michael@0: /* vim: set ts=8 sw=8 noexpandtab: */
michael@0: //  qcms
michael@0: //  Copyright (C) 2009 Mozilla Corporation
michael@0: //  Copyright (C) 1998-2007 Marti Maria
michael@0: //
michael@0: // Permission is hereby granted, free of charge, to any person obtaining
michael@0: // a copy of this software and associated documentation files (the "Software"),
michael@0: // to deal in the Software without restriction, including without limitation
michael@0: // the rights to use, copy, modify, merge, publish, distribute, sublicense,
michael@0: // and/or sell copies of the Software, and to permit persons to whom the Software
michael@0: // is furnished to do so, subject to the following conditions:
michael@0: //
michael@0: // The above copyright notice and this permission notice shall be included in
michael@0: // all copies or substantial portions of the Software.
michael@0: //
michael@0: // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
michael@0: // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO
michael@0: // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
michael@0: // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
michael@0: // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
michael@0: // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
michael@0: // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
michael@0: 
michael@0: #include <altivec.h>
michael@0: 
michael@0: #include "qcmsint.h"
michael@0: 
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 = FLOATSCALE;
michael@0: static const ALIGN float clampMaxValueX4 = CLAMPMAXVAL;
michael@0: 
michael@0: inline vector float load_aligned_float(float *dataPtr)
michael@0: {
michael@0: 	vector float data = vec_lde(0, dataPtr);
michael@0: 	vector unsigned char moveToStart = vec_lvsl(0, dataPtr);
michael@0: 	return vec_perm(data, data, moveToStart);
michael@0: }
michael@0: 
michael@0: void qcms_transform_data_rgb_out_lut_altivec(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 vector float mat0 = vec_ldl(0, (vector float*)mat[0]);
michael@0: 	const vector float mat1 = vec_ldl(0, (vector float*)mat[1]);
michael@0: 	const vector float mat2 = vec_ldl(0, (vector float*)mat[2]);
michael@0: 
michael@0: 	/* these values don't change, either */
michael@0: 	const vector float max = vec_splat(vec_lde(0, (float*)&clampMaxValueX4), 0);
michael@0: 	const vector float min = (vector float)vec_splat_u32(0);
michael@0: 	const vector float scale = vec_splat(vec_lde(0, (float*)&floatScaleX4), 0);
michael@0: 
michael@0: 	/* working variables */
michael@0: 	vector float 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 = load_aligned_float((float*)&igtbl_r[src[0]]);
michael@0: 	vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
michael@0: 	vec_b = load_aligned_float((float*)&igtbl_r[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 = vec_splat(vec_r, 0);
michael@0: 		vec_g = vec_splat(vec_g, 0);
michael@0: 		vec_b = vec_splat(vec_b, 0);
michael@0: 
michael@0: 		/* gamma * matrix */
michael@0: 		vec_r = vec_madd(vec_r, mat0, min);
michael@0: 		vec_g = vec_madd(vec_g, mat1, min);
michael@0: 		vec_b = vec_madd(vec_b, mat2, min);
michael@0: 
michael@0: 		/* crunch, crunch, crunch */
michael@0: 		vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
michael@0: 		vec_r = vec_max(min, vec_r);
michael@0: 		vec_r = vec_min(max, vec_r);
michael@0: 		result = vec_madd(vec_r, scale, min);
michael@0: 
michael@0: 		/* store calc'd output tables indices */
michael@0: 		vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
michael@0: 
michael@0: 		/* load for next loop while store completes */
michael@0: 		vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
michael@0: 		vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
michael@0: 		vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
michael@0: 		src += 3;
michael@0: 
michael@0: 		/* use calc'd indices to output RGB values */
michael@0: 		dest[0] = otdata_r[output[0]];
michael@0: 		dest[1] = otdata_g[output[1]];
