| 1 | /* Copyright (C) 2010-2021 The RetroArch team |
| 2 | * |
| 3 | * --------------------------------------------------------------------------------------- |
| 4 | * The following license statement only applies to this file (s16_to_float.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 | #if defined(__SSE2__) |
| 23 | #include <emmintrin.h> |
| 24 | #elif defined(__ALTIVEC__) |
| 25 | #include <altivec.h> |
| 26 | #endif |
| 27 | |
| 28 | #include <boolean.h> |
| 29 | #include <features/features_cpu.h> |
| 30 | #include <audio/conversion/s16_to_float.h> |
| 31 | |
| 32 | #if (defined(__ARM_NEON__) || defined(HAVE_NEON)) |
| 33 | static bool s16_to_float_neon_enabled = false; |
| 34 | |
| 35 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 36 | /* Avoid potential hard-float/soft-float ABI issues. */ |
| 37 | void convert_s16_float_asm(float *out, const int16_t *in, |
| 38 | size_t samples, const float *gain); |
| 39 | #else |
| 40 | #include <arm_neon.h> |
| 41 | #endif |
| 42 | |
| 43 | void convert_s16_to_float(float *out, |
| 44 | const int16_t *in, size_t samples, float gain) |
| 45 | { |
| 46 | unsigned i = 0; |
| 47 | |
| 48 | if (s16_to_float_neon_enabled) |
| 49 | { |
| 50 | #ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS |
| 51 | size_t aligned_samples = samples & ~7; |
| 52 | if (aligned_samples) |
| 53 | convert_s16_float_asm(out, in, aligned_samples, &gain); |
| 54 | |
| 55 | /* Could do all conversion in ASM, but keep it simple for now. */ |
| 56 | out += aligned_samples; |
| 57 | in += aligned_samples; |
| 58 | samples -= aligned_samples; |
| 59 | i = 0; |
| 60 | #else |
| 61 | float gf = gain / (1 << 15); |
| 62 | float32x4_t vgf = {gf, gf, gf, gf}; |
| 63 | while (samples >= 8) |
| 64 | { |
| 65 | float32x4x2_t oreg; |
| 66 | int16x4x2_t inreg = vld2_s16(in); |
| 67 | int32x4_t p1 = vmovl_s16(inreg.val[0]); |
| 68 | int32x4_t p2 = vmovl_s16(inreg.val[1]); |
| 69 | oreg.val[0] = vmulq_f32(vcvtq_f32_s32(p1), vgf); |
| 70 | oreg.val[1] = vmulq_f32(vcvtq_f32_s32(p2), vgf); |
| 71 | vst2q_f32(out, oreg); |
| 72 | in += 8; |
| 73 | out += 8; |
| 74 | samples -= 8; |
| 75 | } |
| 76 | #endif |
| 77 | } |
| 78 | |
| 79 | gain /= 0x8000; |
| 80 | |
| 81 | for (; i < samples; i++) |
| 82 | out[i] = (float)in[i] * gain; |
| 83 | } |
| 84 | |
| 85 | void convert_s16_to_float_init_simd(void) |
| 86 | { |
| 87 | uint64_t cpu = cpu_features_get(); |
| 88 | |
| 89 | if (cpu & RETRO_SIMD_NEON) |
| 90 | s16_to_float_neon_enabled = true; |
| 91 | } |
| 92 | #else |
| 93 | void convert_s16_to_float(float *out, |
| 94 | const int16_t *in, size_t samples, float gain) |
| 95 | { |
| 96 | unsigned i = 0; |
| 97 | |
| 98 | #if defined(__SSE2__) |
| 99 | float fgain = gain / UINT32_C(0x80000000); |
| 100 | __m128 factor = _mm_set1_ps(fgain); |
| 101 | |
| 102 | for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8) |
| 103 | { |
| 104 | __m128i input = _mm_loadu_si128((const __m128i *)in); |
| 105 | __m128i regs_l = _mm_unpacklo_epi16(_mm_setzero_si128(), input); |
| 106 | __m128i regs_r = _mm_unpackhi_epi16(_mm_setzero_si128(), input); |
