| 1 | /* |
|---|
| 2 | * Copyright (C) 2007 Michael Lewis |
|---|
| 3 | * Author: Mike Lewis <mikelikespie@gmail.com> |
|---|
| 4 | * |
|---|
| 5 | * This work is provide AS IS, and has no warranty. |
|---|
| 6 | * The author is NOT responsible for anything that happens |
|---|
| 7 | * due to use of this code, either using it or running it on |
|---|
| 8 | * your system |
|---|
| 9 | */ |
|---|
| 10 | #ifndef _FAKE_SSE_INCLUDED |
|---|
| 11 | #define _FAKE_SSE_INCLUDED |
|---|
| 12 | #ifndef __SSE2__ |
|---|
| 13 | #include <stdlib.h> |
|---|
| 14 | #include <string.h> |
|---|
| 15 | |
|---|
| 16 | #ifdef __APPLE__ |
|---|
| 17 | #include <inttypes.h> |
|---|
| 18 | #endif |
|---|
| 19 | /* |
|---|
| 20 | typedef struct { |
|---|
| 21 | union { |
|---|
| 22 | unsigned char byte[16]; |
|---|
| 23 | uint32_t int32[4]; |
|---|
| 24 | } u; |
|---|
| 25 | } __m128i; |
|---|
| 26 | */ |
|---|
| 27 | //Shifts this to the left all the way |
|---|
| 28 | static inline __m128i _mm_srli_epi32( __m128i A, int count ) { |
|---|
| 29 | //printf("_mm_srli_epi32\n"); |
|---|
| 30 | |
|---|
| 31 | int i; |
|---|
| 32 | for( i = 0; i < 4; i++ ) { |
|---|
| 33 | A.u.int32[i] = A.u.int32[i] >> count; |
|---|
| 34 | } |
|---|
| 35 | return A; |
|---|
| 36 | } |
|---|
| 37 | |
|---|
| 38 | static inline __m128i _mm_slli_epi32( __m128i A, int count ) { |
|---|
| 39 | //printf("_mm_slli_epi32\n"); |
|---|
| 40 | int i; |
|---|
| 41 | for( i = 0; i < 4; i++ ) { |
|---|
| 42 | A.u.int32[i] = A.u.int32[i] << count; |
|---|
| 43 | } |
|---|
| 44 | return A; |
|---|
| 45 | } |
|---|
| 46 | |
|---|
| 47 | static inline __m128i _mm_slli_si128( __m128i A, int imm ) { |
|---|
| 48 | //printf("_mm_srli_si128\n"); |
|---|
| 49 | int i; |
|---|
| 50 | __m128i ret; |
|---|
| 51 | for( i = imm; i < 16; i++ ) { |
|---|
| 52 | ret.u.byte[i] = A.u.byte[i-imm]; |
|---|
| 53 | } |
|---|
| 54 | for( i = 0; i < imm; i++ ) { |
|---|
| 55 | ret.u.byte[i] = 0x00; |
|---|
| 56 | } |
|---|
| 57 | return ret; |
|---|
| 58 | } |
|---|
| 59 | |
|---|
| 60 | static inline __m128i _mm_srli_si128( __m128i A, int imm ) { |
|---|
| 61 | //printf("_mm_slli_si128\n"); |
|---|
| 62 | int i; |
|---|
| 63 | __m128i ret; |
|---|
| 64 | for( i = 0; i < 16-imm; i++ ) { |
|---|
| 65 | ret.u.byte[i] = A.u.byte[i+imm]; |
|---|
| 66 | } |
|---|
| 67 | for( i = 16-imm; i < 16; i++ ) { |
|---|
| 68 | ret.u.byte[i] = 0x00; |
|---|
| 69 | } |
|---|
| 70 | return ret; |
|---|
| 71 | } |
|---|
| 72 | |
|---|
| 73 | static inline __m128i _mm_setzero_si128() { |
|---|
| 74 | //printf("_mm_setzero_si128\n"); |
|---|
| 75 | __m128i r; |
|---|
| 76 | memset( &r, 0, sizeof( r )); |
|---|
| 77 | return r; |
|---|
| 78 | } |
|---|
| 79 | |
|---|
| 80 | static inline __m128i _mm_cmpeq_epi8( __m128i A, __m128i B ) { |
|---|
