wu-arch/vortex
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

adding tracking for SFU stalls

Browse Source
This commit is contained in:
Blaise Tine
2023-12-28 12:12:11 -08:00
parent c7a81d1493
commit e217bc2c23
27 changed files with 1266 additions and 1166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

81
runtime/common/utils.cpp
View File

@@ -208,6 +208,8 @@ extern int vx_dump_perf(vx_device_h hdevice, FILE* stream) {
uint64_t scrb_fpu = 0;
uint64_t scrb_lsu = 0;
uint64_t scrb_sfu = 0;
uint64_t scrb_wctl = 0;
uint64_t scrb_csrs = 0;
uint64_t ifetches = 0;
uint64_t loads = 0;
uint64_t stores = 0;
@@ -268,44 +270,69 @@ extern int vx_dump_perf(vx_device_h hdevice, FILE* stream) {
// PERF: pipeline
// scheduler idles
{
uint64_t sched_idles_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCHED_ID);
int idles_percent_per_core = calcAvgPercent(sched_idles_per_core, cycles_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: scheduler idles=%ld (%d%%)\n", core_id, sched_idles_per_core, idles_percent_per_core);
uint64_t sched_idles_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCHED_ID);
if (num_cores > 1) {
int idles_percent_per_core = calcAvgPercent(sched_idles_per_core, cycles_per_core);
fprintf(stream, "PERF: core%d: scheduler idle=%ld (%d%%)\n", core_id, sched_idles_per_core, idles_percent_per_core);
}
sched_idles += sched_idles_per_core;
}
// scheduler stalls
{
uint64_t sched_stalls_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCHED_ST);
int stalls_percent_per_core = calcAvgPercent(sched_stalls_per_core, cycles_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: scheduler stalls=%ld (%d%%)\n", core_id, sched_stalls_per_core, stalls_percent_per_core);
uint64_t sched_stalls_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCHED_ST);
if (num_cores > 1) {
int stalls_percent_per_core = calcAvgPercent(sched_stalls_per_core, cycles_per_core);
fprintf(stream, "PERF: core%d: scheduler stalls=%ld (%d%%)\n", core_id, sched_stalls_per_core, stalls_percent_per_core);
}
sched_stalls += sched_stalls_per_core;
}
// ibuffer_stalls
{
uint64_t ibuffer_stalls_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_IBUF_ST);
int ibuffer_percent_per_core = calcAvgPercent(ibuffer_stalls_per_core, cycles_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: ibuffer stalls=%ld (%d%%)\n", core_id, ibuffer_stalls_per_core, ibuffer_percent_per_core);
uint64_t ibuffer_stalls_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_IBUF_ST);
if (num_cores > 1) {
int ibuffer_percent_per_core = calcAvgPercent(ibuffer_stalls_per_core, cycles_per_core);
fprintf(stream, "PERF: core%d: ibuffer stalls=%ld (%d%%)\n", core_id, ibuffer_stalls_per_core, ibuffer_percent_per_core);
}
ibuffer_stalls += ibuffer_stalls_per_core;
}
// scrb_stalls
// issue_stalls
{
uint64_t scrb_stalls_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_ST);
uint64_t scrb_alu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_ALU);
uint64_t scrb_fpu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_FPU);
uint64_t scrb_lsu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_LSU);
uint64_t scrb_sfu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_SFU);
uint64_t scrb_total = scrb_alu_per_core + scrb_fpu_per_core + scrb_lsu_per_core + scrb_sfu_per_core;
uint64_t scrb_sfu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_SFU);
scrb_alu += scrb_alu_per_core;
scrb_fpu += scrb_fpu_per_core;
scrb_lsu += scrb_lsu_per_core;
scrb_sfu += scrb_sfu_per_core;
if (num_cores > 1) fprintf(stream, "PERF: core%d: scoreboard stalls=%ld (alu=%d%%, fpu=%d%%, lsu=%d%%, sfu=%d%%)\n", core_id, scrb_stalls_per_core,
if (num_cores > 1) {
uint64_t scrb_total = scrb_alu_per_core + scrb_fpu_per_core + scrb_lsu_per_core + scrb_sfu_per_core;
fprintf(stream, "PERF: core%d: issue stalls=%ld (alu=%d%%, fpu=%d%%, lsu=%d%%, sfu=%d%%)\n", core_id, scrb_stalls_per_core,
