static inline vec_float4 vec_dot3( vec_float4 vec0, vec_float4 vec1 )
{
vec_float4 result;
result = spu_mul( vec0, vec1 );
result = spu_madd( spu_rlqwbyte( vec0, 4 ), spu_rlqwbyte( vec1, 4 ), result );
return spu_madd( spu_rlqwbyte( vec0, 8 ), spu_rlqwbyte( vec1, 8 ), result );
}
/**
* Setup fragment shader inputs by evaluating triangle's vertex
* attribute coefficient info.
* \param x quad x pos
* \param y quad y pos
* \param fragZ returns quad Z values
* \param fragInputs returns fragment program inputs
* Note: this code could be incorporated into the fragment program
* itself to avoid the loop and switch.
*/
static void
eval_inputs(float x, float y, vector float *fragZ, vector float fragInputs[])
{
static const vector float deltaX = (const vector float) {0, 1, 0, 1};
static const vector float deltaY = (const vector float) {0, 0, 1, 1};
const uint posSlot = 0;
const vector float pos = setup.coef[posSlot].a0;
const vector float dposdx = setup.coef[posSlot].dadx;
const vector float dposdy = setup.coef[posSlot].dady;
const vector float fragX = spu_splats(x) + deltaX;
const vector float fragY = spu_splats(y) + deltaY;
vector float fragW, wInv;
uint i;
*fragZ = splatz(pos) + fragX * splatz(dposdx) + fragY * splatz(dposdy);
fragW = splatw(pos) + fragX * splatw(dposdx) + fragY * splatw(dposdy);
wInv = spu_re(fragW); /* 1 / w */
/* loop over fragment program inputs */
for (i = 0; i < spu.vertex_info.num_attribs; i++) {
uint attr = i + 1;
enum interp_mode interp = spu.vertex_info.attrib[attr].interp_mode;
/* constant term */
vector float a0 = setup.coef[attr].a0;
vector float r0 = splatx(a0);
vector float r1 = splaty(a0);
vector float r2 = splatz(a0);
vector float r3 = splatw(a0);
if (interp == INTERP_LINEAR || interp == INTERP_PERSPECTIVE) {
/* linear term */
vector float dadx = setup.coef[attr].dadx;
vector float dady = setup.coef[attr].dady;
/* Use SPU intrinsics here to get slightly better code.
* originally: r0 += fragX * splatx(dadx) + fragY * splatx(dady);
*/
r0 = spu_madd(fragX, splatx(dadx), spu_madd(fragY, splatx(dady), r0));
r1 = spu_madd(fragX, splaty(dadx), spu_madd(fragY, splaty(dady), r1));
r2 = spu_madd(fragX, splatz(dadx), spu_madd(fragY, splatz(dady), r2));
r3 = spu_madd(fragX, splatw(dadx), spu_madd(fragY, splatw(dady), r3));
if (interp == INTERP_PERSPECTIVE) {
/* perspective term */
r0 *= wInv;
r1 *= wInv;
r2 *= wInv;
r3 *= wInv;
}
}
fragInputs[CHAN0] = r0;
fragInputs[CHAN1] = r1;
fragInputs[CHAN2] = r2;
fragInputs[CHAN3] = r3;
fragInputs += 4;
}
}
/**
* Emit a quad (pass to next stage). No clipping is done.
* Note: about 1/5 to 1/7 of the time, mask is zero and this function
* should be skipped. But adding the test for that slows things down
* overall.
*/
static INLINE void
emit_quad( int x, int y, mask_t mask)
{
/* If any bits in mask are set... */
if (spu_extract(spu_orx(mask), 0)) {
const int ix = x - setup.cliprect_minx;
const int iy = y - setup.cliprect_miny;
spu.cur_ctile_status = TILE_STATUS_DIRTY;
spu.cur_ztile_status = TILE_STATUS_DIRTY;
{
/*
* Run fragment shader, execute per-fragment ops, update fb/tile.
*/
vector float inputs[4*4], outputs[2*4];
vector unsigned int kill_mask;
vector float fragZ;
eval_inputs((float) x, (float) y, &fragZ, inputs);
ASSERT(spu.fragment_program);
ASSERT(spu.fragment_ops);
/* Execute the current fragment program */
kill_mask = spu.fragment_program(inputs, outputs, spu.constants);
mask = spu_andc(mask, kill_mask);
/* Execute per-fragment/quad operations, including:
* alpha test, z test, stencil test, blend and framebuffer writing.
