* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- * Authors:
- * Marek Olšák <maraeo@gmail.com>
*/
#include "ac_debug.h"
#include <memcheck.h>
#define VG(x) x
#else
-#define VG(x)
+#define VG(x) ((void)0)
#endif
#include <inttypes.h>
#include "sid.h"
-#include "gfx9d.h"
#include "sid_tables.h"
#include "util/u_math.h"
#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.
*/
print_value(file, value, bits);
}
-static const struct si_reg *find_register(const struct si_reg *table,
- unsigned table_size,
- unsigned offset)
+static const struct si_reg *find_register(enum chip_class chip_class, unsigned offset)
{
+ const struct si_reg *table;
+ unsigned table_size;
+
+ switch (chip_class) {
+ case GFX10_3:
+ case GFX10:
+ table = gfx10_reg_table;
+ table_size = ARRAY_SIZE(gfx10_reg_table);
+ break;
+ case GFX9:
+ table = gfx9_reg_table;
+ table_size = ARRAY_SIZE(gfx9_reg_table);
+ break;
+ case GFX8:
+ table = gfx8_reg_table;
+ table_size = ARRAY_SIZE(gfx8_reg_table);
+ break;
+ case GFX7:
+ table = gfx7_reg_table;
+ table_size = ARRAY_SIZE(gfx7_reg_table);
+ break;
+ case GFX6:
+ table = gfx6_reg_table;
+ table_size = ARRAY_SIZE(gfx6_reg_table);
+ break;
+ default:
+ return NULL;
+ }
+
for (unsigned i = 0; i < table_size; i++) {
const struct si_reg *reg = &table[i];
return NULL;
}
+const char *ac_get_register_name(enum chip_class chip_class, unsigned offset)
+{
+ const struct si_reg *reg = find_register(chip_class, offset);
+
+ return reg ? sid_strings + reg->name_offset : "(no name)";
+}
+
void ac_dump_reg(FILE *file, enum chip_class chip_class, unsigned offset,
uint32_t value, uint32_t field_mask)
{
- const struct si_reg *reg = NULL;
-
- if (chip_class >= GFX9)
- reg = find_register(gfx9d_reg_table, ARRAY_SIZE(gfx9d_reg_table), offset);
- if (!reg)
- reg = find_register(sid_reg_table, ARRAY_SIZE(sid_reg_table), offset);
+ const struct si_reg *reg = find_register(chip_class, offset);
if (reg) {
const char *reg_name = sid_strings + reg->name_offset;
if (op == PKT3_SET_CONTEXT_REG ||
op == PKT3_SET_CONFIG_REG ||
op == PKT3_SET_UCONFIG_REG ||
+ op == PKT3_SET_UCONFIG_REG_INDEX ||
op == PKT3_SET_SH_REG)
fprintf(f, COLOR_CYAN "%s%s" COLOR_CYAN ":\n",
name, predicate);
ac_parse_set_reg_packet(f, count, SI_CONFIG_REG_OFFSET, ib);
break;
case PKT3_SET_UCONFIG_REG:
+ case PKT3_SET_UCONFIG_REG_INDEX:
ac_parse_set_reg_packet(f, count, CIK_UCONFIG_REG_OFFSET, ib);
break;
case PKT3_SET_SH_REG:
ac_dump_reg(f, ib->chip_class, R_0301F8_CP_COHER_BASE, ac_ib_get(ib), ~0);
ac_dump_reg(f, ib->chip_class, R_0301E4_CP_COHER_BASE_HI, ac_ib_get(ib), ~0);
print_named_value(f, "POLL_INTERVAL", ac_ib_get(ib), 16);
+ if (ib->chip_class >= GFX10)
+ ac_dump_reg(f, ib->chip_class, R_586_GCR_CNTL, ac_ib_get(ib), ~0);
break;
case PKT3_SURFACE_SYNC:
- if (ib->chip_class >= CIK) {
+ if (ib->chip_class >= GFX7) {
ac_dump_reg(f, ib->chip_class, R_0301F0_CP_COHER_CNTL, ac_ib_get(ib), ~0);
ac_dump_reg(f, ib->chip_class, R_0301F4_CP_COHER_SIZE, ac_ib_get(ib), ~0);
