michael@0: #include <xmmintrin.h>
michael@0: 
michael@0: #include "qcmsint.h"
michael@0: 
michael@0: /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
michael@0: #define FLOATSCALE  (float)(PRECACHE_OUTPUT_SIZE)
michael@0: #define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE )
michael@0: static const ALIGN float floatScaleX4[4] =
michael@0:     { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
michael@0: static const ALIGN float clampMaxValueX4[4] =
michael@0:     { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
michael@0: 
michael@0: void qcms_transform_data_rgb_out_lut_sse1(qcms_transform *transform,
michael@0:                                           unsigned char *src,
michael@0:                                           unsigned char *dest,
michael@0:                                           size_t length)
michael@0: {
michael@0:     unsigned int i;
michael@0:     float (*mat)[4] = transform->matrix;
michael@0:     char input_back[32];
michael@0:     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
michael@0:      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
michael@0:      * because they don't work on stack variables. gcc 4.4 does do the right thing
michael@0:      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
michael@0:     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
michael@0:     /* share input and output locations to save having to keep the
michael@0:      * locations in separate registers */
michael@0:     uint32_t const * output = (uint32_t*)input;
michael@0: 
michael@0:     /* deref *transform now to avoid it in loop */
michael@0:     const float *igtbl_r = transform->input_gamma_table_r;
michael@0:     const float *igtbl_g = transform->input_gamma_table_g;
michael@0:     const float *igtbl_b = transform->input_gamma_table_b;
michael@0: 
michael@0:     /* deref *transform now to avoid it in loop */
michael@0:     const uint8_t *otdata_r = &transform->output_table_r->data[0];
michael@0:     const uint8_t *otdata_g = &transform->output_table_g->data[0];
michael@0:     const uint8_t *otdata_b = &transform->output_table_b->data[0];
michael@0: 
michael@0:     /* input matrix values never change */
michael@0:     const __m128 mat0  = _mm_load_ps(mat[0]);
michael@0:     const __m128 mat1  = _mm_load_ps(mat[1]);
michael@0:     const __m128 mat2  = _mm_load_ps(mat[2]);
michael@0: 
michael@0:     /* these values don't change, either */
michael@0:     const __m128 max   = _mm_load_ps(clampMaxValueX4);
michael@0:     const __m128 min   = _mm_setzero_ps();
michael@0:     const __m128 scale = _mm_load_ps(floatScaleX4);
michael@0: 
michael@0:     /* working variables */
michael@0:     __m128 vec_r, vec_g, vec_b, result;
michael@0: 
michael@0:     /* CYA */
michael@0:     if (!length)
michael@0:         return;
michael@0: 
michael@0:     /* one pixel is handled outside of the loop */
michael@0:     length--;
michael@0: 
michael@0:     /* setup for transforming 1st pixel */
michael@0:     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
michael@0:     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
michael@0:     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
michael@0:     src += 3;
michael@0: 
michael@0:     /* transform all but final pixel */
michael@0: 
michael@0:     for (i=0; i<length; i++)
michael@0:     {
michael@0:         /* position values from gamma tables */
michael@0:         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
michael@0:         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
michael@0:         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
michael@0: 
michael@0:         /* gamma * matrix */
michael@0:         vec_r = _mm_mul_ps(vec_r, mat0);
michael@0:         vec_g = _mm_mul_ps(vec_g, mat1);
michael@0:         vec_b = _mm_mul_ps(vec_b, mat2);
michael@0: 
michael@0:         /* crunch, crunch, crunch */
michael@0:         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
michael@0:         vec_r  = _mm_max_ps(min, vec_r);
michael@0:         vec_r  = _mm_min_ps(max, vec_r);
michael@0:         result = _mm_mul_ps(vec_r, scale);
michael@0: 
michael@0:         /* store calc'd output tables indices */
michael@0:         *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
michael@0:         result = _mm_movehl_ps(result, result);
michael@0:         *((__m64 *)&output[2]) = _mm_cvtps_pi32(result) ;
michael@0: 
michael@0:         /* load for next loop while store completes */
michael@0:         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
michael@0:         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
michael@0:         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
michael@0:         src += 3;
michael@0: 
michael@0:         /* use calc'd indices to output RGB values */
michael@0:         dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
michael@0:         dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
michael@0:         dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
michael@0:         dest += RGB_OUTPUT_COMPONENTS;
michael@0:     }
michael@0: 
michael@0:     /* handle final (maybe only) pixel */
michael@0: 
michael@0:     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
michael@0:     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
michael@0:     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
michael@0: 
michael@0:     vec_r = _mm_mul_ps(vec_r, mat0);
michael@0:     vec_g = _mm_mul_ps(vec_g, mat1);
michael@0:     vec_b = _mm_mul_ps(vec_b, mat2);
michael@0: 
michael@0:     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
michael@0:     vec_r  = _mm_max_ps(min, vec_r);
michael@0:     vec_r  = _mm_min_ps(max, vec_r);
michael@0:     result = _mm_mul_ps(vec_r, scale);
michael@0: 
michael@0:     *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
michael@0:     result = _mm_movehl_ps(result, result);
michael@0:     *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
michael@0: 
michael@0:     dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
michael@0:     dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
michael@0:     dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
michael@0: 
michael@0:     _mm_empty();
michael@0: }
michael@0: 
michael@0: void qcms_transform_data_rgba_out_lut_sse1(qcms_transform *transform,
michael@0:                                            unsigned char *src,
michael@0:                                            unsigned char *dest,
michael@0:                                            size_t length)
michael@0: {
michael@0:     unsigned int i;
michael@0:     float (*mat)[4] = transform->matrix;
michael@0:     char input_back[32];
michael@0:     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
michael@0:      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
michael@0:      * because they don't work on stack variables. gcc 4.4 does do the right thing
michael@0:      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
michael@0:     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
michael@0:     /* share input and output locations to save having to keep the
michael@0:      * locations in separate registers */
michael@0:     uint32_t const * output = (uint32_t*)input;
michael@0: 
michael@0:     /* deref *transform now to avoid it in loop */
michael@0:     const float *igtbl_r = transform->input_gamma_table_r;
michael@0:     const float *igtbl_g = transform->input_gamma_table_g;
michael@0:     const float *igtbl_b = transform->input_gamma_table_b;
michael@0: 
michael@0:     /* deref *transform now to avoid it in loop */
michael@0:     const uint8_t *otdata_r = &transform->output_table_r->data[0];
michael@0:     const uint8_t *otdata_g = &transform->output_table_g->data[0];
michael@0:     const uint8_t *otdata_b = &transform->output_table_b->data[0];
michael@0: 
michael@0:     /* input matrix values never change */
michael@0:     const __m128 mat0  = _mm_load_ps(mat[0]);
michael@0:     const __m128 mat1  = _mm_load_ps(mat[1]);
michael@0:     const __m128 mat2  = _mm_load_ps(mat[2]);
michael@0: 
michael@0:     /* these values don't change, either */
michael@0:     const __m128 max   = _mm_load_ps(clampMaxValueX4);
michael@0:     const __m128 min   = _mm_setzero_ps();
michael@0:     const __m128 scale = _mm_load_ps(floatScaleX4);
michael@0: 
michael@0:     /* working variables */
michael@0:     __m128 vec_r, vec_g, vec_b, result;
michael@0:     unsigned char alpha;
michael@0: 
michael@0:     /* CYA */
michael@0:     if (!length)
michael@0:         return;
michael@0: 
michael@0:     /* one pixel is handled outside of the loop */
michael@0:     length--;
michael@0: 
michael@0:     /* setup for transforming 1st pixel */
michael@0:     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
michael@0:     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
michael@0:     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
michael@0:     alpha = src[3];
michael@0:     src += 4;
michael@0: 
michael@0:     /* transform all but final pixel */
michael@0: 
michael@0:     for (i=0; i<length; i++)
michael@0:     {
michael@0:         /* position values from gamma tables */
michael@0:         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
michael@0:         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
michael@0:         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
michael@0: 
michael@0:         /* gamma * matrix */
michael@0:         vec_r = _mm_mul_ps(vec_r, mat0);
michael@0:         vec_g = _mm_mul_ps(vec_g, mat1);
michael@0:         vec_b = _mm_mul_ps(vec_b, mat2);
michael@0: 
michael@0:         /* store alpha for this pixel; load alpha for next */
michael@0:         dest[OUTPUT_A_INDEX] = alpha;
michael@0:         alpha   = src[3];
michael@0: 
michael@0:         /* crunch, crunch, crunch */
michael@0:         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
michael@0:         vec_r  = _mm_max_ps(min, vec_r);
michael@0:         vec_r  = _mm_min_ps(max, vec_r);
michael@0:         result = _mm_mul_ps(vec_r, scale);
michael@0: 
michael@0:         /* store calc'd output tables indices */
michael@0:         *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
michael@0:         result = _mm_movehl_ps(result, result);
michael@0:         *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
michael@0: 
michael@0:         /* load gamma values for next loop while store completes */
michael@0:         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
michael@0:         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
michael@0:         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
michael@0:         src += 4;
michael@0: 
michael@0:         /* use calc'd indices to output RGB values */
michael@0:         dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
michael@0:         dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
michael@0:         dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
michael@0:         dest += 4;
michael@0:     }
michael@0: 
michael@0:     /* handle final (maybe only) pixel */
michael@0: 
michael@0:     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
michael@0:     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
michael@0:     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
michael@0: 
michael@0:     vec_r = _mm_mul_ps(vec_r, mat0);
michael@0:     vec_g = _mm_mul_ps(vec_g, mat1);
michael@0:     vec_b = _mm_mul_ps(vec_b, mat2);
michael@0: 
michael@0:     dest[OUTPUT_A_INDEX] = alpha;
michael@0: 
michael@0:     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
michael@0:     vec_r  = _mm_max_ps(min, vec_r);
michael@0:     vec_r  = _mm_min_ps(max, vec_r);
michael@0:     result = _mm_mul_ps(vec_r, scale);
michael@0: 
michael@0:     *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
michael@0:     result = _mm_movehl_ps(result, result);
michael@0:     *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
michael@0: 
michael@0:     dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
michael@0:     dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
michael@0:     dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
michael@0: 
michael@0:     _mm_empty();
michael@0: }