| 1 | /* Copyright (C) 2010-2021 The RetroArch team |
| 2 | * |
| 3 | * --------------------------------------------------------------------------------------- |
| 4 | * The following license statement only applies to this file (float_to_s16.c). |
| 5 | * --------------------------------------------------------------------------------------- |
| 6 | * |
| 7 | * Permission is hereby granted, free of charge, |
| 8 | * to any person obtaining a copy of this software and associated documentation files (the "Software"), |
| 9 | * to deal in the Software without restriction, including without limitation the rights to |
| 10 | * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, |
| 11 | * and to permit persons to whom the Software is furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, |
| 16 | * INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. |
| 18 | * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, |
| 19 | * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 20 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. |
| 21 | */ |
| 22 | #include <stdint.h> |
| 23 | #include <stddef.h> |
| 24 | |
| 25 | #if defined(__SSE2__) |
| 26 | #include <emmintrin.h> |
| 27 | #elif defined(__ALTIVEC__) |
| 28 | #include <altivec.h> |
| 29 | #endif |
| 30 | |
| 31 | #include <features/features_cpu.h> |
| 32 | #include <audio/conversion/float_to_s16.h> |
| 33 | |
| 34 | #if (defined(__ARM_NEON__) || defined(HAVE_NEON)) |
| 35 | static bool float_to_s16_neon_enabled = false; |
| 36 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 37 | void convert_float_s16_asm(int16_t *out, |
| 38 | const float *in, size_t samples); |
| 39 | #else |
| 40 | #include <arm_neon.h> |
| 41 | #endif |
| 42 | |
| 43 | void convert_float_to_s16(int16_t *out, |
| 44 | const float *in, size_t samples) |
| 45 | { |
| 46 | size_t i = 0; |
| 47 | if (float_to_s16_neon_enabled) |
| 48 | { |
| 49 | float gf = (1<<15); |
| 50 | float32x4_t vgf = {gf, gf, gf, gf}; |
| 51 | while (samples >= 8) |
| 52 | { |
| 53 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 54 | size_t aligned_samples = samples & ~7; |
| 55 | if (aligned_samples) |
| 56 | convert_float_s16_asm(out, in, aligned_samples); |
| 57 | |
| 58 | out += aligned_samples; |
| 59 | in += aligned_samples; |
| 60 | samples -= aligned_samples; |
| 61 | i = 0; |
| 62 | #else |
| 63 | int16x4x2_t oreg; |
| 64 | int32x4x2_t creg; |
| 65 | float32x4x2_t inreg = vld2q_f32(in); |
| 66 | creg.val[0] = vcvtq_s32_f32(vmulq_f32(inreg.val[0], vgf)); |
| 67 | creg.val[1] = vcvtq_s32_f32(vmulq_f32(inreg.val[1], vgf)); |
| 68 | oreg.val[0] = vqmovn_s32(creg.val[0]); |
| 69 | oreg.val[1] = vqmovn_s32(creg.val[1]); |
| 70 | vst2_s16(out, oreg); |
| 71 | in += 8; |
| 72 | out += 8; |
| 73 | samples -= 8; |
| 74 | #endif |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | for (; i < samples; i++) |
| 79 | { |
| 80 | int32_t val = (int32_t)(in[i] * 0x8000); |
| 81 | out[i] = (val > 0x7FFF) ? 0x7FFF : |
| 82 | (val < -0x8000 ? -0x8000 : (int16_t)val); |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | void convert_float_to_s16_init_simd(void) |
| 87 | { |
| 88 | uint64_t cpu = cpu_features_get(); |
| 89 | |
| 90 | if (cpu & RETRO_SIMD_NEON) |
| 91 | float_to_s16_neon_enabled = true; |
| 92 | } |
| 93 | #else |
| 94 | void convert_float_to_s16(int16_t *out, |
| 95 | const float *in, size_t samples) |
| 96 | { |
