inline void MinMaxBinCount3SIMD(aabb_t *aabb, minmaxbin_t *mmb, aabb_t *baabb)
{
vector float *baabb_min = (vector float*)baabb->min;
vector float *baabb_max = (vector float*)baabb->max;
vector float *aabb_min = (vector float*)aabb->min;
vector float *aabb_max = (vector float*)aabb->max;
vector float nbins = spu_splats((float)nsamplepoints);
vector float invnbins = spu_re(nbins);
nbins = spu_sub(nbins, spu_splats(1.0f));
vector float width = spu_abs(spu_sub(*baabb_max, *baabb_min));
vector float invdelta = spu_re(spu_mul(width, invnbins));
vector int minindex = GetBinSIMD(*baabb_min, *baabb_max, *aabb_min, invdelta, nbins);
vector int maxindex = GetBinSIMD(*baabb_min, *baabb_max, *aabb_max, invdelta, nbins);
mmb->minbins[minindex[0]].b[0]++;
mmb->minbins[minindex[1]].b[1]++;
mmb->minbins[minindex[2]].b[2]++;
mmb->maxbins[maxindex[0]].b[0]++;
mmb->maxbins[maxindex[1]].b[1]++;
mmb->maxbins[maxindex[2]].b[2]++;
}
/**
* 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 draw_frame(uint64_t buf_ea) {
vec_uint4 buf[2*1920/4];
int row, col, i, tag = 0;
float step = 4.0f/spu.width*spu.zoom;
float xbeg = spu.xc - spu.width*step*0.5f;
vec_float4 vxbeg = spu_splats(xbeg)
+ spu_splats(step) * (vec_float4) {
0.f,1.f,2.f,3.f
};
vec_float4 xstep = spu_splats(step)*spu_splats(4.f);
vec_float4 vyp = spu_splats(spu.yc - spu.height*step*0.5f + step*spu.rank);
const vec_float4 vinc = spu_splats(spu.count * step);
const vec_float4 esc2 = spu_splats(BAILOUT*BAILOUT);
#if BAILBITS != 1
const vec_float4 esc21 = spu_splats(4.f/(BAILOUT*BAILOUT));
#endif
const vec_float4 two = spu_splats(2.f);
const vec_float4 zero = spu_splats(0.f);
const vec_float4 colsc = spu_splats(255.f);
const vec_float4 ccr = spu_splats(4.f*BAILOUT/(3.5f*3.141592654f));
const vec_float4 ccg = spu_splats(4.f*BAILOUT/(5.f*3.141592654f));
const vec_float4 ccb = spu_splats(4.f*BAILOUT/(9.f*3.141592654f));
vec_float4 x, y, x2, y2, m2, vxp;
vec_uint4 cmp, inc;
vec_uint4 vi;
vec_uint4 *p, *b;
vec_float4 co;
/* Process the full image. As there are 6 SPUs working in parallel, each with
* a different rank from 0 to 5, each SPU processes only the line numbers:
* rank, rank+6, rank+12, ...
* The program uses a SPU DMA programming technique known as "double buffering",
* where the previously generated line is transmitted to main memory while we
* compute the next one, hence the need for a local buffer containing two lines.
*/
for (row = spu.rank; row < spu.height; row += spu.count) {
/* Pixel buffer address (in local memory) of the next line to be drawn */
b = p = buf + ((1920/4)&-tag);
vxp = vxbeg; /* first four x coordinates */
/* Process a whole screen line by packets of 4 pixels */
for (col = spu.width/4; col > 0 ; col--) {
vi = spu_splats(0u);
x = vxp;
y = vyp;
i = 0;
cmp = spu_splats(-1u);
inc = spu_splats(1u);
m2 = zero;
/* This loop processes the Mandelbrot suite for the four complex numbers
* whose real part are the components of the x vector, and the imaginary
* part are in y (as we process the same line, all initial values of y
* are equal).
* We perform loop unrolling for SPU performance optimization reasons,
* hence the 4x replication of the same computation block.
*/
do {
x2 = x*x;
y2 = y*y;
m2 = spu_sel(m2, x2+y2, cmp);
cmp = spu_cmpgt(esc2, m2);
inc = spu_and(inc, cmp); /* increment the iteration count only if */
vi = vi + inc; /* we're still inside the bailout radius */
y = two*x*y + vyp;
x = x2-y2 + vxp;
x2 = x*x;
y2 = y*y;
m2 = spu_sel(m2, x2+y2, cmp);
cmp = spu_cmpgt(esc2, m2);
inc = spu_and(inc, cmp);
vi = vi + inc;
y = two*x*y + vyp;
x = x2-y2 + vxp;
x2 = x*x;
y2 = y*y;
m2 = spu_sel(m2, x2+y2, cmp);
cmp = spu_cmpgt(esc2, m2);
inc = spu_and(inc, cmp);
vi = vi + inc;
y = two*x*y + vyp;
x = x2-y2 + vxp;
x2 = x*x;
y2 = y*y;
m2 = spu_sel(m2, x2+y2, cmp);
cmp = spu_cmpgt(esc2, m2);
inc = spu_and(inc, cmp);
vi = vi + inc;
y = two*x*y + vyp;
x = x2-y2 + vxp;
i += 4;
}
/* Exit the loop only if the iteration limit of 128 has been reached,
* or all current four points are outside the bailout radius.
