File size: 6,013 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
#include <benchmark/benchmark.h>

#include <fp16.h>
#ifndef EMSCRIPTEN
	#include <fp16/psimd.h>
#endif

#include <vector>
#include <random>
#include <chrono>
#include <functional>
#include <algorithm>

#if defined(__ARM_NEON__) || defined(__aarch64__)
	#include <arm_neon.h>
#endif


static void fp16_alt_to_fp32_bits(benchmark::State& state) {
	const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
	auto rng = std::bind(std::uniform_int_distribution<uint16_t>(0, 0x7BFF), std::mt19937(seed));

	std::vector<uint16_t> fp16(state.range(0));
	std::vector<uint32_t> fp32(state.range(0));
	std::generate(fp16.begin(), fp16.end(),
		[&rng]{ return fp16_alt_from_fp32_value(rng()); });

	while (state.KeepRunning()) {
		uint16_t* input = fp16.data();
		benchmark::DoNotOptimize(input);

		uint32_t* output = fp32.data();
		const size_t n = state.range(0);
		for (size_t i = 0; i < n; i++) {
			output[i] = fp16_alt_to_fp32_bits(input[i]);
		}

		benchmark::DoNotOptimize(output);
	}
	state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
}
BENCHMARK(fp16_alt_to_fp32_bits)->RangeMultiplier(2)->Range(1<<10, 64<<20);

static void fp16_alt_to_fp32_value(benchmark::State& state) {
	const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
	auto rng = std::bind(std::uniform_int_distribution<uint16_t>(0, 0x7BFF), std::mt19937(seed));

	std::vector<uint16_t> fp16(state.range(0));
	std::vector<float> fp32(state.range(0));
	std::generate(fp16.begin(), fp16.end(),
		[&rng]{ return fp16_alt_from_fp32_value(rng()); });

	while (state.KeepRunning()) {
		uint16_t* input = fp16.data();
		benchmark::DoNotOptimize(input);

		float* output = fp32.data();
		const size_t n = state.range(0);
		for (size_t i = 0; i < n; i++) {
			output[i] = fp16_alt_to_fp32_value(input[i]);
		}

		benchmark::DoNotOptimize(output);
	}
	state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
}
BENCHMARK(fp16_alt_to_fp32_value)->RangeMultiplier(2)->Range(1<<10, 64<<20);

#ifndef EMSCRIPTEN
	static void fp16_alt_to_fp32_psimd(benchmark::State& state) {
		const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
		auto rng = std::bind(std::uniform_int_distribution<uint16_t>(0, 0x7BFF), std::mt19937(seed));

		std::vector<uint16_t> fp16(state.range(0));
		std::vector<float> fp32(state.range(0));
		std::generate(fp16.begin(), fp16.end(),
			[&rng]{ return fp16_alt_from_fp32_value(rng()); });

		while (state.KeepRunning()) {
			uint16_t* input = fp16.data();
			benchmark::DoNotOptimize(input);

			float* output = fp32.data();
			const size_t n = state.range(0);
			for (size_t i = 0; i < n - 4; i += 4) {
				psimd_store_f32(&output[i],
					fp16_alt_to_fp32_psimd(
						psimd_load_u16(&input[i])));
			}
			const psimd_u16 last_vector = { input[n - 4], input[n - 3], input[n - 2], input[n - 1] };
			psimd_store_f32(&output[n - 4],
				fp16_alt_to_fp32_psimd(last_vector));

			benchmark::DoNotOptimize(output);
		}
		state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
	}
	BENCHMARK(fp16_alt_to_fp32_psimd)->RangeMultiplier(2)->Range(1<<10, 64<<20);

	static void fp16_alt_to_fp32x2_psimd(benchmark::State& state) {
		const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
		auto rng = std::bind(std::uniform_int_distribution<uint16_t>(0, 0x7BFF), std::mt19937(seed));

		std::vector<uint16_t> fp16(state.range(0));
		std::vector<float> fp32(state.range(0));
		std::generate(fp16.begin(), fp16.end(),
			[&rng]{ return fp16_alt_from_fp32_value(rng()); });

		while (state.KeepRunning()) {
			uint16_t* input = fp16.data();
			benchmark::DoNotOptimize(input);

			float* output = fp32.data();
			const size_t n = state.range(0);
			for (size_t i = 0; i < n; i += 8) {
				const psimd_f32x2 data =
					fp16_alt_to_fp32x2_psimd(
						psimd_load_u16(&input[i]));
				psimd_store_f32(&output[i], data.lo);
				psimd_store_f32(&output[i + 4], data.hi);
			}

			benchmark::DoNotOptimize(output);
		}
		state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
	}
	BENCHMARK(fp16_alt_to_fp32x2_psimd)->RangeMultiplier(2)->Range(1<<10, 64<<20);
#endif

#if defined(__ARM_NEON_FP) && (__ARM_NEON_FP & 0x2) || defined(__aarch64__)
	static void hardware_vcvt_f32_f16(benchmark::State& state) {
		const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
		auto rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), std::mt19937(seed));

		std::vector<uint16_t> fp16(state.range(0));
		std::vector<float> fp32(state.range(0));
		std::generate(fp16.begin(), fp16.end(),
			[&rng]{ return fp16_ieee_from_fp32_value(rng()); });

		while (state.KeepRunning()) {
			uint16_t* input = fp16.data();
			benchmark::DoNotOptimize(input);

			float* output = fp32.data();
			const size_t n = state.range(0);
			#if defined(__aarch64__)
				const unsigned int fpcr = __builtin_aarch64_get_fpcr();
				/* Disable flush-to-zero (bit 24) and enable Alternative FP16 format (bit 26) */
				__builtin_aarch64_set_fpcr((fpcr & 0xFEFFFFFFu) | 0x08000000u);
			#else
				unsigned int fpscr;
				__asm__ __volatile__ ("VMRS %[fpscr], fpscr" : [fpscr] "=r" (fpscr));
				/* Disable flush-to-zero (bit 24) and enable Alternative FP16 format (bit 26) */
				__asm__ __volatile__ ("VMSR fpscr, %[fpscr]" :
					: [fpscr] "r" ((fpscr & 0xFEFFFFFFu) | 0x08000000u));
			#endif
			for (size_t i = 0; i < n; i += 4) {
				vst1q_f32(&output[i],
					vcvt_f32_f16(
						(float16x4_t) vld1_u16(&input[i])));
			}
			#if defined(__aarch64__)
				__builtin_aarch64_set_fpcr(fpcr);
			#else
				__asm__ __volatile__ ("VMSR fpscr, %[fpscr]" :: [fpscr] "r" (fpscr));
			#endif

			benchmark::DoNotOptimize(output);
		}
		state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
	}
	BENCHMARK(hardware_vcvt_f32_f16)->RangeMultiplier(2)->Range(1<<10, 64<<20);
#endif

BENCHMARK_MAIN();