gfx/qcms/transform-avx.cpp

   1 #include <emmintrin.h>
   2 #include <immintrin.h>
   3
   4 #include "qcmsint.h"
   5
   6 template <size_t kRIndex, size_t kGIndex, size_t kBIndex, size_t kAIndex = NO_A_INDEX>
   7 static void qcms_transform_data_template_lut_avx(const qcms_transform *transform,
   8                                                  const unsigned char *src,
   9                                                  unsigned char *dest,
  10                                                  size_t length)
  11 {
  12     const float (*mat)[4] = transform->matrix;
  13     char input_back[64];
  14     /* Ensure we have a buffer that's 32 byte aligned regardless of the original
  15      * stack alignment. We can't use __attribute__((aligned(32))) or __declspec(align(32))
  16      * because they don't work on stack variables. gcc 4.4 does do the right thing
  17      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  18     float const * input = (float*)(((uintptr_t)&input_back[32]) & ~0x1f);
  19     /* share input and output locations to save having to keep the
  20      * locations in separate registers */
  21     uint32_t const * output = (uint32_t*)input;
  22
  23     /* deref *transform now to avoid it in loop */
  24     const float *igtbl_r = transform->input_gamma_table_r;
  25     const float *igtbl_g = transform->input_gamma_table_g;
  26     const float *igtbl_b = transform->input_gamma_table_b;
  27
  28     /* deref *transform now to avoid it in loop */
  29     const uint8_t *otdata_r = &transform->output_table_r->data[0];
  30     const uint8_t *otdata_g = &transform->output_table_g->data[0];
  31     const uint8_t *otdata_b = &transform->output_table_b->data[0];
  32
  33     /* input matrix values never change */
  34     const __m256 mat0  = _mm256_broadcast_ps(reinterpret_cast<const __m128*>(mat[0]));
  35     const __m256 mat1  = _mm256_broadcast_ps(reinterpret_cast<const __m128*>(mat[1]));
  36     const __m256 mat2  = _mm256_broadcast_ps(reinterpret_cast<const __m128*>(mat[2]));
  37
  38     /* these values don't change, either */
  39     const __m256 max   = _mm256_set1_ps(CLAMPMAXVAL);
  40     const __m256 min   = _mm256_setzero_ps();
  41     const __m256 scale = _mm256_set1_ps(FLOATSCALE);
  42     const unsigned int components = A_INDEX_COMPONENTS(kAIndex);
  43
  44     /* working variables */
  45     __m256 vec_r, vec_g, vec_b, result;
  46     __m128 vec_r0, vec_g0, vec_b0, vec_r1, vec_g1, vec_b1;
  47     unsigned char alpha1, alpha2;
  48
  49     /* CYA */
  50     if (!length)
  51         return;
  52
  53     /* If there are at least 2 pixels, then we can load their components into
  54        a single 256-bit register for processing. */
  55     if (length > 1) {
  56         vec_r0 = _mm_broadcast_ss(&igtbl_r[src[kRIndex]]);
  57         vec_g0 = _mm_broadcast_ss(&igtbl_g[src[kGIndex]]);
  58         vec_b0 = _mm_broadcast_ss(&igtbl_b[src[kBIndex]]);
  59         vec_r1 = _mm_broadcast_ss(&igtbl_r[src[kRIndex + components]]);
  60         vec_g1 = _mm_broadcast_ss(&igtbl_g[src[kGIndex + components]]);
  61         vec_b1 = _mm_broadcast_ss(&igtbl_b[src[kBIndex + components]]);
  62         vec_r = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_r0), vec_r1, 1);
  63         vec_g = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_g0), vec_g1, 1);
  64         vec_b = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_b0), vec_b1, 1);
  65
  66         if (kAIndex != NO_A_INDEX) {
  67             alpha1 = src[kAIndex];
  68             alpha2 = src[kAIndex + components];
  69         }
  70     }
  71
  72     /* If there are at least 4 pixels, then we can iterate and preload the
  73        next 2 while we store the result of the current 2. */
  74     while (length > 3) {
  75         /* Ensure we are pointing at the next 2 pixels for the next load. */
  76         src += 2 * components;
  77
  78         /* gamma * matrix */
  79         vec_r = _mm256_mul_ps(vec_r, mat0);
  80         vec_g = _mm256_mul_ps(vec_g, mat1);
  81         vec_b = _mm256_mul_ps(vec_b, mat2);
  82
  83         /* store alpha for these pixels; load alpha for next two */
  84         if (kAIndex != NO_A_INDEX) {
  85             dest[kAIndex] = alpha1;
  86             dest[kAIndex + components] = alpha2;
  87             alpha1 = src[kAIndex];
  88             alpha2 = src[kAIndex + components];
  89         }
  90
  91         /* crunch, crunch, crunch */
  92         vec_r = _mm256_add_ps(vec_r, _mm256_add_ps(vec_g, vec_b));
  93         vec_r = _mm256_max_ps(min, vec_r);
  94         vec_r = _mm256_min_ps(max, vec_r);
  95         result = _mm256_mul_ps(vec_r, scale);
  96
  97         /* store calc'd output tables indices */
  98         _mm256_store_si256((__m256i*)output, _mm256_cvtps_epi32(result));
  99
 100         /* load gamma values for next loop while store completes */