* The __builtin_expect(xxx, 1) construct hints the compiler that the xxx
* test has greater chance of being true (1), so a branch hinting
* instruction is inserted into the binary code to make the conditional
* branch faster in most cases (except the last one when we exit the
* loop). This results in performance increase.
*/
while (__builtin_expect((i < 128) &
(si_to_int((qword)spu_gather(cmp)) != 0), 1));
/* smooth coloring: compute the fractional part */
co = spu_convtf(vi, 0) + spu_splats(1.f);
co -= fast_logf(fast_logf(m2) * spu_splats(.5f));
#if BAILBITS != 1
co = spu_re(spu_rsqrte(co*esc21));
#endif
/* Compute the red, green an blue pixel components */
vec_uint4 cr = spu_convtu(mcos(co * ccr) * colsc, 0);
vec_uint4 cg = spu_convtu(mcos(co * ccg) * colsc, 0);
vec_uint4 cb = spu_convtu(mcos(co * ccb) * colsc, 0);
/* Put the 4 pixel values in the buffer */
*p++ = (spu_sl(cr, 16) | spu_sl(cg, 8) | cb) & ~-inc;
vxp += xstep;
}
/* double-buffered dma: initiate a dma transfer of last computed scanline
* then wait for completion of the second last transfer (previous computed
* line). This is done by changing the tag value.
*/
mfc_put(b, buf_ea+(spu.width*4)*row, spu.width*4, tag, 0, 0);
tag = 1 - tag;
wait_for_completion(tag);
vyp += vinc;
}
/* wait for completion of last sent image line */
wait_for_completion(1-tag);
}
vector double
__divv2df3 (vector double a_in, vector double b_in)
{
/* Variables */
vec_int4 exp, exp_bias;
vec_uint4 no_underflow, overflow;
vec_float4 mant_bf, inv_bf;
vec_ullong2 exp_a, exp_b;
vec_ullong2 a_nan, a_zero, a_inf, a_denorm, a_denorm0;
vec_ullong2 b_nan, b_zero, b_inf, b_denorm, b_denorm0;
vec_ullong2 nan;
vec_uint4 a_exp, b_exp;
vec_ullong2 a_mant_0, b_mant_0;
vec_ullong2 a_exp_1s, b_exp_1s;
vec_ullong2 sign_exp_mask;
vec_double2 a, b;
vec_double2 mant_a, mant_b, inv_b, q0, q1, q2, mult;
/* Constants */
vec_uint4 exp_mask_u32 = spu_splats((unsigned int)0x7FF00000);
vec_uchar16 splat_hi = (vec_uchar16) {
0,1,2,3, 0,1,2,3, 8, 9,10,11, 8,9,10,11
};
vec_uchar16 swap_32 = (vec_uchar16) {
4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11
};
vec_ullong2 exp_mask = spu_splats(0x7FF0000000000000ULL);
vec_ullong2 sign_mask = spu_splats(0x8000000000000000ULL);
vec_float4 onef = spu_splats(1.0f);
vec_double2 one = spu_splats(1.0);
vec_double2 exp_53 = (vec_double2)spu_splats(0x0350000000000000ULL);
sign_exp_mask = spu_or(sign_mask, exp_mask);
/* Extract the floating point components from each of the operands including
* exponent and mantissa.
*/
a_exp = (vec_uint4)spu_and((vec_uint4)a_in, exp_mask_u32);
a_exp = spu_shuffle(a_exp, a_exp, splat_hi);
b_exp = (vec_uint4)spu_and((vec_uint4)b_in, exp_mask_u32);
b_exp = spu_shuffle(b_exp, b_exp, splat_hi);
a_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)a_in, sign_exp_mask), 0);
a_mant_0 = spu_and(a_mant_0, spu_shuffle(a_mant_0, a_mant_0, swap_32));
b_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)b_in, sign_exp_mask), 0);
b_mant_0 = spu_and(b_mant_0, spu_shuffle(b_mant_0, b_mant_0, swap_32));
a_exp_1s = (vec_ullong2)spu_cmpeq(a_exp, exp_mask_u32);
b_exp_1s = (vec_ullong2)spu_cmpeq(b_exp, exp_mask_u32);
/* Identify all possible special values that must be accommodated including:
* +-denorm, +-0, +-infinity, and NaNs.