| 107 | __m128 output_l = _mm_mul_ps(_mm_cvtepi32_ps(regs_l), factor); |
| 108 | __m128 output_r = _mm_mul_ps(_mm_cvtepi32_ps(regs_r), factor); |
| 109 | |
| 110 | _mm_storeu_ps(out + 0, output_l); |
| 111 | _mm_storeu_ps(out + 4, output_r); |
| 112 | } |
| 113 | |
| 114 | samples = samples - i; |
| 115 | i = 0; |
| 116 | #elif defined(__ALTIVEC__) |
| 117 | size_t samples_in = samples; |
| 118 | |
| 119 | /* Unaligned loads/store is a bit expensive, so we |
| 120 | * optimize for the good path (very likely). */ |
| 121 | if (((uintptr_t)out & 15) + ((uintptr_t)in & 15) == 0) |
| 122 | { |
| 123 | const vector float gain_vec = { gain, gain , gain, gain }; |
| 124 | const vector float zero_vec = { 0.0f, 0.0f, 0.0f, 0.0f}; |
| 125 | |
| 126 | for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8) |
| 127 | { |
| 128 | vector signed short input = vec_ld(0, in); |
| 129 | vector signed int hi = vec_unpackh(input); |
| 130 | vector signed int lo = vec_unpackl(input); |
| 131 | vector float out_hi = vec_madd(vec_ctf(hi, 15), gain_vec, zero_vec); |
| 132 | vector float out_lo = vec_madd(vec_ctf(lo, 15), gain_vec, zero_vec); |
| 133 | |
| 134 | vec_st(out_hi, 0, out); |
| 135 | vec_st(out_lo, 16, out); |
| 136 | } |
| 137 | |
| 138 | samples_in -= i; |
| 139 | } |
| 140 | |
| 141 | samples = samples_in; |
| 142 | i = 0; |
| 143 | #endif |
| 144 | |
| 145 | gain /= 0x8000; |
| 146 | |
| 147 | #if defined(_MIPS_ARCH_ALLEGREX) |
| 148 | #ifdef DEBUG |
| 149 | /* Make sure the buffer is 16 byte aligned, this should be the |
| 150 | * default behaviour of malloc in the PSPSDK. |
| 151 | * Only the output buffer can be assumed to be 16-byte aligned. */ |
| 152 | retro_assert(((uintptr_t)out & 0xf) == 0); |
| 153 | #endif |
| 154 | |
| 155 | __asm__ ( |
| 156 | ".set push \n" |
| 157 | ".set noreorder \n" |
| 158 | "mtv %0, s200 \n" |
| 159 | ".set pop \n" |
| 160 | ::"r"(gain)); |
| 161 | |
| 162 | for (i = 0; i + 16 <= samples; i += 16) |
| 163 | { |
| 164 | __asm__ ( |
| 165 | ".set push \n" |
| 166 | ".set noreorder \n" |
| 167 | |
| 168 | "lv.s s100, 0(%0) \n" |
| 169 | "lv.s s101, 4(%0) \n" |
| 170 | "lv.s s110, 8(%0) \n" |
| 171 | "lv.s s111, 12(%0) \n" |
| 172 | "lv.s s120, 16(%0) \n" |
| 173 | "lv.s s121, 20(%0) \n" |
| 174 | "lv.s s130, 24(%0) \n" |
| 175 | "lv.s s131, 28(%0) \n" |
| 176 | |
| 177 | "vs2i.p c100, c100 \n" |
| 178 | "vs2i.p c110, c110 \n" |
| 179 | "vs2i.p c120, c120 \n" |
| 180 | "vs2i.p c130, c130 \n" |
| 181 | |
| 182 | "vi2f.q c100, c100, 16 \n" |
| 183 | "vi2f.q c110, c110, 16 \n" |
| 184 | "vi2f.q c120, c120, 16 \n" |
| 185 | "vi2f.q c130, c130, 16 \n" |
| 186 | |
| 187 | "vmscl.q e100, e100, s200 \n" |
| 188 | |
| 189 | "sv.q c100, 0(%1) \n" |
| 190 | "sv.q c110, 16(%1) \n" |
| 191 | "sv.q c120, 32(%1) \n" |
| 192 | "sv.q c130, 48(%1) \n" |
| 193 | |
| 194 | ".set pop \n" |
| 195 | :: "r"(in + i), "r"(out + i)); |
| 196 | } |
| 197 | #endif |
| 198 | |
| 199 | for (; i < samples; i++) |
| 200 | out[i] = (float)in[i] * gain; |
| 201 | } |
| 202 | |
| 203 | void convert_s16_to_float_init_simd(void) { } |
| 204 | #endif |
| 205 | |