1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/qcms/transform-sse2.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,243 @@
1.4 +#include <emmintrin.h>
1.5 +
1.6 +#include "qcmsint.h"
1.7 +
1.8 +/* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
1.9 +#define FLOATSCALE (float)(PRECACHE_OUTPUT_SIZE)
1.10 +#define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE )
1.11 +static const ALIGN float floatScaleX4[4] =
1.12 + { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
1.13 +static const ALIGN float clampMaxValueX4[4] =
1.14 + { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
1.15 +
1.16 +void qcms_transform_data_rgb_out_lut_sse2(qcms_transform *transform,
1.17 + unsigned char *src,
1.18 + unsigned char *dest,
1.19 + size_t length)
1.20 +{
1.21 + unsigned int i;
1.22 + float (*mat)[4] = transform->matrix;
1.23 + char input_back[32];
1.24 + /* Ensure we have a buffer that's 16 byte aligned regardless of the original
1.25 + * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
1.26 + * because they don't work on stack variables. gcc 4.4 does do the right thing
1.27 + * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
1.28 + float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
1.29 + /* share input and output locations to save having to keep the
1.30 + * locations in separate registers */
1.31 + uint32_t const * output = (uint32_t*)input;
1.32 +
1.33 + /* deref *transform now to avoid it in loop */
1.34 + const float *igtbl_r = transform->input_gamma_table_r;
1.35 + const float *igtbl_g = transform->input_gamma_table_g;
1.36 + const float *igtbl_b = transform->input_gamma_table_b;
1.37 +
1.38 + /* deref *transform now to avoid it in loop */
1.39 + const uint8_t *otdata_r = &transform->output_table_r->data[0];
1.40 + const uint8_t *otdata_g = &transform->output_table_g->data[0];
1.41 + const uint8_t *otdata_b = &transform->output_table_b->data[0];
1.42 +
1.43 + /* input matrix values never change */
1.44 + const __m128 mat0 = _mm_load_ps(mat[0]);
1.45 + const __m128 mat1 = _mm_load_ps(mat[1]);
1.46 + const __m128 mat2 = _mm_load_ps(mat[2]);
1.47 +
1.48 + /* these values don't change, either */
1.49 + const __m128 max = _mm_load_ps(clampMaxValueX4);
1.50 + const __m128 min = _mm_setzero_ps();
1.51 + const __m128 scale = _mm_load_ps(floatScaleX4);
1.52 +
1.53 + /* working variables */
1.54 + __m128 vec_r, vec_g, vec_b, result;
1.55 +
1.56 + /* CYA */
1.57 + if (!length)
1.58 + return;
1.59 +
1.60 + /* one pixel is handled outside of the loop */
1.61 + length--;
1.62 +
1.63 + /* setup for transforming 1st pixel */
1.64 + vec_r = _mm_load_ss(&igtbl_r[src[0]]);
1.65 + vec_g = _mm_load_ss(&igtbl_g[src[1]]);
1.66 + vec_b = _mm_load_ss(&igtbl_b[src[2]]);
1.67 + src += 3;
1.68 +
1.69 + /* transform all but final pixel */
1.70 +
1.71 + for (i=0; i<length; i++)
1.72 + {
1.73 + /* position values from gamma tables */
1.74 + vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
1.75 + vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
1.76 + vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
1.77 +
1.78 + /* gamma * matrix */
1.79 + vec_r = _mm_mul_ps(vec_r, mat0);
1.80 + vec_g = _mm_mul_ps(vec_g, mat1);
1.81 + vec_b = _mm_mul_ps(vec_b, mat2);
1.82 +
1.83 + /* crunch, crunch, crunch */
1.84 + vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
1.85 + vec_r = _mm_max_ps(min, vec_r);
1.86 + vec_r = _mm_min_ps(max, vec_r);
1.87 + result = _mm_mul_ps(vec_r, scale);
1.88 +
1.89 + /* store calc'd output tables indices */
1.90 + _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
1.91 +
1.92 + /* load for next loop while store completes */
1.93 + vec_r = _mm_load_ss(&igtbl_r[src[0]]);
1.94 + vec_g = _mm_load_ss(&igtbl_g[src[1]]);
1.95 + vec_b = _mm_load_ss(&igtbl_b[src[2]]);
1.96 + src += 3;
1.97 +
1.98 + /* use calc'd indices to output RGB values */
1.99 + dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
1.100 + dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
1.101 + dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
1.102 + dest += RGB_OUTPUT_COMPONENTS;
1.103 + }
1.104 +
1.105 + /* handle final (maybe only) pixel */
1.106 +
1.107 + vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
1.108 + vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
1.109 + vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
1.110 +
1.111 + vec_r = _mm_mul_ps(vec_r, mat0);
1.112 + vec_g = _mm_mul_ps(vec_g, mat1);
1.113 + vec_b = _mm_mul_ps(vec_b, mat2);
1.114 +
1.115 + vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
1.116 + vec_r = _mm_max_ps(min, vec_r);
1.117 + vec_r = _mm_min_ps(max, vec_r);
1.118 + result = _mm_mul_ps(vec_r, scale);
1.119 +
1.120 + _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
1.121 +
1.122 + dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
