gfx/qcms/transform-sse1.cpp

   1 #include <xmmintrin.h>
   2
   3 #include "qcmsint.h"
   4
   5 /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
   6 static const ALIGN float floatScaleX4[4] =
   7     { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
   8 static const ALIGN float clampMaxValueX4[4] =
   9     { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
  10
  11 template <size_t kRIndex, size_t kGIndex, size_t kBIndex, size_t kAIndex = NO_A_INDEX>
  12 static void qcms_transform_data_template_lut_sse1(const qcms_transform *transform,
  13                                                   const unsigned char *src,
  14                                                   unsigned char *dest,
  15                                                   size_t length)
  16 {
  17     unsigned int i;
  18     const float (*mat)[4] = transform->matrix;
  19     char input_back[32];
  20     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
  21      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
  22      * because they don't work on stack variables. gcc 4.4 does do the right thing
  23      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  24     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
  25     /* share input and output locations to save having to keep the
  26      * locations in separate registers */
  27     uint32_t const * output = (uint32_t*)input;
  28
  29     /* deref *transform now to avoid it in loop */
  30     const float *igtbl_r = transform->input_gamma_table_r;
  31     const float *igtbl_g = transform->input_gamma_table_g;
  32     const float *igtbl_b = transform->input_gamma_table_b;
  33
  34     /* deref *transform now to avoid it in loop */
  35     const uint8_t *otdata_r = &transform->output_table_r->data[0];
  36     const uint8_t *otdata_g = &transform->output_table_g->data[0];
  37     const uint8_t *otdata_b = &transform->output_table_b->data[0];
  38
  39     /* input matrix values never change */
  40     const __m128 mat0  = _mm_load_ps(mat[0]);
  41     const __m128 mat1  = _mm_load_ps(mat[1]);
  42     const __m128 mat2  = _mm_load_ps(mat[2]);
  43
  44     /* these values don't change, either */
  45     const __m128 max   = _mm_load_ps(clampMaxValueX4);
  46     const __m128 min   = _mm_setzero_ps();
  47     const __m128 scale = _mm_load_ps(floatScaleX4);
  48     const unsigned int components = A_INDEX_COMPONENTS(kAIndex);
  49
  50     /* working variables */
  51     __m128 vec_r, vec_g, vec_b, result;
  52     unsigned char alpha;
  53
  54     /* CYA */
  55     if (!length)
  56         return;
  57
  58     /* one pixel is handled outside of the loop */
  59     length--;
  60
  61     /* setup for transforming 1st pixel */
  62     vec_r = _mm_load_ss(&igtbl_r[src[kRIndex]]);
  63     vec_g = _mm_load_ss(&igtbl_g[src[kGIndex]]);
  64     vec_b = _mm_load_ss(&igtbl_b[src[kBIndex]]);
  65     if (kAIndex != NO_A_INDEX) {
  66         alpha = src[kAIndex];
  67     }
  68     src += components;
  69
  70     /* transform all but final pixel */
  71
  72     for (i=0; i<length; i++)
  73     {
  74         /* position values from gamma tables */
  75         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
  76         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
  77         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
  78
  79         /* gamma * matrix */
  80         vec_r = _mm_mul_ps(vec_r, mat0);
  81         vec_g = _mm_mul_ps(vec_g, mat1);
  82         vec_b = _mm_mul_ps(vec_b, mat2);
  83
  84         /* store alpha for this pixel; load alpha for next */
  85         if (kAIndex != NO_A_INDEX) {
  86             dest[kAIndex] = alpha;
  87             alpha = src[kAIndex];
  88         }
  89
  90         /* crunch, crunch, crunch */
  91         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
  92         vec_r  = _mm_max_ps(min, vec_r);
  93         vec_r  = _mm_min_ps(max, vec_r);
  94         result = _mm_mul_ps(vec_r, scale);
  95
  96         /* store calc'd output tables indices */
  97         *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
  98         result = _mm_movehl_ps(result, result);
  99         *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
 100
 101         /* load gamma values for next loop while store completes */
 102         vec_r = _mm_load_ss(&igtbl_r[src[kRIndex]]);
 103         vec_g = _mm_load_ss(&igtbl_g[src[kGIndex]]);
 104         vec_b = _mm_load_ss(&igtbl_b[src[kBIndex]]);
 105         src += components;
 106
 107         /* use calc'd indices to output RGB values */
 108         dest[kRIndex] = otdata_r[output[0]];
 109         dest[kGIndex] = otdata_g[output[1]];
 110         dest[kBIndex] = otdata_b[output[2]];
 111         dest += components;
 112     }
 113
 114     /* handle final (maybe only) pixel */
 115
 116     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 117     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 118     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 119
 120     vec_r = _mm_mul_ps(vec_r, mat0);
 121     vec_g = _mm_mul_ps(vec_g, mat1);
 122     vec_b = _mm_mul_ps(vec_b, mat2);
 123
 124     if (kAIndex != NO_A_INDEX) {
 125         dest[kAIndex] = alpha;
 126     }
 127
 128     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 129     vec_r  = _mm_max_ps(min, vec_r);
 130     vec_r  = _mm_min_ps(max, vec_r);
 131     result = _mm_mul_ps(vec_r, scale);
 132
 133     *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
 134     result = _mm_movehl_ps(result, result);
 135     *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
 136
 137     dest[kRIndex] = otdata_r[output[0]];
 138     dest[kGIndex] = otdata_g[output[1]];
 139     dest[kBIndex] = otdata_b[output[2]];
 140
 141     _mm_empty();
 142 }
 143
 144 void qcms_transform_data_rgb_out_lut_sse1(const qcms_transform *transform,
 145                                           const unsigned char *src,
 146                                           unsigned char *dest,
 147                                           size_t length)
 148 {
 149     qcms_transform_data_template_lut_sse1<RGBA_R_INDEX, RGBA_G_INDEX, RGBA_B_INDEX>(transform, src, dest, length);
 150 }
 151
 152 void qcms_transform_data_rgba_out_lut_sse1(const qcms_transform *transform,
 153                                            const unsigned char *src,
 154                                            unsigned char *dest,
 155                                            size_t length)
 156 {
 157     qcms_transform_data_template_lut_sse1<RGBA_R_INDEX, RGBA_G_INDEX, RGBA_B_INDEX, RGBA_A_INDEX>(transform, src, dest, length);
 158 }
 159
 160 void qcms_transform_data_bgra_out_lut_sse1(const qcms_transform *transform,
 161                                            const unsigned char *src,
 162                                            unsigned char *dest,
 163                                            size_t length)
 164 {
 165     qcms_transform_data_template_lut_sse1<BGRA_R_INDEX, BGRA_G_INDEX, BGRA_B_INDEX, BGRA_A_INDEX>(transform, src, dest, length);
 166 }