 101         vec_r0 = _mm_broadcast_ss(&igtbl_r[src[kRIndex]]);
 102         vec_g0 = _mm_broadcast_ss(&igtbl_g[src[kGIndex]]);
 103         vec_b0 = _mm_broadcast_ss(&igtbl_b[src[kBIndex]]);
 104         vec_r1 = _mm_broadcast_ss(&igtbl_r[src[kRIndex + components]]);
 105         vec_g1 = _mm_broadcast_ss(&igtbl_g[src[kGIndex + components]]);
 106         vec_b1 = _mm_broadcast_ss(&igtbl_b[src[kBIndex + components]]);
 107         vec_r = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_r0), vec_r1, 1);
 108         vec_g = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_g0), vec_g1, 1);
 109         vec_b = _mm256_insertf128_ps(_mm256_castps128_ps256(vec_b0), vec_b1, 1);
 110
 111         /* use calc'd indices to output RGB values */
 112         dest[kRIndex] = otdata_r[output[0]];
 113         dest[kGIndex] = otdata_g[output[1]];
 114         dest[kBIndex] = otdata_b[output[2]];
 115         dest[kRIndex + components] = otdata_r[output[4]];
 116         dest[kGIndex + components] = otdata_g[output[5]];
 117         dest[kBIndex + components] = otdata_b[output[6]];
 118
 119         dest += 2 * components;
 120         length -= 2;
 121     }
 122
 123     /* There are 0-3 pixels remaining. If there are 2-3 remaining, then we know
 124        we have already populated the necessary registers to start the transform. */
 125     if (length > 1) {
 126         vec_r = _mm256_mul_ps(vec_r, mat0);
 127         vec_g = _mm256_mul_ps(vec_g, mat1);
 128         vec_b = _mm256_mul_ps(vec_b, mat2);
 129
 130         if (kAIndex != NO_A_INDEX) {
 131             dest[kAIndex] = alpha1;
 132             dest[kAIndex + components] = alpha2;
 133         }
 134
 135         vec_r = _mm256_add_ps(vec_r, _mm256_add_ps(vec_g, vec_b));
 136         vec_r = _mm256_max_ps(min, vec_r);
 137         vec_r = _mm256_min_ps(max, vec_r);
 138         result = _mm256_mul_ps(vec_r, scale);
 139
 140         _mm256_store_si256((__m256i*)output, _mm256_cvtps_epi32(result));
 141
 142         dest[kRIndex] = otdata_r[output[0]];
 143         dest[kGIndex] = otdata_g[output[1]];
 144         dest[kBIndex] = otdata_b[output[2]];
 145         dest[kRIndex + components] = otdata_r[output[4]];
 146         dest[kGIndex + components] = otdata_g[output[5]];
 147         dest[kBIndex + components] = otdata_b[output[6]];
 148
 149         src += 2 * components;
 150         dest += 2 * components;
 151         length -= 2;
 152     }
 153
 154     /* There may be 0-1 pixels remaining. */
 155     if (length == 1) {
 156         vec_r0 = _mm_broadcast_ss(&igtbl_r[src[kRIndex]]);
 157         vec_g0 = _mm_broadcast_ss(&igtbl_g[src[kGIndex]]);
 158         vec_b0 = _mm_broadcast_ss(&igtbl_b[src[kBIndex]]);
 159
 160         vec_r0 = _mm_mul_ps(vec_r0, _mm256_castps256_ps128(mat0));
 161         vec_g0 = _mm_mul_ps(vec_g0, _mm256_castps256_ps128(mat1));
 162         vec_b0 = _mm_mul_ps(vec_b0, _mm256_castps256_ps128(mat2));
 163
 164         if (kAIndex != NO_A_INDEX) {
 165             dest[kAIndex] = src[kAIndex];
 166         }
 167
 168         vec_r0 = _mm_add_ps(vec_r0, _mm_add_ps(vec_g0, vec_b0));
 169         vec_r0 = _mm_max_ps(_mm256_castps256_ps128(min), vec_r0);
 170         vec_r0 = _mm_min_ps(_mm256_castps256_ps128(max), vec_r0);
 171         vec_r0 = _mm_mul_ps(vec_r0, _mm256_castps256_ps128(scale));
 172
 173         _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(vec_r0));
 174
 175         dest[kRIndex] = otdata_r[output[0]];
 176         dest[kGIndex] = otdata_g[output[1]];
 177         dest[kBIndex] = otdata_b[output[2]];
 178     }
 179 }
 180
 181 void qcms_transform_data_rgb_out_lut_avx(const qcms_transform *transform,
 182                                          const unsigned char *src,
 183                                          unsigned char *dest,
 184                                          size_t length)
 185 {
 186   qcms_transform_data_template_lut_avx<RGBA_R_INDEX, RGBA_G_INDEX, RGBA_B_INDEX>(transform, src, dest, length);
 187 }
 188
 189 void qcms_transform_data_rgba_out_lut_avx(const qcms_transform *transform,
 190                                           const unsigned char *src,
 191                                           unsigned char *dest,
 192                                           size_t length)
 193 {
 194   qcms_transform_data_template_lut_avx<RGBA_R_INDEX, RGBA_G_INDEX, RGBA_B_INDEX, RGBA_A_INDEX>(transform, src, dest, length);
 195 }
 196
 197 void qcms_transform_data_bgra_out_lut_avx(const qcms_transform *transform,
 198                                           const unsigned char *src,
 199                                           unsigned char *dest,
 200                                           size_t length)
 201 {
 202   qcms_transform_data_template_lut_avx<BGRA_R_INDEX, BGRA_G_INDEX, BGRA_B_INDEX, BGRA_A_INDEX>(transform, src, dest, length);
 203 }