radeonsi: move si_get_wave_info() to AMD common code

This will allow us to use it from radv.

Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Marek Olšák <[email protected]>
Reviewed-by: Bas Nieuwenhuizen <[email protected]>

2 files changed, 94 insertions, 0 deletions

diff --git a/src/amd/common/ac_debug.c b/src/amd/common/ac_debug.c
index 54685356f1d..26d1f7dcedd 100644
--- a/src/amd/common/ac_debug.c
+++ b/src/amd/common/ac_debug.c
@@ -43,6 +43,8 @@
 #include "util/u_memory.h"
 #include "util/u_string.h"
 
+#include <assert.h>
+
 /* Parsed IBs are difficult to read without colors. Use "less -R file" to
  * read them, or use "aha -b -f file" to convert them to html.
  */
@@ -721,3 +723,77 @@ bool ac_vm_fault_occured(enum chip_class chip_class,
 
 	return fault;
 }
+
+static int compare_wave(const void *p1, const void *p2)
+{
+	struct ac_wave_info *w1 = (struct ac_wave_info *)p1;
+	struct ac_wave_info *w2 = (struct ac_wave_info *)p2;
+
+	/* Sort waves according to PC and then SE, SH, CU, etc. */
+	if (w1->pc < w2->pc)
+		return -1;
+	if (w1->pc > w2->pc)
+		return 1;
+	if (w1->se < w2->se)
+		return -1;
+	if (w1->se > w2->se)
+		return 1;
+	if (w1->sh < w2->sh)
+		return -1;
+	if (w1->sh > w2->sh)
+		return 1;
+	if (w1->cu < w2->cu)
+		return -1;
+	if (w1->cu > w2->cu)
+		return 1;
+	if (w1->simd < w2->simd)
+		return -1;
+	if (w1->simd > w2->simd)
+		return 1;
+	if (w1->wave < w2->wave)
+		return -1;
+	if (w1->wave > w2->wave)
+		return 1;
+
+	return 0;
+}
+
+/* Return wave information. "waves" should be a large enough array. */
+unsigned ac_get_wave_info(struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP])
+{
+	char line[2000];
+	unsigned num_waves = 0;
+
+	FILE *p = popen("umr -wa", "r");
+	if (!p)
+		return 0;
+
+	if (!fgets(line, sizeof(line), p) ||
+	    strncmp(line, "SE", 2) != 0) {
+		pclose(p);
+		return 0;
+	}
+
+	while (fgets(line, sizeof(line), p)) {
+		struct ac_wave_info *w;
+		uint32_t pc_hi, pc_lo, exec_hi, exec_lo;
+
+		assert(num_waves < AC_MAX_WAVES_PER_CHIP);
+		w = &waves[num_waves];
+
+		if (sscanf(line, "%u %u %u %u %u %x %x %x %x %x %x %x",
+			   &w->se, &w->sh, &w->cu, &w->simd, &w->wave,
+			   &w->status, &pc_hi, &pc_lo, &w->inst_dw0,
+			   &w->inst_dw1, &exec_hi, &exec_lo) == 12) {
+			w->pc = ((uint64_t)pc_hi << 32) | pc_lo;
+			w->exec = ((uint64_t)exec_hi << 32) | exec_lo;
+			w->matched = false;
+			num_waves++;
+		}
+	}
+
+	qsort(waves, num_waves, sizeof(struct ac_wave_info), compare_wave);
+
+	pclose(p);
+	return num_waves;
+}
diff --git a/src/amd/common/ac_debug.h b/src/amd/common/ac_debug.h
index 35e950014b0..df207526531 100644
--- a/src/amd/common/ac_debug.h
+++ b/src/amd/common/ac_debug.h
@@ -36,6 +36,22 @@
 #define AC_IS_TRACE_POINT(x)            (((x) & 0xcafe0000) == 0xcafe0000)
 #define AC_GET_TRACE_POINT_ID(x)        ((x) & 0xffff)
 
+#define AC_MAX_WAVES_PER_CHIP (64 * 40)
+
+struct ac_wave_info {
+	unsigned se; /* shader engine */
+	unsigned sh; /* shader array */
+	unsigned cu; /* compute unit */
+	unsigned simd;
+	unsigned wave;
+	uint32_t status;
+	uint64_t pc; /* program counter */
+	uint32_t inst_dw0;
+	uint32_t inst_dw1;
+	uint64_t exec;
+	bool matched; /* whether the wave is used by a currently-bound shader */
+};
+
 typedef void *(*ac_debug_addr_callback)(void *data, uint64_t addr);
 
 void ac_dump_reg(FILE *file, enum chip_class chip_class, unsigned offset,
@@ -50,4 +66,6 @@ void ac_parse_ib(FILE *f, uint32_t *ib, int num_dw, const int *trace_ids,
 bool ac_vm_fault_occured(enum chip_class chip_class,
 			 uint64_t *old_dmesg_timestamp, uint64_t *out_addr);
 
+unsigned ac_get_wave_info(struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP]);
+
 #endif