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

CSRs I/O refactoring

Browse Source
This commit is contained in:
Blaise Tine
2021-06-11 03:08:07 -07:00
parent a46d6cb606
commit 3cc1190cd7
33 changed files with 881 additions and 1385 deletions
Download Patch File Download Diff File Expand all files Collapse all files

115
driver/common/vx_utils.cpp
View File

@@ -76,21 +76,20 @@ extern int vx_upload_kernel_file(vx_device_h device, const char* filename) {
return err;
}
int vx_csr_get_l(vx_device_h device, int core_id, int addr, int addr_h, uint64_t* value) {
int ret = 0;
unsigned value_lo, value_hi;
ret |= vx_csr_get(device, core_id, addr, &value_lo);
ret |= vx_csr_get(device, core_id, addr_h, &value_hi);
*value = (uint64_t(value_hi) << 32) | value_lo;
return ret;
/*static uint32_t get_csr_32(const uint32_t* buffer, int addr) {
uint32_t value_lo = buffer[addr - CSR_MPM_BASE];
return value_lo;
}*/
static uint64_t get_csr_64(const uint32_t* buffer, int addr) {
uint32_t value_lo = buffer[addr - CSR_MPM_BASE];
uint32_t value_hi = buffer[addr - CSR_MPM_BASE + 32];
return (uint64_t(value_hi) << 32) | value_lo;
}
extern int vx_dump_perf(vx_device_h device, FILE* stream) {
int ret = 0;
unsigned num_cores;
vx_csr_get(device, 0, CSR_NC, &num_cores);
uint64_t instrs = 0;
uint64_t cycles = 0;
@@ -127,12 +126,23 @@ extern int vx_dump_perf(vx_device_h device, FILE* stream) {
uint64_t mem_stalls = 0;
uint64_t mem_lat = 0;
#endif
for (unsigned core_id = 0; core_id < num_cores; ++core_id) {
uint64_t instrs_per_core, cycles_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_INSTRET, CSR_INSTRET_H, &instrs_per_core);
ret |= vx_csr_get_l(device, core_id, CSR_CYCLE, CSR_CYCLE_H, &cycles_per_core);
unsigned num_cores;
ret = vx_dev_caps(device, VX_CAPS_MAX_CORES, &num_cores);
if (ret)
return ret;
vx_buffer_h staging_buf;
ret |= vx_alloc_shared_mem(device, 64 * sizeof(uint32_t), &staging_buf);
if (ret)
return ret;
auto staging_ptr = (uint32_t*)vx_host_ptr(staging_buf);
for (unsigned core_id = 0; core_id < num_cores; ++core_id) {
ret |= vx_copy_from_dev(staging_buf, IO_ADDR_CSR + 64 * sizeof(uint32_t) * core_id, 64 * sizeof(uint32_t), 0);
uint64_t instrs_per_core = get_csr_64(staging_ptr, CSR_MINSTRET);
uint64_t cycles_per_core = get_csr_64(staging_ptr, CSR_MCYCLE);
float IPC = (float)(double(instrs_per_core) / double(cycles_per_core));
if (num_cores > 1) fprintf(stream, "PERF: core%d: instrs=%ld, cycles=%ld, IPC=%f\n", core_id, instrs_per_core, cycles_per_core, IPC);
instrs += instrs_per_core;
@@ -141,133 +151,110 @@ extern int vx_dump_perf(vx_device_h device, FILE* stream) {
#ifdef PERF_ENABLE
// PERF: pipeline
// ibuffer_stall
uint64_t ibuffer_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_IBUF_ST, CSR_MPM_IBUF_ST_H, &ibuffer_stalls_per_core);
uint64_t ibuffer_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_IBUF_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: ibuffer stalls=%ld\n", core_id, ibuffer_stalls_per_core);
ibuffer_stalls += ibuffer_stalls_per_core;
// scoreboard_stall
uint64_t scoreboard_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_SCRB_ST, CSR_MPM_SCRB_ST_H, &scoreboard_stalls_per_core);
uint64_t scoreboard_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_SCRB_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: scoreboard stalls=%ld\n", core_id, scoreboard_stalls_per_core);
scoreboard_stalls += scoreboard_stalls_per_core;
// alu_stall
uint64_t alu_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_ALU_ST, CSR_MPM_ALU_ST_H, &alu_stalls_per_core);
uint64_t alu_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_ALU_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: alu unit stalls=%ld\n", core_id, alu_stalls_per_core);
alu_stalls += alu_stalls_per_core;
// lsu_stall
uint64_t lsu_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_LSU_ST, CSR_MPM_LSU_ST_H, &lsu_stalls_per_core);
uint64_t lsu_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_LSU_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: lsu unit stalls=%ld\n", core_id, lsu_stalls_per_core);
lsu_stalls += lsu_stalls_per_core;
// csr_stall
uint64_t csr_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_CSR_ST, CSR_MPM_CSR_ST_H, &csr_stalls_per_core);
uint64_t csr_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_CSR_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: csr unit stalls=%ld\n", core_id, csr_stalls_per_core);