* Note that there are two different fragment operations functions
* that can be called, one for front-facing fragments, and one
* for back-facing fragments. (Often the two are the same;
* but in some cases, like two-sided stenciling, they can be
* very different.) So choose the correct function depending
* on the calculated facing.
*/
spu.fragment_ops[setup.facing](ix, iy, &spu.ctile, &spu.ztile,
fragZ,
outputs[0*4+0],
outputs[0*4+1],
outputs[0*4+2],
outputs[0*4+3],
mask);
}
}
}
/**
* Given an X or Y coordinate, return the block/quad coordinate that it
* belongs to.
*/
static INLINE int
block(int x)
{
return x & ~1;
}
/**
* Render a horizontal span of quads
*/
static void
flush_spans(void)
{
int minleft, maxright;
const int l0 = spu_extract(setup.span.quad, 0);
const int l1 = spu_extract(setup.span.quad, 1);
const int r0 = spu_extract(setup.span.quad, 2);
const int r1 = spu_extract(setup.span.quad, 3);
switch (setup.span.y_flags) {
case 0x3:
/* both odd and even lines written (both quad rows) */
minleft = MIN2(l0, l1);
maxright = MAX2(r0, r1);
break;
case 0x1:
/* only even line written (quad top row) */
minleft = l0;
maxright = r0;
break;
case 0x2:
/* only odd line written (quad bottom row) */
minleft = l1;
maxright = r1;
break;
default:
return;
}
/* OK, we're very likely to need the tile data now.
* clear or finish waiting if needed.
*/
if (spu.cur_ctile_status == TILE_STATUS_GETTING) {
/* wait for mfc_get() to complete */
//printf("SPU: %u: waiting for ctile\n", spu.init.id);
wait_on_mask(1 << TAG_READ_TILE_COLOR);
spu.cur_ctile_status = TILE_STATUS_CLEAN;
}
else if (spu.cur_ctile_status == TILE_STATUS_CLEAR) {
//printf("SPU %u: clearing C tile %u, %u\n", spu.init.id, setup.tx, setup.ty);
clear_c_tile(&spu.ctile);
spu.cur_ctile_status = TILE_STATUS_DIRTY;
}
ASSERT(spu.cur_ctile_status != TILE_STATUS_DEFINED);
if (spu.read_depth_stencil) {
if (spu.cur_ztile_status == TILE_STATUS_GETTING) {
/* wait for mfc_get() to complete */
//printf("SPU: %u: waiting for ztile\n", spu.init.id);
wait_on_mask(1 << TAG_READ_TILE_Z);
spu.cur_ztile_status = TILE_STATUS_CLEAN;
}
else if (spu.cur_ztile_status == TILE_STATUS_CLEAR) {
//printf("SPU %u: clearing Z tile %u, %u\n", spu.init.id, setup.tx, setup.ty);
clear_z_tile(&spu.ztile);
spu.cur_ztile_status = TILE_STATUS_DIRTY;
}
ASSERT(spu.cur_ztile_status != TILE_STATUS_DEFINED);
}
/* XXX this loop could be moved into the above switch cases... */
/* Setup for mask calculation */
const vec_int4 quad_LlRr = setup.span.quad;
const vec_int4 quad_RrLl = spu_rlqwbyte(quad_LlRr, 8);
const vec_int4 quad_LLll = spu_shuffle(quad_LlRr, quad_LlRr, SHUFFLE4(A,A,B,B));
const vec_int4 quad_RRrr = spu_shuffle(quad_RrLl, quad_RrLl, SHUFFLE4(A,A,B,B));
const vec_int4 twos = spu_splats(2);
const int x = block(minleft);
vec_int4 xs = {x, x+1, x, x+1};
for (; spu_extract(xs, 0) <= block(maxright); xs += twos) {
/**
* Computes mask to indicate which pixels in the 2x2 quad are actually
* inside the triangle's bounds.