*/
a_denorm0= (vec_ullong2)spu_cmpeq(a_exp, 0);
a_nan = spu_andc(a_exp_1s, a_mant_0);
a_zero = spu_and (a_denorm0, a_mant_0);
a_inf = spu_and (a_exp_1s, a_mant_0);
a_denorm = spu_andc(a_denorm0, a_zero);
b_denorm0= (vec_ullong2)spu_cmpeq(b_exp, 0);
b_nan = spu_andc(b_exp_1s, b_mant_0);
b_zero = spu_and (b_denorm0, b_mant_0);
b_inf = spu_and (b_exp_1s, b_mant_0);
b_denorm = spu_andc(b_denorm0, b_zero);
/* Scale denorm inputs to into normalized numbers by conditionally scaling the
* input parameters.
*/
a = spu_sub(spu_or(a_in, exp_53), spu_sel(exp_53, a_in, sign_mask));
a = spu_sel(a_in, a, a_denorm);
b = spu_sub(spu_or(b_in, exp_53), spu_sel(exp_53, b_in, sign_mask));
b = spu_sel(b_in, b, b_denorm);
/* Extract the divisor and dividend exponent and force parameters into the signed
* range [1.0,2.0) or [-1.0,2.0).
*/
exp_a = spu_and((vec_ullong2)a, exp_mask);
exp_b = spu_and((vec_ullong2)b, exp_mask);
mant_a = spu_sel(a, one, (vec_ullong2)exp_mask);
mant_b = spu_sel(b, one, (vec_ullong2)exp_mask);
/* Approximate the single reciprocal of b by using
* the single precision reciprocal estimate followed by one
* single precision iteration of Newton-Raphson.
*/
mant_bf = spu_roundtf(mant_b);
inv_bf = spu_re(mant_bf);
inv_bf = spu_madd(spu_nmsub(mant_bf, inv_bf, onef), inv_bf, inv_bf);
/* Perform 2 more Newton-Raphson iterations in double precision. The
* result (q1) is in the range (0.5, 2.0).
*/
inv_b = spu_extend(inv_bf);
inv_b = spu_madd(spu_nmsub(mant_b, inv_b, one), inv_b, inv_b);
q0 = spu_mul(mant_a, inv_b);
q1 = spu_madd(spu_nmsub(mant_b, q0, mant_a), inv_b, q0);
/* Determine the exponent correction factor that must be applied
* to q1 by taking into account the exponent of the normalized inputs
* and the scale factors that were applied to normalize them.
*/
exp = spu_rlmaska(spu_sub((vec_int4)exp_a, (vec_int4)exp_b), -20);
exp = spu_add(exp, (vec_int4)spu_add(spu_and((vec_int4)a_denorm, -0x34), spu_and((vec_int4)b_denorm, 0x34)));
/* Bias the quotient exponent depending on the sign of the exponent correction
* factor so that a single multiplier will ensure the entire double precision
* domain (including denorms) can be achieved.
*
* exp bias q1 adjust exp
* ===== ======== ==========
* positive 2^+65 -65
* negative 2^-64 +64
*/
exp_bias = spu_xor(spu_rlmaska(exp, -31), 64);
exp = spu_sub(exp, exp_bias);
q1 = spu_sel(q1, (vec_double2)spu_add((vec_int4)q1, spu_sl(exp_bias, 20)), exp_mask);
/* Compute a multiplier (mult) to applied to the quotient (q1) to produce the
* expected result. On overflow, clamp the multiplier to the maximum non-infinite
* number in case the rounding mode is not round-to-nearest.
*/
exp = spu_add(exp, 0x3FF);
no_underflow = spu_cmpgt(exp, 0);
overflow = spu_cmpgt(exp, 0x7FE);
exp = spu_and(spu_sl(exp, 20), (vec_int4)no_underflow);
exp = spu_and(exp, (vec_int4)exp_mask);
mult = spu_sel((vec_double2)exp, (vec_double2)(spu_add((vec_uint4)exp_mask, -1)), (vec_ullong2)overflow);
/* Handle special value conditions. These include:
*
* 1) IF either operand is a NaN OR both operands are 0 or INFINITY THEN a NaN
* results.
* 2) ELSE IF the dividend is an INFINITY OR the divisor is 0 THEN a INFINITY results.
* 3) ELSE IF the dividend is 0 OR the divisor is INFINITY THEN a 0 results.
*/
mult = spu_andc(mult, (vec_double2)spu_or(a_zero, b_inf));
mult = spu_sel(mult, (vec_double2)exp_mask, spu_or(a_inf, b_zero));
nan = spu_or(a_nan, b_nan);
nan = spu_or(nan, spu_and(a_zero, b_zero));
nan = spu_or(nan, spu_and(a_inf, b_inf));
mult = spu_or(mult, (vec_double2)nan);
/* Scale the final quotient */
q2 = spu_mul(q1, mult);
return (q2);
}