csr_stalls += csr_stalls_per_core;
// fpu_stall
uint64_t fpu_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_FPU_ST, CSR_MPM_FPU_ST_H, &fpu_stalls_per_core);
uint64_t fpu_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_FPU_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: fpu unit stalls=%ld\n", core_id, fpu_stalls_per_core);
fpu_stalls += fpu_stalls_per_core;
// gpu_stall
uint64_t gpu_stalls_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_GPU_ST, CSR_MPM_GPU_ST_H, &gpu_stalls_per_core);
uint64_t gpu_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_GPU_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: gpu unit stalls=%ld\n", core_id, gpu_stalls_per_core);
gpu_stalls += gpu_stalls_per_core;
// PERF: Icache
// total reads
uint64_t icache_reads_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_ICACHE_READS, CSR_MPM_ICACHE_READS_H, &icache_reads_per_core);
uint64_t icache_reads_per_core = get_csr_64(staging_ptr, CSR_MPM_ICACHE_READS);
if (num_cores > 1) fprintf(stream, "PERF: core%d: icache reads=%ld\n", core_id, icache_reads_per_core);
icache_reads += icache_reads_per_core;
// read misses
uint64_t icache_miss_r_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_ICACHE_MISS_R, CSR_MPM_ICACHE_MISS_R_H, &icache_miss_r_per_core);
uint64_t icache_miss_r_per_core = get_csr_64(staging_ptr, CSR_MPM_ICACHE_MISS_R);
int icache_read_hit_ratio = (int)((1.0 - (double(icache_miss_r_per_core) / double(icache_reads_per_core))) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: icache read misses=%ld (hit ratio=%d%%)\n", core_id, icache_miss_r_per_core, icache_read_hit_ratio);
icache_read_misses += icache_miss_r_per_core;
// pipeline stalls
uint64_t icache_pipe_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_ICACHE_PIPE_ST, CSR_MPM_ICACHE_PIPE_ST_H, &icache_pipe_st_per_core);
uint64_t icache_pipe_st_per_core = get_csr_64(staging_ptr, CSR_MPM_ICACHE_PIPE_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: icache pipeline stalls=%ld\n", core_id, icache_pipe_st_per_core);
icache_pipe_stalls += icache_pipe_st_per_core;
// response stalls
uint64_t icache_crsp_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_ICACHE_CRSP_ST, CSR_MPM_ICACHE_CRSP_ST_H, &icache_crsp_st_per_core);
uint64_t icache_crsp_st_per_core = get_csr_64(staging_ptr, CSR_MPM_ICACHE_CRSP_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: icache reponse stalls=%ld\n", core_id, icache_crsp_st_per_core);
icache_rsp_stalls += icache_crsp_st_per_core;
// PERF: Dcache
// total reads
uint64_t dcache_reads_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_READS, CSR_MPM_DCACHE_READS_H, &dcache_reads_per_core);
uint64_t dcache_reads_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_READS);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache reads=%ld\n", core_id, dcache_reads_per_core);
dcache_reads += dcache_reads_per_core;
// total write
uint64_t dcache_writes_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_WRITES, CSR_MPM_DCACHE_WRITES_H, &dcache_writes_per_core);
uint64_t dcache_writes_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_WRITES);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache writes=%ld\n", core_id, dcache_writes_per_core);
dcache_writes += dcache_writes_per_core;
// read misses
uint64_t dcache_miss_r_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_MISS_R, CSR_MPM_DCACHE_MISS_R_H, &dcache_miss_r_per_core);
uint64_t dcache_miss_r_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_MISS_R);
int dcache_read_hit_ratio = (int)((1.0 - (double(dcache_miss_r_per_core) / double(dcache_reads_per_core))) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache read misses=%ld (hit ratio=%d%%)\n", core_id, dcache_miss_r_per_core, dcache_read_hit_ratio);
dcache_read_misses += dcache_miss_r_per_core;
// read misses
uint64_t dcache_miss_w_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_MISS_W, CSR_MPM_DCACHE_MISS_W_H, &dcache_miss_w_per_core);
uint64_t dcache_miss_w_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_MISS_W);
int dcache_write_hit_ratio = (int)((1.0 - (double(dcache_miss_w_per_core) / double(dcache_writes_per_core))) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache write misses=%ld (hit ratio=%d%%)\n", core_id, dcache_miss_w_per_core, dcache_write_hit_ratio);
dcache_write_misses += dcache_miss_w_per_core;