*/
/* Calculate ({x,x+1,x,x+1} >= {l[0],l[0],l[1],l[1]}) */
const mask_t gt_LLll_xs = spu_cmpgt(quad_LLll, xs);
const mask_t gte_xs_LLll = spu_nand(gt_LLll_xs, gt_LLll_xs);
/* Calculate ({r[0],r[0],r[1],r[1]} > {x,x+1,x,x+1}) */
const mask_t gt_RRrr_xs = spu_cmpgt(quad_RRrr, xs);
/* Combine results to create mask */
const mask_t mask = spu_and(gte_xs_LLll, gt_RRrr_xs);
emit_quad(spu_extract(xs, 0), setup.span.y, mask);
}
setup.span.y = 0;
setup.span.y_flags = 0;
/* Zero right elements */
setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0));
}
#if DEBUG_VERTS
static void
print_vertex(const struct vertex_header *v)
{
uint i;
fprintf(stderr, " Vertex: (%p)\n", v);
for (i = 0; i < spu.vertex_info.num_attribs; i++) {
fprintf(stderr, " %d: %f %f %f %f\n", i,
spu_extract(v->data[i], 0),
spu_extract(v->data[i], 1),
spu_extract(v->data[i], 2),
spu_extract(v->data[i], 3));
}
}
void* libvector_pointwise_multiply_32fc_unaligned(void* target, void* src0, void* src1, unsigned int num_bytes){
//loop iterator i
int i = 0;
void* retval = target;
//put the target and source addresses into qwords
vector unsigned int address_counter_tgt = {(unsigned int)target, 0, 0, 0};
vector unsigned int address_counter_src0 = {(unsigned int)src0, 0, 0 ,0};
vector unsigned int address_counter_src1 = {(unsigned int)src1, 0, 0, 0};
//create shuffle masks
//shuffle mask building blocks:
//all from the first vector
vector unsigned char oneup = {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f};
//all from the second vector
vector unsigned char second_oneup = {0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f};
//gamma: second half of the second, first half of the first, break at (unsigned int)src0%16
vector unsigned char src_cmp = spu_splats((unsigned char)((unsigned int)src0%16));
vector unsigned char gt_res = spu_cmpgt(oneup, src_cmp);
vector unsigned char eq_res = spu_cmpeq(oneup, src_cmp);
vector unsigned char cmp_res = spu_or(gt_res, eq_res);
vector unsigned char sixteen_uchar = spu_splats((unsigned char)16);
vector unsigned char phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_gamma = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
shuffle_mask_gamma = spu_rlqwbyte(shuffle_mask_gamma, (unsigned int)src0%16);
//eta: second half of the second, first half of the first, break at (unsigned int)src1%16
src_cmp = spu_splats((unsigned char)((unsigned int)src1%16));
gt_res = spu_cmpgt(oneup, src_cmp);
eq_res = spu_cmpeq(oneup, src_cmp);
cmp_res = spu_or(gt_res, eq_res);
sixteen_uchar = spu_splats((unsigned char)16);
phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_eta = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
shuffle_mask_eta = spu_rlqwbyte(shuffle_mask_eta, (unsigned int)src1%16);
vector unsigned char tgt_second = spu_rlqwbyte(second_oneup, -((unsigned int)target%16));
vector unsigned char tgt_first = spu_rlqwbyte(oneup, -((unsigned int)target%16));
//alpha: first half of first, second half of second, break at (unsigned int)target%16
src_cmp = spu_splats((unsigned char)((unsigned int)target%16));
gt_res = spu_cmpgt(oneup, src_cmp);
eq_res = spu_cmpeq(oneup, src_cmp);
cmp_res = spu_or(gt_res, eq_res);
phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_alpha = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
//delta: first half of first, first half of second, break at (unsigned int)target%16
vector unsigned char shuffle_mask_delta = spu_shuffle(oneup, tgt_second, (vector unsigned char)shuffle_mask_alpha);
//epsilon: second half of second, second half of first, break at (unsigned int)target%16
vector unsigned char shuffle_mask_epsilon = spu_shuffle(tgt_second, oneup, (vector unsigned char)shuffle_mask_alpha);
//zeta: second half of second, first half of first, break at 16 - (unsigned int)target%16
vector unsigned int shuffle_mask_zeta = spu_rlqwbyte(shuffle_mask_alpha, (unsigned int)target%16);
//beta: first half of first, second half of second, break at num_bytes%16
