Benchmark SIMD sincos32_simd (4 angles per call)
| 52 | |
| 53 | // Benchmark SIMD sincos32_simd (4 angles per call) |
| 54 | __attribute__((noinline)) |
| 55 | void benchmark_sincos32_simd(int calls) { |
| 56 | // Pre-generate test angle sets (16 different sets) |
| 57 | FL_ALIGNAS(16) u32 angle_sets[16][4]; |
| 58 | for (int set = 0; set < 16; set++) { |
| 59 | angle_sets[set][0] = set * 1048576; |
| 60 | angle_sets[set][1] = set * 1048576 + 4194304; |
| 61 | angle_sets[set][2] = set * 1048576 + 8388608; |
| 62 | angle_sets[set][3] = set * 1048576 + 12582912; |
| 63 | } |
| 64 | |
| 65 | i32 local_accumulator = 0; |
| 66 | |
| 67 | for (int i = 0; i < calls; i++) { |
| 68 | int set_idx = (i + g_angle_offset) & 15; |
| 69 | |
| 70 | asm volatile("" : : : "memory"); |
| 71 | simd::simd_u32x4 angles = simd::load_u32_4(angle_sets[set_idx]); |
| 72 | |
| 73 | SinCos32_simd result = sincos32_simd(angles); |
| 74 | |
| 75 | FL_ALIGNAS(16) u32 sin_vals[4]; |
| 76 | FL_ALIGNAS(16) u32 cos_vals[4]; |
| 77 | simd::store_u32_4(sin_vals, result.sin_vals); |
| 78 | simd::store_u32_4(cos_vals, result.cos_vals); |
| 79 | |
| 80 | local_accumulator ^= static_cast<i32>(sin_vals[0]); |
| 81 | local_accumulator ^= static_cast<i32>(cos_vals[0]); |
| 82 | |
| 83 | asm volatile("" : "+r"(local_accumulator) : : "memory"); |
| 84 | } |
| 85 | |
| 86 | g_accumulator = local_accumulator; |
| 87 | } |
| 88 | |
| 89 | int main(int argc, char *argv[]) { |
| 90 | bool json_output = (argc > 1 && fl::strcmp(argv[1], "baseline") == 0); |