// bank_stalls
uint64_t dcache_bank_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_BANK_ST, CSR_MPM_DCACHE_BANK_ST_H, &dcache_bank_st_per_core);
uint64_t dcache_bank_st_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_BANK_ST);
int dcache_bank_utilization = (int)((double(dcache_reads_per_core + dcache_writes_per_core) / double(dcache_reads_per_core + dcache_writes_per_core + dcache_bank_st_per_core)) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache bank stalls=%ld (utilization=%d%%)\n", core_id, dcache_bank_st_per_core, dcache_bank_utilization);
dcache_bank_stalls += dcache_bank_st_per_core;
// mshr_stalls
uint64_t dcache_mshr_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_MSHR_ST, CSR_MPM_DCACHE_MSHR_ST_H, &dcache_mshr_st_per_core);
uint64_t dcache_mshr_st_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_MSHR_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache mshr stalls=%ld\n", core_id, dcache_mshr_st_per_core);
dcache_mshr_stalls += dcache_mshr_st_per_core;
// pipeline stalls
uint64_t dcache_pipe_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_PIPE_ST, CSR_MPM_DCACHE_PIPE_ST_H, &dcache_pipe_st_per_core);
uint64_t dcache_pipe_st_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_PIPE_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache pipeline stalls=%ld\n", core_id, dcache_pipe_st_per_core);
dcache_pipe_stalls += dcache_pipe_st_per_core;
// response stalls
uint64_t dcache_crsp_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_DCACHE_CRSP_ST, CSR_MPM_DCACHE_CRSP_ST_H, &dcache_crsp_st_per_core);
uint64_t dcache_crsp_st_per_core = get_csr_64(staging_ptr, CSR_MPM_DCACHE_CRSP_ST);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dcache reponse stalls=%ld\n", core_id, dcache_crsp_st_per_core);
dcache_rsp_stalls += dcache_crsp_st_per_core;
// PERF: SMEM
// total reads
uint64_t smem_reads_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_SMEM_READS, CSR_MPM_SMEM_READS_H, &smem_reads_per_core);
uint64_t smem_reads_per_core = get_csr_64(staging_ptr, CSR_MPM_SMEM_READS);
if (num_cores > 1) fprintf(stream, "PERF: core%d: smem reads=%ld\n", core_id, smem_reads_per_core);
smem_reads += smem_reads_per_core;
// total write
uint64_t smem_writes_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_SMEM_WRITES, CSR_MPM_SMEM_WRITES_H, &smem_writes_per_core);
uint64_t smem_writes_per_core = get_csr_64(staging_ptr, CSR_MPM_SMEM_WRITES);
if (num_cores > 1) fprintf(stream, "PERF: core%d: smem writes=%ld\n", core_id, smem_writes_per_core);
smem_writes += smem_writes_per_core;
// bank_stalls
uint64_t smem_bank_st_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_SMEM_BANK_ST, CSR_MPM_SMEM_BANK_ST_H, &smem_bank_st_per_core);
uint64_t smem_bank_st_per_core = get_csr_64(staging_ptr, CSR_MPM_SMEM_BANK_ST);
int smem_bank_utilization = (int)((double(smem_reads_per_core + smem_writes_per_core) / double(smem_reads_per_core + smem_writes_per_core + smem_bank_st_per_core)) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: smem bank stalls=%ld (utilization=%d%%)\n", core_id, smem_bank_st_per_core, smem_bank_utilization);
smem_bank_stalls += smem_bank_st_per_core;
// PERF: memory
uint64_t mem_reads_per_core, mem_writes_per_core, mem_stalls_per_core, mem_lat_per_core;
ret |= vx_csr_get_l(device, core_id, CSR_MPM_MEM_READS, CSR_MPM_MEM_READS_H, &mem_reads_per_core);
ret |= vx_csr_get_l(device, core_id, CSR_MPM_MEM_WRITES, CSR_MPM_MEM_WRITES_H, &mem_writes_per_core);
ret |= vx_csr_get_l(device, core_id, CSR_MPM_MEM_ST, CSR_MPM_MEM_ST_H, &mem_stalls_per_core);
ret |= vx_csr_get_l(device, core_id, CSR_MPM_MEM_LAT, CSR_MPM_MEM_LAT_H, &mem_lat_per_core);
uint64_t mem_reads_per_core = get_csr_64(staging_ptr, CSR_MPM_MEM_READS);
uint64_t mem_writes_per_core = get_csr_64(staging_ptr, CSR_MPM_MEM_WRITES);
uint64_t mem_stalls_per_core = get_csr_64(staging_ptr, CSR_MPM_MEM_ST);
uint64_t mem_lat_per_core = get_csr_64(staging_ptr, CSR_MPM_MEM_LAT);
int mem_utilization = (int)((double(mem_reads_per_core + mem_writes_per_core) / double(mem_reads_per_core + mem_writes_per_core + mem_stalls_per_core)) * 100);
int mem_avg_lat = (int)(double(mem_lat_per_core) / double(mem_reads_per_core));
if (num_cores > 1) fprintf(stream, "PERF: core%d: memory requests=%ld (reads=%ld, writes=%ld)\n", core_id, (mem_reads_per_core + mem_writes_per_core), mem_reads_per_core, mem_writes_per_core);