1.123 + dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
1.124 + dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
1.125 +}
1.126 +
1.127 +void qcms_transform_data_rgba_out_lut_sse2(qcms_transform *transform,
1.128 + unsigned char *src,
1.129 + unsigned char *dest,
1.130 + size_t length)
1.131 +{
1.132 + unsigned int i;
1.133 + float (*mat)[4] = transform->matrix;
1.134 + char input_back[32];
1.135 + /* Ensure we have a buffer that's 16 byte aligned regardless of the original
1.136 + * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
1.137 + * because they don't work on stack variables. gcc 4.4 does do the right thing
1.138 + * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
1.139 + float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
1.140 + /* share input and output locations to save having to keep the
1.141 + * locations in separate registers */
1.142 + uint32_t const * output = (uint32_t*)input;
1.143 +
1.144 + /* deref *transform now to avoid it in loop */
1.145 + const float *igtbl_r = transform->input_gamma_table_r;
1.146 + const float *igtbl_g = transform->input_gamma_table_g;
1.147 + const float *igtbl_b = transform->input_gamma_table_b;
1.148 +
1.149 + /* deref *transform now to avoid it in loop */
1.150 + const uint8_t *otdata_r = &transform->output_table_r->data[0];
1.151 + const uint8_t *otdata_g = &transform->output_table_g->data[0];
1.152 + const uint8_t *otdata_b = &transform->output_table_b->data[0];
1.153 +
1.154 + /* input matrix values never change */
1.155 + const __m128 mat0 = _mm_load_ps(mat[0]);
1.156 + const __m128 mat1 = _mm_load_ps(mat[1]);
1.157 + const __m128 mat2 = _mm_load_ps(mat[2]);
1.158 +
1.159 + /* these values don't change, either */
1.160 + const __m128 max = _mm_load_ps(clampMaxValueX4);
1.161 + const __m128 min = _mm_setzero_ps();
1.162 + const __m128 scale = _mm_load_ps(floatScaleX4);
1.163 +
1.164 + /* working variables */
1.165 + __m128 vec_r, vec_g, vec_b, result;
1.166 + unsigned char alpha;
1.167 +
1.168 + /* CYA */
1.169 + if (!length)
1.170 + return;
1.171 +
1.172 + /* one pixel is handled outside of the loop */
1.173 + length--;
1.174 +
1.175 + /* setup for transforming 1st pixel */
1.176 + vec_r = _mm_load_ss(&igtbl_r[src[0]]);
1.177 + vec_g = _mm_load_ss(&igtbl_g[src[1]]);
1.178 + vec_b = _mm_load_ss(&igtbl_b[src[2]]);
1.179 + alpha = src[3];
1.180 + src += 4;
1.181 +
1.182 + /* transform all but final pixel */
1.183 +
1.184 + for (i=0; i<length; i++)
1.185 + {
1.186 + /* position values from gamma tables */
1.187 + vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
1.188 + vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
1.189 + vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
1.190 +
1.191 + /* gamma * matrix */
1.192 + vec_r = _mm_mul_ps(vec_r, mat0);
1.193 + vec_g = _mm_mul_ps(vec_g, mat1);
1.194 + vec_b = _mm_mul_ps(vec_b, mat2);
1.195 +
1.196 + /* store alpha for this pixel; load alpha for next */
1.197 + dest[OUTPUT_A_INDEX] = alpha;
1.198 + alpha = src[3];
1.199 +
1.200 + /* crunch, crunch, crunch */
1.201 + vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
1.202 + vec_r = _mm_max_ps(min, vec_r);
1.203 + vec_r = _mm_min_ps(max, vec_r);
1.204 + result = _mm_mul_ps(vec_r, scale);
1.205 +
1.206 + /* store calc'd output tables indices */
1.207 + _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
1.208 +
1.209 + /* load gamma values for next loop while store completes */
1.210 + vec_r = _mm_load_ss(&igtbl_r[src[0]]);
1.211 + vec_g = _mm_load_ss(&igtbl_g[src[1]]);
1.212 + vec_b = _mm_load_ss(&igtbl_b[src[2]]);
1.213 + src += 4;
1.214 +
1.215 + /* use calc'd indices to output RGB values */
1.216 + dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
1.217 + dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
1.218 + dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
1.219 + dest += RGBA_OUTPUT_COMPONENTS;
1.220 + }
1.221 +
1.222 + /* handle final (maybe only) pixel */
1.223 +
1.224 + vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
1.225 + vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
1.226 + vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
1.227 +
1.228 + vec_r = _mm_mul_ps(vec_r, mat0);
1.229 + vec_g = _mm_mul_ps(vec_g, mat1);
1.230 + vec_b = _mm_mul_ps(vec_b, mat2);
1.231 +
1.232 + dest[OUTPUT_A_INDEX] = alpha;
1.233 +
1.234 + vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
1.235 + vec_r = _mm_max_ps(min, vec_r);
1.236 + vec_r = _mm_min_ps(max, vec_r);
1.237 + result = _mm_mul_ps(vec_r, scale);
1.238 +
1.239 + _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
1.240 +
1.241 + dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
1.242 + dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
1.243 + dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
1.244 +}
1.245 +
1.246 +