calcAvgPercent(scrb_alu_per_core, scrb_total),
calcAvgPercent(scrb_fpu_per_core, scrb_total),
calcAvgPercent(scrb_lsu_per_core, scrb_total),
calcAvgPercent(scrb_sfu_per_core, scrb_total));
}
scrb_stalls += scrb_stalls_per_core;
}
// sfu_stalls
{
uint64_t scrb_sfu_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_SFU);
uint64_t scrb_wctl_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_WCTL);
uint64_t scrb_csrs_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_SCRB_CSRS);
if (num_cores > 1) {
uint64_t sfu_total = scrb_wctl_per_core + scrb_csrs_per_core + scrb_tex_per_core + scrb_raster_per_core + scrb_om_per_core;
fprintf(stream, "PERF: core%d: sfu stalls=%ld (scrs=%d%%, wctl=%d%%)\n"
, core_id
, scrb_sfu_per_core
, calcAvgPercent(scrb_csrs_per_core, sfu_total)
, calcAvgPercent(scrb_wctl_per_core, sfu_total)
);
}
scrb_wctl += scrb_wctl_per_core;
scrb_csrs += scrb_csrs_per_core;
}
// PERF: memory
// ifetches
{
@@ -313,9 +340,11 @@ extern int vx_dump_perf(vx_device_h hdevice, FILE* stream) {
if (num_cores > 1) fprintf(stream, "PERF: core%d: ifetches=%ld\n", core_id, ifetches_per_core);
ifetches += ifetches_per_core;
uint64_t ifetch_lat_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_IFETCH_LT);
int mem_avg_lat = caclAverage(ifetch_lat_per_core, ifetches_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: ifetch latency=%d cycles\n", core_id, mem_avg_lat);
uint64_t ifetch_lat_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_IFETCH_LT);
if (num_cores > 1) {
int mem_avg_lat = caclAverage(ifetch_lat_per_core, ifetches_per_core);
fprintf(stream, "PERF: core%d: ifetch latency=%d cycles\n", core_id, mem_avg_lat);
}
ifetch_lat += ifetch_lat_per_core;
}
// loads
@@ -324,9 +353,11 @@ extern int vx_dump_perf(vx_device_h hdevice, FILE* stream) {
if (num_cores > 1) fprintf(stream, "PERF: core%d: loads=%ld\n", core_id, loads_per_core);
loads += loads_per_core;
uint64_t load_lat_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_LOAD_LT);
int mem_avg_lat = caclAverage(load_lat_per_core, loads_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: load latency=%d cycles\n", core_id, mem_avg_lat);
uint64_t load_lat_per_core = get_csr_64(staging_buf.data(), VX_CSR_MPM_LOAD_LT);
if (num_cores > 1) {
int mem_avg_lat = caclAverage(load_lat_per_core, loads_per_core);
fprintf(stream, "PERF: core%d: load latency=%d cycles\n", core_id, mem_avg_lat);
}
load_lat += load_lat_per_core;
}
// stores
@@ -428,14 +459,20 @@ extern int vx_dump_perf(vx_device_h hdevice, FILE* stream) {
int ifetch_avg_lat = (int)(double(ifetch_lat) / double(ifetches));
int load_avg_lat = (int)(double(load_lat) / double(loads));
uint64_t scrb_total = scrb_alu + scrb_fpu + scrb_lsu + scrb_sfu;
fprintf(stream, "PERF: scheduler idles=%ld (%d%%)\n", sched_idles, sched_idles_percent);
uint64_t sfu_total = scrb_wctl + scrb_csrs;
fprintf(stream, "PERF: scheduler idle=%ld (%d%%)\n", sched_idles, sched_idles_percent);
fprintf(stream, "PERF: scheduler stalls=%ld (%d%%)\n", sched_stalls, sched_stalls_percent);
fprintf(stream, "PERF: ibuffer stalls=%ld (%d%%)\n", ibuffer_stalls, ibuffer_percent);
fprintf(stream, "PERF: scoreboard stalls=%ld (alu=%d%%, fpu=%d%%, lsu=%d%%, sfu=%d%%)\n", scrb_stalls,
fprintf(stream, "PERF: issue stalls=%ld (alu=%d%%, fpu=%d%%, lsu=%d%%, sfu=%d%%)\n", scrb_stalls,
calcAvgPercent(scrb_alu, scrb_total),
calcAvgPercent(scrb_fpu, scrb_total),
calcAvgPercent(scrb_lsu, scrb_total),
calcAvgPercent(scrb_sfu, scrb_total));
calcAvgPercent(scrb_sfu, scrb_total));
fprintf(stream, "PERF: sfu stalls=%ld (scrs=%d%%, wctl=%d%%)\n"
, scrb_sfu
, calcAvgPercent(scrb_csrs, sfu_total)
, calcAvgPercent(scrb_wctl, sfu_total)
);
fprintf(stream, "PERF: ifetches=%ld\n", ifetches);
fprintf(stream, "PERF: loads=%ld\n", loads);
fprintf(stream, "PERF: stores=%ld\n", stores);