| 97 | size_t i = 0; |
| 98 | #if defined(__SSE2__) |
| 99 | __m128 factor = _mm_set1_ps((float)0x8000); |
| 100 | /* Initialize a 4D vector with 32768.0 for its elements */ |
| 101 | |
| 102 | for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8) |
| 103 | { /* Skip forward 8 samples at a time... */ |
| 104 | __m128 input_a = _mm_loadu_ps(in + 0); /* Create a 4-float vector from the next four samples... */ |
| 105 | __m128 input_b = _mm_loadu_ps(in + 4); /* ...and another from the *next* next four. */ |
| 106 | __m128 res_a = _mm_mul_ps(input_a, factor); |
| 107 | __m128 res_b = _mm_mul_ps(input_b, factor); /* Multiply these samples by 32768 */ |
| 108 | __m128i ints_a = _mm_cvtps_epi32(res_a); |
| 109 | __m128i ints_b = _mm_cvtps_epi32(res_b); /* Convert the samples to 32-bit integers */ |
| 110 | __m128i packed = _mm_packs_epi32(ints_a, ints_b); /* Then convert them to 16-bit ints, clamping to [-32768, 32767] */ |
| 111 | |
| 112 | _mm_storeu_si128((__m128i *)out, packed); /* Then put the result in the output array */ |
| 113 | } |
| 114 | |
| 115 | samples = samples - i; |
| 116 | i = 0; |
| 117 | /* If there are any stray samples at the end, we need to convert them |
| 118 | * (maybe the original array didn't contain a multiple of 8 samples) */ |
| 119 | #elif defined(__ALTIVEC__) |
| 120 | int samples_in = samples; |
| 121 | |
| 122 | /* Unaligned loads/store is a bit expensive, |
| 123 | * so we optimize for the good path (very likely). */ |
| 124 | if (((uintptr_t)out & 15) + ((uintptr_t)in & 15) == 0) |
| 125 | { |
| 126 | size_t i; |
| 127 | for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8) |
| 128 | { |
| 129 | vector float input0 = vec_ld( 0, in); |
| 130 | vector float input1 = vec_ld(16, in); |
| 131 | vector signed int result0 = vec_cts(input0, 15); |
| 132 | vector signed int result1 = vec_cts(input1, 15); |
| 133 | vec_st(vec_packs(result0, result1), 0, out); |
| 134 | } |
| 135 | |
| 136 | samples_in -= i; |
| 137 | } |
| 138 | |
| 139 | samples = samples_in; |
| 140 | i = 0; |
| 141 | #elif defined(_MIPS_ARCH_ALLEGREX) |
| 142 | #ifdef DEBUG |
| 143 | /* Make sure the buffers are 16 byte aligned, this should be |
| 144 | * the default behaviour of malloc in the PSPSDK. |
| 145 | * Assume alignment. */ |
| 146 | retro_assert(((uintptr_t)in & 0xf) == 0); |
| 147 | retro_assert(((uintptr_t)out & 0xf) == 0); |
| 148 | #endif |
| 149 | |
| 150 | for (i = 0; i + 8 <= samples; i += 8) |
| 151 | { |
| 152 | __asm__ ( |
| 153 | ".set push \n" |
| 154 | ".set noreorder \n" |
| 155 | |
| 156 | "lv.q c100, 0(%0) \n" |
| 157 | "lv.q c110, 16(%0) \n" |
| 158 | |
| 159 | "vf2in.q c100, c100, 31 \n" |
| 160 | "vf2in.q c110, c110, 31 \n" |
| 161 | "vi2s.q c100, c100 \n" |
| 162 | "vi2s.q c102, c110 \n" |
| 163 | |
| 164 | "sv.q c100, 0(%1) \n" |
| 165 | |
| 166 | ".set pop \n" |
| 167 | :: "r"(in + i), "r"(out + i)); |
| 168 | } |
| 169 | #endif |
| 170 | |
| 171 | /* This loop converts stray samples to the right format, |
| 172 | * but it's also a fallback in case no SIMD instructions are available. */ |
| 173 | for (; i < samples; i++) |
| 174 | { |
| 175 | int32_t val = (int32_t)(in[i] * 0x8000); |
| 176 | out[i] = (val > 0x7FFF) |
| 177 | ? 0x7FFF |
| 178 | : (val < -0x8000 ? -0x8000 : (int16_t)val); |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | void convert_float_to_s16_init_simd(void) { } |
| 183 | #endif |