PM4 packet streaming support (v2)

index d6571e77b74d..1710b2cdf4a9 100644
@@ -274,6 +274,12 @@ int main(int argc, char **argv)
 		} else if (!strcmp(argv[i], "--waves") || !strcmp(argv[i], "-wa")) {
 			if (!asic)
 				asic = get_asic();
+			if (i + 1 < argc) {
+				if (argv[i+1][0] != '-') {
+					strcpy(asic->options.ring_name, argv[i+1]);
+					++i;
+				}
+			}
 			umr_print_waves(asic);
 		} else if (!strcmp(argv[i], "--scan") || !strcmp(argv[i], "-s")) {
 			if (i + 1 < argc) {

@@ -573,9 +579,11 @@ int main(int argc, char **argv)
 	"\n\t\t'-O bits,follow,empty_log' to continually dump the trace log.)\n"
 "\n\t--top, -t\n\t\tSummarize GPU utilization.  Can select a SE block with --bank.  Can use"
 	"\n\t\toptions 'use_colour' to colourize output and 'use_pci' to improve efficiency.\n"
-"\n\t--waves, -wa\n\t\tPrint out information about any active CU waves.  Can use '-O bits'"
+"\n\t--waves, -wa <ring_name>\n\t\tPrint out information about any active CU waves.  Can use '-O bits'"
 	"\n\t\tto see decoding of various wave fields.  Can use the '-O halt_waves' option"
-	"\n\t\tto halt the SQ while reading registers.\n"
+	"\n\t\tto halt the SQ while reading registers.  An optional ring name can be specified"
+	"\n\t\twhich will then search a given ring for pointers to active shaders.  It will"
+	"\n\t\tdefault to the 'gfx' ring if nothing is specified.\n"
 "\n\t--vm-decode, -vm vmid@<address> <num_of_pages>"
 	"\n\t\tDecode page mappings at a specified address (in hex) from the VMID specified."
 	"\n\t\tThe VMID can be specified in hexadecimal (with leading '0x') or in decimal."

index 6965f7f31854..3847d69a3b8d 100644
@@ -41,10 +41,19 @@ void umr_print_waves(struct umr_asic *asic)
 	uint64_t pgm_addr;
 	struct umr_wave_data *wd, *owd;
 	int first = 1, col = 0;
+	struct umr_shaders_pgm *shader = NULL;
+	struct umr_pm4_stream *stream;
 
 	if (asic->options.halt_waves)
 		umr_sq_cmd_halt_waves(asic, UMR_SQ_CMD_HALT);
 
+	// scan a ring but don't trigger the halt/resume
+	// since it would have already been done
+	x = asic->options.halt_waves;
+	asic->options.halt_waves = 0;
+	stream = umr_pm4_decode_ring(asic, asic->options.ring_name[0] ? asic->options.ring_name : "gfx");
+	asic->options.halt_waves = x;
+
 	if (asic->family <= FAMILY_CIK)
 		shift = 3;  // on SI..CIK allocations were done in 8-dword blocks
 	else

@@ -220,8 +229,28 @@ void umr_print_waves(struct umr_asic *asic)
 				}
 			}
 
-			printf("\n\nPGM_MEM:\n");
-			pgm_addr = (((uint64_t)wd->ws.pc_hi << 32) | wd->ws.pc_lo) - (NUM_OPCODE_WORDS*4)/2;
+			printf("\n\nPGM_MEM:");
+			pgm_addr = (((uint64_t)wd->ws.pc_hi << 32) | wd->ws.pc_lo);
+			if (stream)
+				shader = umr_find_shader_in_stream(stream, wd->ws.hw_id.vm_id, pgm_addr);
+			if (shader) {
+				printf(" (found shader at: %s%u%s@0x%s%llx%s of %s%u%s bytes)\n",
+					BLUE, shader->vmid, RST,
+					YELLOW, (unsigned long long)shader->addr, RST,
+					BLUE, shader->size, RST);
+
+				// start decoding a bit before PC if possible
+				if (shader->addr + ((NUM_OPCODE_WORDS*4)/2) < pgm_addr)
+					pgm_addr -= (NUM_OPCODE_WORDS*4)/2;
+				else
+					pgm_addr = shader->addr;
+
+				free(shader);
+			} else {
+				pgm_addr -= (NUM_OPCODE_WORDS*4)/2;
+				printf("\n");
+			}
+
 			umr_vm_disasm(asic, wd->ws.hw_id.vm_id, pgm_addr, (((uint64_t)wd->ws.pc_hi << 32) | wd->ws.pc_lo), NUM_OPCODE_WORDS*4, NULL);
 
 			Hv("LDS_ALLOC", wd->ws.lds_alloc.value);

@@ -253,6 +282,9 @@ void umr_print_waves(struct umr_asic *asic)
 		wd = owd;
 	}
 
+	if (stream)
+		umr_free_pm4_stream(stream);
+
 	if (asic->options.halt_waves)
 		umr_sq_cmd_halt_waves(asic, UMR_SQ_CMD_RESUME);
 }

244 lines not on this page · raw message

index 4c9f95e1b87b..326c13af694f 100644
@@ -213,7 +213,8 @@ struct umr_options {
 	char
 		*scanblock,
 		dev_name[32],
-		hub_name[32];
+		hub_name[32],
+		ring_name[32];
 	struct {
 		int domain,
 		    bus,

@@ -621,7 +622,26 @@ int umr_grbm_select_index(struct umr_asic *asic, uint32_t se, uint32_t sh, uint3
 int umr_sq_cmd_halt_waves(struct umr_asic *asic, enum umr_sq_cmd_halt_resume mode);
 
 /* IB/ring decoding/dumping/etc */
+struct umr_pm4_stream {
+	uint32_t pkttype,				// packet type (0==simple write, 3 == packet)
+			 pkt0off,				// base address for PKT0 writes
+			 opcode,
+			 n_words,				// number of words ignoring header
+			 *words;				// words following header word
+
+	struct umr_pm4_stream *next,	// adjacent PM4 packet if any
+						  *ib;		// IB this packet might point to
+
+	struct umr_shaders_pgm *shader; // shader program if any
+};
+
 void *umr_read_ring_data(struct umr_asic *asic, char *ringname, uint32_t *ringsize);
+struct umr_pm4_stream *umr_pm4_decode_ring(struct umr_asic *asic, char *ringname);
+struct umr_pm4_stream *umr_pm4_decode_stream(struct umr_asic *asic, int vmid, uint32_t *stream, uint32_t nwords);
+void umr_free_pm4_stream(struct umr_pm4_stream *stream);
+struct umr_shaders_pgm *umr_find_shader_in_stream(struct umr_pm4_stream *stream, unsigned vmid, uint64_t addr);
+struct umr_shaders_pgm *umr_find_shader_in_ring(struct umr_asic *asic, char *ringname, unsigned vmid, uint64_t addr, int no_halt);
+
 void umr_print_decode(struct umr_asic *asic, struct umr_ring_decoder *decoder, uint32_t ib);
 void umr_dump_ib(struct umr_asic *asic, struct umr_ring_decoder *decoder);
 void umr_dump_shaders(struct umr_asic *asic, struct umr_ring_decoder *decoder, struct umr_wave_data *wd);