ac_dump_reg(f, ib->chip_class, R_0301F8_CP_COHER_BASE, ac_ib_get(ib), ~0);
}
case PKT3_RELEASE_MEM: {
uint32_t event_dw = ac_ib_get(ib);
- ac_dump_reg(f, ib->chip_class, R_028A90_VGT_EVENT_INITIATOR, event_dw,
- S_028A90_EVENT_TYPE(~0));
- print_named_value(f, "EVENT_INDEX", (event_dw >> 8) & 0xf, 4);
- print_named_value(f, "TCL1_VOL_ACTION_ENA", (event_dw >> 12) & 0x1, 1);
- print_named_value(f, "TC_VOL_ACTION_ENA", (event_dw >> 13) & 0x1, 1);
- print_named_value(f, "TC_WB_ACTION_ENA", (event_dw >> 15) & 0x1, 1);
- print_named_value(f, "TCL1_ACTION_ENA", (event_dw >> 16) & 0x1, 1);
- print_named_value(f, "TC_ACTION_ENA", (event_dw >> 17) & 0x1, 1);
- print_named_value(f, "TC_NC_ACTION_ENA", (event_dw >> 19) & 0x1, 1);
- print_named_value(f, "TC_WC_ACTION_ENA", (event_dw >> 20) & 0x1, 1);
- print_named_value(f, "TC_MD_ACTION_ENA", (event_dw >> 21) & 0x1, 1);
+ if (ib->chip_class >= GFX10) {
+ ac_dump_reg(f, ib->chip_class, R_490_RELEASE_MEM_OP, event_dw, ~0u);
+ } else {
+ ac_dump_reg(f, ib->chip_class, R_028A90_VGT_EVENT_INITIATOR, event_dw,
+ S_028A90_EVENT_TYPE(~0));
+ print_named_value(f, "EVENT_INDEX", (event_dw >> 8) & 0xf, 4);
+ print_named_value(f, "TCL1_VOL_ACTION_ENA", (event_dw >> 12) & 0x1, 1);
+ print_named_value(f, "TC_VOL_ACTION_ENA", (event_dw >> 13) & 0x1, 1);
+ print_named_value(f, "TC_WB_ACTION_ENA", (event_dw >> 15) & 0x1, 1);
+ print_named_value(f, "TCL1_ACTION_ENA", (event_dw >> 16) & 0x1, 1);
+ print_named_value(f, "TC_ACTION_ENA", (event_dw >> 17) & 0x1, 1);
+ print_named_value(f, "TC_NC_ACTION_ENA", (event_dw >> 19) & 0x1, 1);
+ print_named_value(f, "TC_WC_ACTION_ENA", (event_dw >> 20) & 0x1, 1);
+ print_named_value(f, "TC_MD_ACTION_ENA", (event_dw >> 21) & 0x1, 1);
+ }
uint32_t sel_dw = ac_ib_get(ib);
print_named_value(f, "DST_SEL", (sel_dw >> 16) & 0x3, 2);
print_named_value(f, "INT_SEL", (sel_dw >> 24) & 0x7, 3);
uint32_t base_hi_dw = ac_ib_get(ib);
ac_dump_reg(f, ib->chip_class, R_3F1_IB_BASE_HI, base_hi_dw, ~0);
uint32_t control_dw = ac_ib_get(ib);
- ac_dump_reg(f, ib->chip_class, R_3F2_CONTROL, control_dw, ~0);
+ ac_dump_reg(f, ib->chip_class, R_3F2_IB_CONTROL, control_dw, ~0);
if (!ib->addr_callback)
break;
case PKT3_PFP_SYNC_ME:
break;
case PKT3_NOP:
- if (header == 0xffff1000) {
+ if (header == PKT3_NOP_PAD) {
count = -1; /* One dword NOP. */
} else if (count == 0 && ib->cur_dw < ib->num_dw &&
AC_IS_TRACE_POINT(ib->ib[ib->cur_dw])) {
if (indent)
print_spaces(f, indent);
- char *end = util_strchrnul(out, '\n');
+ char *end = strchrnul(out, '\n');
fwrite(out, end - out, 1, f);
fputc('\n', f); /* always end with a new line */
if (!*end)
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(enum chip_class chip_class,
+ struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP])
+{
+ char line[2000], cmd[128];
+ unsigned num_waves = 0;
+
+ sprintf(cmd, "umr -O halt_waves -wa %s", chip_class >= GFX10 ? "gfx_0.0.0" : "gfx");
+
+ FILE *p = popen(cmd, "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;
+}