() {return hw_class == HW_CLASS_EVERGREEN;}
>@@ -721,9 +752,6 @@ public:
> 
> };
> 
>-#define SB_DUMP_STAT(a) do { if (sb_context::dump_stat) { a } } while (0)
>-#define SB_DUMP_PASS(a) do { if (sb_context::dump_pass) { a } } while (0)
>-
> class bc_decoder {
> 
> 	sb_context &ctx;
>diff --git a/src/gallium/drivers/r600/sb/sb_core.cpp b/src/gallium/drivers/r600/sb/sb_core.cpp
>index afea818..8745d8f 100644
>--- a/src/gallium/drivers/r600/sb/sb_core.cpp
>+++ b/src/gallium/drivers/r600/sb/sb_core.cpp
>@@ -101,11 +101,10 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 	}
> 
> 	int64_t time_start = 0;
>-	if (sb_context::dump_stat) {
>+	SB_DUMP_STAT(shader_id,
> 		time_start = os_time_get_nano();
>-	}
>-
>-	SB_DUMP_STAT( sblog << "\nsb: shader " << shader_id << "\n"; );
>+		sblog << "\nsb: shader " << shader_id << "\n";
>+	);
> 
> 	bc_parser parser(*ctx, bc, pshader);
> 
>@@ -116,7 +115,7 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 
> 	shader *sh = parser.get_shader();
> 
>-	if (dump_bytecode) {
>+	if (dump_bytecode || SB_SKIP_DO_RANGE(shader_id, DUMP_BC)) {
> 		bc_dump(*sh, bc->bytecode, bc->ndw).run();
> 	}
> 
>@@ -127,25 +126,15 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 
> 	if (sh->target != TARGET_FETCH) {
> 		sh->src_stats.ndw = bc->ndw;
>-		sh->collect_stats(false);
>+		SB_DUMP_STAT(shader_id, sh->collect_stats(false););
> 	}
> 
>-	/* skip some shaders (use shaders from default backend)
>-	 * dskip_start - range start, dskip_end - range_end,
>-	 * e.g. start = 5, end = 6 means shaders 5 & 6
>-	 *
>-	 * dskip_mode == 0 - disabled,
>-	 * dskip_mode == 1 - don't process the shaders from the [start;end] range
>-	 * dskip_mode == 2 - process only the shaders from the range
>-	 */
>-	if (sb_context::dskip_mode) {
>-		if ((sb_context::dskip_start <= shader_id &&
>-				shader_id <= sb_context::dskip_end) ==
>-						(sb_context::dskip_mode == 1)) {
>+	if (sb_context::dskip_mode & SB_SKIP_ENABLED) {
>+		if (!(sb_context::dskip_mode & SB_SKIP_DRY_RUN) &&
>+			sb_context::is_in_range(shader_id)) {
> 			sblog << "sb: skipped shader " << shader_id << " : " << "["
> 					<< sb_context::dskip_start << "; "
>-					<< sb_context::dskip_end << "] mode "
>-					<< sb_context::dskip_mode << "\n";
>+					<< sb_context::dskip_end << "]\n";
> 			return 0;
> 		}
> 	}
>@@ -155,7 +144,9 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 		return r;
> 	}
> 
>-	SB_DUMP_PASS( sblog << "\n\n###### after parse\n"; sh->dump_ir(); );
>+	SB_DUMP_PASS(shader_id, sblog << "\n\n###### after parse\n"; sh->dump_ir(); );
>+
>+	unsigned passno = 0;
> 
> #define SB_RUN_PASS(n, dump) \
> 	do { \
>@@ -169,8 +160,8 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 			return 0; \
> 		} \
> 		if (dump) { \
>-			SB_DUMP_PASS( sblog << "\n\n###### after " << #n << "\n"; \
>-				sh->dump_ir();); \
>+			SB_DUMP_PASS(shader_id, sblog << "\n\n###### after " << #n << "\n"; \
>+				sh->dump_ir();) ; \
> 		} \
> 		assert(!r); \
> 	} while (0)
>@@ -249,12 +240,15 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 
> 	bytecode &nbc = builder.get_bytecode();
> 
>-	if (dump_bytecode) {
>+	if (dump_bytecode || SB_SKIP_DO_RANGE(shader_id, DUMP_BC)) {
> 		bc_dump(*sh, &nbc).run();
> 	}
> 
>-	if (!sb_context::dry_run) {
>-
>+	if (sb_context::dry_run) {
>+		SB_DUMP_STAT(shader_id, sblog << "sb: dry run: optimized bytecode is not used\n"; );
