| 1 | |
| 2 | enum pcounters { |
| 3 | PCNT_ALL, |
| 4 | PCNT_GPU, |
| 5 | PCNT_SPU, |
| 6 | PCNT_BLIT, |
| 7 | PCNT_GTE, |
| 8 | PCNT_TEST, |
| 9 | PCNT_CNT |
| 10 | }; |
| 11 | |
| 12 | #ifdef PCNT |
| 13 | |
| 14 | #if defined(__ARM_ARCH_7A__) || defined(ARM1176) |
| 15 | #define PCNT_DIV 1000 |
| 16 | #else |
| 17 | #include <sys/time.h> |
| 18 | #define PCNT_DIV 1 |
| 19 | #endif |
| 20 | |
| 21 | static const char *pcnt_names[PCNT_CNT] = { "", "gpu", "spu", "blit", "gte", "test" }; |
| 22 | |
| 23 | #define PCNT_FRAMES 10 |
| 24 | |
| 25 | extern unsigned int pcounters[PCNT_CNT]; |
| 26 | extern unsigned int pcounter_starts[PCNT_CNT]; |
| 27 | |
| 28 | #define pcnt_start(id) \ |
| 29 | pcounter_starts[id] = pcnt_get() |
| 30 | |
| 31 | #define pcnt_end(id) \ |
| 32 | pcounters[id] += pcnt_get() - pcounter_starts[id] |
| 33 | |
| 34 | void pcnt_hook_plugins(void); |
| 35 | |
| 36 | static inline void pcnt_print(float fps) |
| 37 | { |
| 38 | static int print_counter; |
| 39 | unsigned int total, rem; |
| 40 | int i; |
| 41 | |
| 42 | for (i = 0; i < PCNT_CNT; i++) |
| 43 | pcounters[i] /= PCNT_DIV * PCNT_FRAMES; |
| 44 | |
| 45 | rem = total = pcounters[PCNT_ALL]; |
| 46 | for (i = 1; i < PCNT_CNT; i++) |
| 47 | rem -= pcounters[i]; |
| 48 | if (!total) |
| 49 | total++; |
| 50 | |
| 51 | if (--print_counter < 0) { |
| 52 | printf(" "); |
| 53 | for (i = 1; i < PCNT_CNT; i++) |
| 54 | printf("%5s ", pcnt_names[i]); |
| 55 | printf("%5s\n", "rem"); |
| 56 | print_counter = 30; |
| 57 | } |
| 58 | |
| 59 | printf("%4.1f ", fps); |
| 60 | #if 0 |
| 61 | static float pcounters_all[PCNT_CNT+1]; |
| 62 | static int pcounter_samples; |
| 63 | pcounter_samples++; |
| 64 | |
| 65 | for (i = 1; i < PCNT_CNT; i++) { |
| 66 | pcounters_all[i] += pcounters[i]; |
| 67 | printf("%5.0f ", pcounters_all[i] / pcounter_samples); |
| 68 | } |
| 69 | pcounters_all[i] += rem; |
| 70 | printf("%5.0f\n", pcounters_all[i] / pcounter_samples); |
| 71 | #else |
| 72 | for (i = 1; i < PCNT_CNT; i++) |
| 73 | printf("%5u ", pcounters[i]); |
| 74 | printf("%5u (", rem); |
| 75 | for (i = 1; i < PCNT_CNT; i++) |
| 76 | printf("%2u ", pcounters[i] * 100 / total); |
| 77 | printf("%2u) %u\n", rem * 100 / total, total); |
| 78 | #endif |
| 79 | memset(pcounters, 0, sizeof(pcounters)); |
| 80 | } |
| 81 | |
| 82 | static inline unsigned int pcnt_get(void) |
| 83 | { |
| 84 | unsigned int val; |
| 85 | #ifdef __ARM_ARCH_7A__ |
| 86 | __asm__ volatile("mrc p15, 0, %0, c9, c13, 0" |
| 87 | : "=r"(val)); |
| 88 | #elif defined(ARM1176) |
| 89 | __asm__ volatile("mrc p15, 0, %0, c15, c12, 1" |
| 90 | : "=r"(val)); |
| 91 | #else |
| 92 | // all slow on ARM :( |
| 93 | //struct timespec tv; |
| 94 | //clock_gettime(CLOCK_MONOTONIC_RAW, &tv); |
| 95 | //val = tv.tv_sec * 1000000000 + tv.tv_nsec; |
| 96 | struct timeval tv; |
| 97 | gettimeofday(&tv, NULL); |
| 98 | val = tv.tv_sec * 1000000 + tv.tv_usec; |
| 99 | #endif |
| 100 | return val; |
| 101 | } |
| 102 | |
| 103 | static inline void pcnt_init(void) |
| 104 | { |
| 105 | #ifdef __ARM_ARCH_7A__ |
| 106 | int v; |
| 107 | asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(v)); |
| 108 | v |= 5; // master enable, ccnt reset |
| 109 | v &= ~8; // ccnt divider 0 |
| 110 | asm volatile("mcr p15, 0, %0, c9, c12, 0" :: "r"(v)); |
| 111 | // enable cycle counter |
| 112 | asm volatile("mcr p15, 0, %0, c9, c12, 1" :: "r"(1<<31)); |
| 113 | #elif defined(ARM1176) |
| 114 | int v; |
| 115 | asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(v)); |
| 116 | v |= 5; // master enable, ccnt reset |
| 117 | v &= ~8; // ccnt divider 0 |
| 118 | asm volatile("mcr p15, 0, %0, c15, c12, 0" :: "r"(v)); |
| 119 | #endif |
| 120 | } |
| 121 | |
| 122 | void pcnt_gte_start(int op); |
| 123 | void pcnt_gte_end(int op); |
| 124 | |
| 125 | #else |
| 126 | |
| 127 | #define pcnt_start(id) |
| 128 | #define pcnt_end(id) |
| 129 | #define pcnt_hook_plugins() |
| 130 | #define pcnt_print(fps) |
| 131 | |
| 132 | #endif |