| 81 | //printf("_mm_cmpeq_epi8\n"); |
|---|
| 82 | __m128i ret; |
|---|
| 83 | int i; |
|---|
| 84 | for( i = 0; i < 16; i++ ) { |
|---|
| 85 | ret.u.byte[i] = ( A.u.byte[i] == B.u.byte[i] ) ? 0xFF : 0x00; |
|---|
| 86 | } |
|---|
| 87 | return ret; |
|---|
| 88 | } |
|---|
| 89 | static inline __m128i _mm_add_epi8( __m128i A, __m128i B ) { |
|---|
| 90 | //printf("_mm_add_epi8\n"); |
|---|
| 91 | __m128i ret; |
|---|
| 92 | int i; |
|---|
| 93 | for( i = 0; i < 16; i++ ) { |
|---|
| 94 | ret.u.byte[i] = A.u.byte[i] + B.u.byte[i]; |
|---|
| 95 | } |
|---|
| 96 | return ret; |
|---|
| 97 | } |
|---|
| 98 | |
|---|
| 99 | static inline __m128i _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12, |
|---|
| 100 | char __q11, char __q10, char __q09, char __q08, |
|---|
| 101 | char __q07, char __q06, char __q05, char __q04, |
|---|
| 102 | char __q03, char __q02, char __q01, char __q00) |
|---|
| 103 | { |
|---|
| 104 | //printf("_mm_set_epi8 \n"); |
|---|
| 105 | __m128i ret; |
|---|
| 106 | ret.u.byte[15] = __q15; |
|---|
| 107 | ret.u.byte[14] = __q14; |
|---|
| 108 | ret.u.byte[13] = __q13; |
|---|
| 109 | ret.u.byte[12] = __q12; |
|---|
| 110 | ret.u.byte[11] = __q11; |
|---|
| 111 | ret.u.byte[10] = __q10; |
|---|
| 112 | ret.u.byte[9] = __q09; |
|---|
| 113 | ret.u.byte[8] = __q08; |
|---|
| 114 | ret.u.byte[7] = __q07; |
|---|
| 115 | ret.u.byte[6] = __q06; |
|---|
| 116 | ret.u.byte[5] = __q05; |
|---|
| 117 | ret.u.byte[4] = __q04; |
|---|
| 118 | ret.u.byte[3] = __q03; |
|---|
| 119 | ret.u.byte[2] = __q02; |
|---|
| 120 | ret.u.byte[1] = __q01; |
|---|
| 121 | ret.u.byte[0] = __q00; |
|---|
| 122 | return ret; |
|---|
| 123 | } |
|---|
| 124 | static inline __m128i _mm_set1_epi8 ( char f ) { |
|---|
| 125 | //printf("_mm_set1_epi8 \n"); |
|---|
| 126 | int i; |
|---|
| 127 | __m128i ret; |
|---|
| 128 | for( i = 0; i < 16; i++ ) { |
|---|
| 129 | ret.u.byte[i] = f; |
|---|
| 130 | } |
|---|
| 131 | return ret; |
|---|
| 132 | } |
|---|
| 133 | |
|---|
| 134 | |
|---|
| 135 | static inline __m128i _mm_and_si128( __m128i A, __m128i B ) { |
|---|
| 136 | //printf("_mm_and_si128\n"); |
|---|
| 137 | __m128i ret; |
|---|
| 138 | int i; |
|---|
| 139 | for( i = 0; i < 4; i++ ) { |
|---|
| 140 | ret.u.int32[i] = A.u.int32[i] & B.u.int32[i]; |
|---|
| 141 | } |
|---|
| 142 | return ret; |
|---|
| 143 | } |
|---|
| 144 | |
|---|
| 145 | static inline __m128i _mm_or_si128( __m128i A, __m128i B ) { |
|---|
| 146 | //printf("_mm_or_si128\n"); |
|---|
| 147 | __m128i ret; |
|---|
| 148 | int i; |
|---|
| 149 | for( i = 0; i < 4; i++ ) { |
|---|
| 150 | ret.u.int32[i] = A.u.int32[i] | B.u.int32[i]; |
|---|
| 151 | } |
|---|
| 152 | return ret; |
|---|
| 153 | } |
|---|
| 154 | |
|---|
| 155 | #endif |
|---|
| 156 | #endif //_FAKE_SSE_INCLUDED |
|---|