>+	} else if (SB_SKIP_DO_RANGE(shader_id, DRY_RUN)) {
>+		sblog << "sb: dry run shader " << shader_id << ": optimized bytecode is not used\n";
>+	} else {
> 		free(bc->bytecode);
> 		bc->ndw = nbc.ndw();
> 		bc->bytecode = (uint32_t*) malloc(bc->ndw << 2);
>@@ -262,11 +256,9 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 
> 		bc->ngpr = sh->ngpr;
> 		bc->nstack = sh->nstack;
>-	} else {
>-		SB_DUMP_STAT( sblog << "sb: dry run: optimized bytecode is not used\n"; );
> 	}
> 
>-	if (sb_context::dump_stat) {
>+	SB_DUMP_STAT(shader_id,
> 		int64_t t = os_time_get_nano() - time_start;
> 
> 		sblog << "sb: processing shader " << shader_id << " done ( "
>@@ -281,7 +273,7 @@ int r600_sb_bytecode_process(struct r600_context *rctx,
> 		sh->opt_stats.dump();
> 		sblog << "diff: ";
> 		sh->src_stats.dump_diff(sh->opt_stats);
>-	}
>+	);
> 
> 	delete sh;
> 	return 0;
>diff --git a/src/gallium/drivers/r600/sb/sb_gcm.cpp b/src/gallium/drivers/r600/sb/sb_gcm.cpp
>index 236b2ea..3eba7a5 100644
>--- a/src/gallium/drivers/r600/sb/sb_gcm.cpp
>+++ b/src/gallium/drivers/r600/sb/sb_gcm.cpp
>@@ -43,7 +43,8 @@ namespace r600_sb {
> 
> int gcm::run() {
> 
>-	GCM_DUMP( sblog << "==== GCM ==== \n"; sh.dump_ir(); );
>+	GCM_DUMP( sblog << "==== GCM ==== \n";
>+			if (sb_context::dump_pass) sh.dump_ir(); );
> 
> 	collect_instructions(sh.root, true);
> 
>@@ -83,7 +84,7 @@ int gcm::run() {
> 
> 	assert(pending.empty());
> 
>-	GCM_DUMP( sh.dump_ir(); );
>+	GCM_DUMP( if (sb_context::dump_pass) sh.dump_ir(); );
> 
> 	GCM_DUMP( sblog << "\n\n ############## gcm late\n\n"; );
> 
>diff --git a/src/gallium/drivers/r600/sb/sb_peephole.cpp b/src/gallium/drivers/r600/sb/sb_peephole.cpp
>index d4b9755..8038697 100644
>--- a/src/gallium/drivers/r600/sb/sb_peephole.cpp
>+++ b/src/gallium/drivers/r600/sb/sb_peephole.cpp
>@@ -76,9 +76,12 @@ void peephole::optimize_cc_op(alu_node* a) {
> 
> 	if (aflags & (AF_PRED | AF_SET | AF_KILL)) {
> 		optimize_cc_op2(a);
>-	} else if (aflags & AF_CMOV) {
>+	}
>+#if 0
>+        else if (aflags & AF_CMOV) {
> 		optimize_CNDcc_op(a);
> 	}
>+#endif
> }
> 
> void peephole::convert_float_setcc(alu_node *f2i, alu_node *s) {
>diff --git a/src/gallium/drivers/r600/sb/sb_shader.cpp b/src/gallium/drivers/r600/sb/sb_shader.cpp
>index 8c7b39b..79e423b 100644
>--- a/src/gallium/drivers/r600/sb/sb_shader.cpp
>+++ b/src/gallium/drivers/r600/sb/sb_shader.cpp
>@@ -358,8 +358,7 @@ shader::~shader() {
> }
> 
> void shader::dump_ir() {
>-	if (ctx.dump_pass)
>-		dump(*this).run();
>+	dump(*this).run();
> }
> 
> value* shader::get_value_version(value* v, unsigned ver) {
>@@ -394,8 +393,8 @@ void shader::add_gpr_array(unsigned gpr_start, unsigned gpr_count,
> 		if (comp_mask & 1) {
> 			gpr_array *a = new gpr_array(
> 					sel_chan(gpr_start, chan), gpr_count);
>-
>-			SB_DUMP_PASS( sblog << "add_gpr_array: @" << a->base_gpr
>+			// FIXME: how to handle shader_id correctly?
>+			SB_DUMP_PASS(0, sblog << "add_gpr_array: @" << a->base_gpr
> 			         << " [" << a->array_size << "]\n";
> 			);
> 
>@@ -491,9 +490,6 @@ alu_node* shader::clone(alu_node* n) {
> }
> 
> void shader::collect_stats(bool opt) {
>-	if (!sb_context::dump_stat)
>-		return;
>-
> 	shader_stats &s = opt ? opt_stats : src_stats;
> 
> 	s.shaders = 1;

Actions: View | Splinter Review

Attachments on bug 93292: 120410 | 121137 | 121138 | 121139

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.