src_cmp = spu_splats((unsigned char)(num_bytes%16));
gt_res = spu_cmpgt(oneup, src_cmp);
eq_res = spu_cmpeq(oneup, src_cmp);
cmp_res = spu_or(gt_res, eq_res);
phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_beta = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
qword src0_past;
qword src0_present;
qword src1_past;
qword src1_present;
qword tgt_past;
qword tgt_present;
qword in_temp0;
qword in_temp1;
qword out_temp0;
qword out_temp1;
src0_past = si_lqd((qword)address_counter_src0, 0);
src1_past = si_lqd((qword)address_counter_src1, 0);
tgt_past = si_lqd((qword)address_counter_tgt, 0);
vector unsigned char shuffle_mask_complexprod0 = {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
0x0c, 0x0d, 0x0e, 0x0f, 0x08, 0x09, 0x0a, 0x0b};
vector unsigned char shuffle_mask_complexprod1 = {0x00, 0x01, 0x02, 0x03, 0x10, 0x11, 0x12, 0x13,
0x08, 0x09, 0x0a, 0x0b, 0x18, 0x19, 0x1a, 0x1b};
vector unsigned char shuffle_mask_complexprod2 = {0x04, 0x05, 0x06, 0x07, 0x14, 0x15, 0x16, 0x17,
0x0c, 0x0d, 0x0e, 0x0f, 0x1c, 0x1d, 0x1e, 0x1f};
vector unsigned char sign_changer = {0x00, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00};
vector float prod0;
qword shuf0;
vector float prod1;
vector float sign_change;
qword summand0;
qword summand1;
vector float sum;
for(i = 0; i < num_bytes/16; ++i) {
src0_present = si_lqd((qword)address_counter_src0, 16);
src1_present = si_lqd((qword)address_counter_src1, 16);
tgt_present = si_lqd((qword)address_counter_tgt, 16);
in_temp0 = spu_shuffle(src0_present, src0_past, (vector unsigned char)shuffle_mask_gamma);
in_temp1 = spu_shuffle(src1_present, src1_past, (vector unsigned char)shuffle_mask_eta);
prod0 = spu_mul((vector float)in_temp0, (vector float)in_temp1);
shuf0 = spu_shuffle((qword)in_temp1, (qword)in_temp1, shuffle_mask_complexprod0);
prod1 = spu_mul((vector float)in_temp0, (vector float)shuf0);
sign_change = spu_xor(prod0, (vector float)sign_changer);
summand0 = spu_shuffle((qword)sign_change, (qword)prod1, shuffle_mask_complexprod1);
summand1 = spu_shuffle((qword)sign_change, (qword)prod1, shuffle_mask_complexprod2);
sum = spu_add((vector float)summand0, (vector float)summand1);
out_temp0 = spu_shuffle(tgt_past, (qword)sum, shuffle_mask_delta);
out_temp1 = spu_shuffle(tgt_present, (qword)sum, shuffle_mask_epsilon);
si_stqd(out_temp0, (qword)address_counter_tgt, 0);
si_stqd(out_temp1, (qword)address_counter_tgt, 16);
tgt_past = out_temp1;
src0_past = src0_present;
src1_past = src1_present;
address_counter_src0 = spu_add(address_counter_src0, 16);
address_counter_src1 = spu_add(address_counter_src1, 16);
address_counter_tgt = spu_add(address_counter_tgt, 16);
}
src0_present = si_lqd((qword)address_counter_src0, 16);
src1_present = si_lqd((qword)address_counter_src1, 16);
tgt_present = si_lqd((qword)address_counter_tgt, 16);
in_temp0 = spu_shuffle(src0_present, src0_past, (vector unsigned char) shuffle_mask_gamma);
in_temp1 = spu_shuffle(src1_present, src1_past, (vector unsigned char) shuffle_mask_eta);
prod0 = spu_mul((vector float)in_temp0, (vector float)in_temp1);
shuf0 = spu_shuffle((qword)in_temp1, (qword)in_temp1, shuffle_mask_complexprod0);
prod1 = spu_mul(prod0, (vector float)shuf0);
sign_change = spu_xor(prod0, (vector float)sign_changer);
summand0 = spu_shuffle((qword)sign_change, (qword)prod1, shuffle_mask_complexprod1);
summand1 = spu_shuffle((qword)sign_change, (qword)prod1, shuffle_mask_complexprod2);
sum = spu_add((vector float)summand0, (vector float)summand1);
qword target_temp = spu_shuffle(tgt_present, tgt_past, (vector unsigned char) shuffle_mask_zeta);
qword meld = spu_shuffle((qword)sum, target_temp, (vector unsigned char)shuffle_mask_beta);
out_temp0 = spu_shuffle(tgt_past, meld, shuffle_mask_delta);
out_temp1 = spu_shuffle(tgt_present, meld, shuffle_mask_epsilon);
si_stqd(out_temp0, (qword)address_counter_tgt, 0);
si_stqd(out_temp1, (qword)address_counter_tgt, 16);
return retval;
}