michael@0: 		dest[2] = otdata_b[output[2]];
michael@0: 		dest += 3;
michael@0: 	}
michael@0: 
michael@0: 	/* handle final (maybe only) pixel */
michael@0: 
michael@0: 	vec_r = vec_splat(vec_r, 0);
michael@0: 	vec_g = vec_splat(vec_g, 0);
michael@0: 	vec_b = vec_splat(vec_b, 0);
michael@0: 
michael@0: 	vec_r = vec_madd(vec_r, mat0, min);
michael@0: 	vec_g = vec_madd(vec_g, mat1, min);
michael@0: 	vec_b = vec_madd(vec_b, mat2, min);
michael@0: 
michael@0: 	vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
michael@0: 	vec_r = vec_max(min, vec_r);
michael@0: 	vec_r = vec_min(max, vec_r);
michael@0: 	result = vec_madd(vec_r, scale, min);
michael@0: 
michael@0: 	vec_st(vec_ctu(vec_round(result),0),0,(vector unsigned int*)output);
michael@0: 
michael@0: 	dest[0] = otdata_r[output[0]];
michael@0: 	dest[1] = otdata_g[output[1]];
michael@0: 	dest[2] = otdata_b[output[2]];
michael@0: }
michael@0: 
michael@0: void qcms_transform_data_rgba_out_lut_altivec(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 vector float mat0 = vec_ldl(0, (vector float*)mat[0]);
michael@0: 	const vector float mat1 = vec_ldl(0, (vector float*)mat[1]);
michael@0: 	const vector float mat2 = vec_ldl(0, (vector float*)mat[2]);
michael@0: 
michael@0: 	/* these values don't change, either */
michael@0: 	const vector float max = vec_splat(vec_lde(0, (float*)&clampMaxValueX4), 0);
michael@0: 	const vector float min = (vector float)vec_splat_u32(0);
michael@0: 	const vector float scale = vec_splat(vec_lde(0, (float*)&floatScaleX4), 0);
michael@0: 
michael@0: 	/* working variables */
michael@0: 	vector float 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 = load_aligned_float((float*)&igtbl_r[src[0]]);
michael@0: 	vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
michael@0: 	vec_b = load_aligned_float((float*)&igtbl_r[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 = vec_splat(vec_r, 0);
michael@0: 		vec_g = vec_splat(vec_g, 0);
michael@0: 		vec_b = vec_splat(vec_b, 0);
michael@0: 
michael@0: 		/* gamma * matrix */
michael@0: 		vec_r = vec_madd(vec_r, mat0, min);
michael@0: 		vec_g = vec_madd(vec_g, mat1, min);
michael@0: 		vec_b = vec_madd(vec_b, mat2, min);
michael@0: 
michael@0: 		/* store alpha for this pixel; load alpha for next */
michael@0: 		dest[3] = alpha;
michael@0: 		alpha = src[3];
michael@0: 
michael@0: 		/* crunch, crunch, crunch */
michael@0: 		vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
michael@0: 		vec_r = vec_max(min, vec_r);
michael@0: 		vec_r = vec_min(max, vec_r);
michael@0: 		result = vec_madd(vec_r, scale, min);
michael@0: 
michael@0: 		/* store calc'd output tables indices */
michael@0: 		vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
michael@0: 
michael@0: 		/* load gamma values for next loop while store completes */
michael@0: 		vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
michael@0: 		vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
michael@0: 		vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
michael@0: 		src += 4;
michael@0: 
michael@0: 		/* use calc'd indices to output RGB values */
michael@0: 		dest[0] = otdata_r[output[0]];
michael@0: 		dest[1] = otdata_g[output[1]];
michael@0: 		dest[2] = 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 = vec_splat(vec_r, 0);
michael@0: 	vec_g = vec_splat(vec_g, 0);
michael@0: 	vec_b = vec_splat(vec_b, 0);
michael@0: 
michael@0: 	vec_r = vec_madd(vec_r, mat0, min);
michael@0: 	vec_g = vec_madd(vec_g, mat1, min);
michael@0: 	vec_b = vec_madd(vec_b, mat2, min);
michael@0: 
michael@0: 	dest[3] = alpha;
michael@0: 
michael@0: 	vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
michael@0: 	vec_r = vec_max(min, vec_r);
michael@0: 	vec_r = vec_min(max, vec_r);
michael@0: 	result = vec_madd(vec_r, scale, min);
michael@0: 
michael@0: 	vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
michael@0: 
michael@0: 	dest[0] = otdata_r[output[0]];
michael@0: 	dest[1] = otdata_g[output[1]];
michael@0: 	dest[2] = otdata_b[output[2]];
michael@0: }
michael@0: