vec_ullong2 bitDiff_d2(vec_double2 ref, vec_double2 vals) {
double ref0, ref1, vals0, vals1;
long long refi0, refi1, valsi0, valsi1, diff0, diff1;
vec_ullong2 bits;
ref0 = spu_extract(ref,0);
ref1 = spu_extract(ref,1);
vals0 = spu_extract(vals,0);
vals1 = spu_extract(vals,1);
refi0 = make_ulonglong(ref0);
refi1 = make_ulonglong(ref1);
valsi0 = make_ulonglong(vals0);
valsi1 = make_ulonglong(vals1);
diff0 = refi0 - valsi0;
diff1 = refi1 - valsi1;
if ( diff0 < 0 )
{
diff0 = valsi0 - refi0;
}
if ( diff1 < 0 )
{
diff1 = valsi1 - refi1;
}
bits = spu_promote( (unsigned long long)ceil(log2((double)diff0)), 0 );
bits = spu_insert( (unsigned long long)ceil(log2((double)diff1)), bits, 1 );
return bits;
}
vec_ullong2 ulpDiff_d2(vec_double2 ref, vec_double2 vals) {
double ref0, ref1, vals0, vals1;
long long refi0, refi1, valsi0, valsi1, diff0, diff1;
vec_ullong2 ulps;
ref0 = spu_extract(ref,0);
ref1 = spu_extract(ref,1);
vals0 = spu_extract(vals,0);
vals1 = spu_extract(vals,1);
refi0 = make_ulonglong(ref0);
refi1 = make_ulonglong(ref1);
valsi0 = make_ulonglong(vals0);
valsi1 = make_ulonglong(vals1);
diff0 = refi0 - valsi0;
diff1 = refi1 - valsi1;
if ( diff0 < 0 )
{
diff0 = valsi0 - refi0;
}
if ( diff1 < 0 )
{
diff1 = valsi1 - refi1;
}
ulps = spu_promote( (unsigned long long)diff0, 0 );
ulps = spu_insert( (unsigned long long)diff1, ulps, 1 );
return ulps;
}
int cacheGetPrime(int n)
{
if ((n < primeCacheStart + primeCacheSize) && (n > primeCacheStart))
{
int r = spu_extract(primeCacheData[(n - primeCacheStart) / 4], n%4);
return r;
}
// Haal op.
uint32_t tag, size;
tag = mfc_tag_reserve();
size = CACHE_PRIME_SIZE*16;
unsigned long long EA = setup.vPrimes + (n - n%4) * 4;
mfc_get(&primeCacheData, EA, size, tag, 0, 0);
mfc_write_tag_mask(1 << tag);
mfc_read_tag_status_all();
mfc_tag_release(tag);
primeCacheStart = n - (n % 4);
int r = spu_extract(primeCacheData[(n - primeCacheStart) / 4], n%4);
return r;
}
int main() {
int i, j;
/* The input and output arrays */
vector float uniform_vec[4];
vector float normal_vec[4];
/* Generate a seed */
struct timeval time;
gettimeofday(&time, NULL);
/* Generate the random numbers */
mc_rand_mt_init(time.tv_sec);
mc_rand_mt_0_to_1_array_f4(4, uniform_vec);
/* Transform the array */
mc_transform_po_array_f4(4, uniform_vec, normal_vec,
&mc_rand_mt_0_to_1_f4);
/* Display the results */
printf("Uniform Distribution: \n");
for(i=0; i<4; i++)
for(j=0; j<4; j++)
printf("%f ",spu_extract(uniform_vec[i], j));
printf("\n\nNormal Distribution: \n");
for(i=0; i<4; i++)
for(j=0; j<4; j++)
printf("%f ",spu_extract(normal_vec[i], j));
printf("\n");
return 0;
}
/* Calculates the length of the string s, not including the terminating
* \0 character.
*/
size_t strlen(const char *s)
{
size_t len;
unsigned int cnt, cmp, skip, mask;
vec_uchar16 *ptr, data;
/* Compensate for initial mis-aligned string.
*/
ptr = (vec_uchar16 *)s;
skip = (unsigned int)(ptr) & 15;
mask = 0xFFFF >> skip;
data = *ptr++;
cmp = spu_extract(spu_gather(spu_cmpeq(data, 0)), 0);
cmp &= mask;
cnt = spu_extract(spu_cntlz(spu_promote(cmp, 0)), 0);
len = cnt - (skip + 16);
while (cnt == 32) {
data = *ptr++;
len -= 16;
cnt = spu_extract(spu_cntlz(spu_gather(spu_cmpeq(data, 0))), 0);
len += cnt;
}
return (len);
}
int allequal_bits_float4( vec_float4 x, vec_float4 y, int tolerance )
{
vec_uint4 bits = bitDiff_f4( x, y );
return ( (int)spu_extract(bits,0) <= tolerance &&
(int)spu_extract(bits,1) <= tolerance &&
(int)spu_extract(bits,2) <= tolerance &&
(int)spu_extract(bits,3) <= tolerance );
}
int allequal_ulps_float4( vec_float4 x, vec_float4 y, int tolerance )
{
vec_uint4 ulps = ulpDiff_f4( x, y );
return ( (int)spu_extract(ulps,0) <= tolerance &&
(int)spu_extract(ulps,1) <= tolerance &&
(int)spu_extract(ulps,2) <= tolerance &&
(int)spu_extract(ulps,3) <= tolerance );
}
inline void merge_cache_blocks(RenderableCacheLine* cache)
{
vec_uchar16 next = cache->chunkNext;
for (;;) {
vec_uchar16 nextnext = spu_shuffle(next, next, next);
vec_uchar16 nextmask = spu_and(next, spu_splats((unsigned char)CHUNKNEXT_MASK));
vec_ushort8 firstblock0 = spu_cmpeq( cache->chunkStart[0], 0);
vec_ushort8 firstblock1 = spu_cmpeq( cache->chunkStart[1], 0);
// change next to word offset, note we don't care what the low bit shifted in is
vec_uchar16 firstshuf = (vec_uchar16) spu_sl( (vec_ushort8)nextmask, 1 );
vec_uchar16 first = (vec_uchar16) spu_shuffle( firstblock0, firstblock1, firstshuf );
vec_ushort8 tri0 = cache->chunkTriangle[0];
vec_ushort8 tri1 = cache->chunkTriangle[1];
vec_uchar16 trishufhi = spu_or ( firstshuf, spu_splats((unsigned char) 1));
vec_uchar16 trishuflo = spu_and( firstshuf, spu_splats((unsigned char) 254));
vec_ushort8 ntri0 = spu_shuffle( tri0, tri1, spu_shuffle( trishuflo, trishufhi, SHUF0 ) );
vec_ushort8 ntri1 = spu_shuffle( tri0, tri1, spu_shuffle( trishuflo, trishufhi, SHUF1 ) );
vec_ushort8 trieq0 = spu_cmpeq( tri0, ntri0 );
vec_ushort8 trieq1 = spu_cmpeq( tri1, ntri1 );
vec_uchar16 trieq = (vec_uchar16) spu_shuffle( trieq0, trieq1, MERGE );
vec_uchar16 combi = spu_orc(first, trieq);
vec_uchar16 canmerge = spu_cmpgt( spu_nor(spu_or(next, nextnext), combi), 256-CHUNKNEXT_BUSY_BIT );
vec_uint4 gather = spu_gather( canmerge );
vec_uint4 mergeid = spu_sub( spu_cntlz( gather ), spu_promote((unsigned int)16, 0));
if( !spu_extract(gather, 0) ) {
return;
}
// unsigned int firstchunk = spu_extract(mergeid, 0);
// unsigned int nextchunk = cache->chunkNextArray[firstchunk];
vec_uint4 v_chunkNext = (vec_uint4) si_rotqby( (qword) next, (qword) spu_add(mergeid,13) );
vec_uint4 v_chunkNextNext = (vec_uint4) si_rotqby( (qword) next, (qword) spu_add(v_chunkNext,13) );
// cache->chunkNextArray[firstchunk] = cache->chunkNextArray[nextchunk];
next = spu_shuffle( (vec_uchar16) v_chunkNextNext, next, (vec_uchar16) si_cbd( (qword) mergeid, 0 ) );
// cache->chunkNextArray[nextchunk] = CHUNKNEXT_FREE_BLOCK;
next = spu_shuffle( spu_splats( (unsigned char) CHUNKNEXT_FREE_BLOCK), next, (vec_uchar16) si_cbd( (qword) v_chunkNext, 0 ) );
// this is for debug use only, it's not really needed...
// cache->chunkStartArray[nextchunk] = -1;
cache->chunkStartArray[ spu_extract(v_chunkNext,0) & 255 ] = -1;
cache->chunkNext = next;
}
}
void GenerateFrustum()
{
// g_R2OCon.m_frustum_planes has 6 planes in the following order:
//
// 0: near (there is igg code that assumes this is first)
// 1: left
// 2: right
// 3: bottom
// 4: top
// 5: far (there is igg code that assumes this is last)
// the plane normals point into the interior of the frustum
// copy T,B,L,R planes directly
for (u32 i=1; i<5; i++)
{
g_Planes[i] = g_pViewData->m_frustum_planes[i];
// loosen frustum based on lod, amplitude, and a fudgefactor
f32 nw = spu_extract(g_Planes[i], 3);
f32 ny = spu_extract(g_Planes[i], 1);
//f32 slop = g_R2OCon.m_frustum_fudge_factor * g_WaterObject.m_amplitude * powf(0.5f, 0.5f*(f32)lod)
// * (g_R2OCon.m_step * 0.0625f); // readjust based on changed in lod 0 from 16m to 128m stepsize
//f32 slop = g_R2OCon.m_frustum_fudge_factor * powf(0.5f, 0.5f*(f32)lod);
f32 slop = g_R2OCon.m_frustum_fudge_factor * step;
nw += slop * fabsf(ny);
g_Planes[i] = spu_insert(nw, g_Planes[i], 3);
}
// compute near subfrustum plane based on lod
//f32 d = g_R2OCon.m_near0 * powf(0.5f, 0.5f*(f32)lod);
f32 d = g_R2OCon.m_near0 * step * (1.0f/128.0f);
// test
if (lod==g_R2OCon.m_num_lods-1)
{
d = step;
}
vf32 camera_direction = (vf32){0,0,1,0};
camera_direction = spu_insert(g_pViewData->m_world_to_camera_matrix.m_v0.z, camera_direction, 0);
camera_direction = spu_insert(g_pViewData->m_world_to_camera_matrix.m_v1.z, camera_direction, 1);
camera_direction = spu_insert(g_pViewData->m_world_to_camera_matrix.m_v2.z, camera_direction, 2);
f32 dot = spu_extract(spu_dot3(camera_direction, g_pViewData->m_camera_position), 0);
f32 w = -(dot + d);
g_Planes[0] = spu_insert(w, camera_direction, 3);
// reverse the sense of the near plane
g_Planes[0] = -g_Planes[0];
}
int allequal_llroundf4( llroundf4_t x, llroundf4_t y )
{
return ( spu_extract(x.vll[0],0) == spu_extract(y.vll[0],0) &&
spu_extract(x.vll[0],1) == spu_extract(y.vll[0],1) &&
spu_extract(x.vll[1],0) == spu_extract(y.vll[1],0) &&
spu_extract(x.vll[1],1) == spu_extract(y.vll[1],1) );
}
int EulerLagrangeLifchitzPrimalityTest(vec_uint4 N[3], int fSophieGermain)
{
const vec_uint4 one[3] = { { 0,0,0,0 }, { 0,0,0,0 }, { 0,0,0,1 } };
vec_uint4 r[3];
vec_uint4 N_1[3];
vec_uint4 N_1_2[3];
_mpm_sub(N_1, N, one, 3);
MPM_SHR_BITS_LARGE(N_1_2, N_1, 3, zero, 1);
_mpm_mod_exp_2(r, N_1_2, 3, N, 3, 6);
int nMod8 = spu_extract(N[2], 3) & 7;
if (fSophieGermain && (nMod8 == 7))
{
if (_mpm_cmpeq(r, one, 3)) return 1;
}
else if (fSophieGermain && (nMod8 == 3))
{
if (_mpm_cmpeq(r, N_1, 3)) return 1;
}
else if (!fSophieGermain && (nMod8 == 5))
{
if (_mpm_cmpeq(r, N_1, 3)) return 1;
}
else if (!fSophieGermain && (nMod8 == 1))
{
if (_mpm_cmpeq(r, one, 3)) return 1;
}
return 0;
}
void writeTriangleBuffer(Triangle* endTriangle)
{
if (endTriangle != _currentTriangle) {
int length = ( ((char*)endTriangle) - _currentTriangleBuffer + 127) & ~127;
unsigned short endTriangleBase = (((char*)endTriangle) - ((char*)_currentTriangle)) + _currentTriangleOffset;
vec_ushort8 v_new_end = spu_promote(endTriangleBase, 1);
// calculate genuine next pointer ( rewind==0 -> next, rewind!=0 -> 0 )
unsigned short next_pointer = spu_extract( spu_andc( v_new_end, _currentTriangleRewind ), 1 );
_currentTriangle->next_triangle = next_pointer;
// printf("current=0x%x, endTriBase=0x%x, next_pointer=0x%x\n", _currentTriangleOffset, endTriangleBase, next_pointer);
// DMA the triangle data out
spu_mfcdma64(_currentTriangleBuffer, mfc_ea2h(_currentTriangleBufferEA), mfc_ea2l(_currentTriangleBufferEA), length, 0, MFC_PUT_CMD);
// update the information in the cache line
_currentTriangleRewind = spu_splats(next_pointer); // re-use this variable as we don't need it anymore
char* dstart = ((char*)&_currentTriangleRewind) + (_currentTriangleCacheEndTriangleEAL & 15);
spu_mfcdma64(dstart, _currentTriangleCacheEndTriangleEAH, _currentTriangleCacheEndTriangleEAL, sizeof(short), 0, MFC_PUTB_CMD);
// printf("writing from %x to %x:%x\n", dstart, _currentTriangleCacheEndTriangleEAH, _currentTriangleCacheEndTriangleEAL);
// finally invalidate the triangle info
_currentTriangle = NULL;
// and make sure the DMA completed
mfc_write_tag_mask(1<<0);
mfc_read_tag_status_all();
}
}
unsigned int
__mfc_tag_release (unsigned int tag)
{
vector unsigned int is_invalid;
vector unsigned int mask = (vector unsigned int)
{ 0x80000000, 0x80000000, 0x80000000, 0x80000000 };
vector signed int zero = (vector signed int) { 0, 0, 0, 0 };
vector signed int has_been_reserved;
/* Check if the tag is out of range. */
is_invalid = spu_cmpgt (spu_promote (tag, 0), 31);
/* Check whether the tag has been reserved, set to all 1 if has not
been reserved, 0 otherwise. */
has_been_reserved = (vector signed int) spu_rl (__mfc_tag_table, tag);
has_been_reserved = (vector signed int) spu_cmpgt (zero, has_been_reserved);
/* Set invalid. */
is_invalid = spu_or ((vector unsigned int) has_been_reserved, is_invalid);
mask = spu_rlmask (mask, (int)(-tag));
__mfc_tag_table = spu_or (__mfc_tag_table, mask);
return spu_extract(is_invalid, 0);
}
int num_in_buffer(int side){
volatile vector signed int *head_idx, *tail_idx;
int buffer_size;
if(side == OUT && mcb[am].local[OUT] < 255){
int parent_idx = mcb[am].local[OUT];
int side = (mcb[am].id+1)&1;
head_idx = &md[parent_idx].idx[side][HEAD];
tail_idx = &md[parent_idx].idx[side][TAIL];
buffer_size = mcb[parent_idx].buffer_size[side];
} else {
head_idx = &md[am].idx[side][HEAD];
tail_idx = &md[am].idx[side][TAIL];
buffer_size = mcb[am].buffer_size[side];
}
vector signed int diff = spu_sub(*head_idx,*tail_idx);
int num = spu_extract(diff,0);
if(num < 0)
num = num + buffer_size;
return num;
}
unsigned int
__mfc_multi_tag_reserve (unsigned int number_of_tags)
{
vector unsigned int table_copy;
vector unsigned int one = (vector unsigned int)
{ 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF };
vector unsigned int count_busy, is_valid;
vector unsigned int count_total;
vector unsigned int count_avail = (vector unsigned int) { 0, 0, 0, 0 };
vector unsigned int index = (vector unsigned int) { 0, 0, 0, 0 };
table_copy = __mfc_tag_table;
/* count_busy: number of consecutive busy tags
count_avail: number of consecutive free tags
table_copy: temporary copy of the tag table
count_total: sum of count_busy and count_avail
index: index of the current working tag */
do
{
table_copy = spu_sl (table_copy, count_avail);
count_busy = spu_cntlz (table_copy);
table_copy = spu_sl (table_copy, count_busy);
count_avail = spu_cntlz (spu_xor(table_copy, -1));
count_total = spu_add (count_busy, count_avail);
index = spu_add (index, count_total);
}
while (spu_extract (count_avail, 0) < number_of_tags
&& spu_extract (table_copy, 0) != 0);
index = spu_sub (index, count_avail);
/* is_valid is set to 0xFFFFFFFF if table_copy == 0, 0 otherwise. */
is_valid = spu_cmpeq (table_copy, 0);
index = spu_sel (index, is_valid, is_valid);
/* Now I need to actually mark the tags as used. */
table_copy = spu_sl (one, number_of_tags);
table_copy = spu_rl (table_copy, -number_of_tags - spu_extract (index, 0));
table_copy = spu_sel (table_copy, __mfc_tag_table, table_copy);
__mfc_tag_table = spu_sel (table_copy, __mfc_tag_table, is_valid);
return spu_extract (index, 0);
}
static int
kernel_zfft_f(lwp_functions* pf,
void* params,
void* inout,
unsigned int iter,
unsigned int iter_max)
{
static fft1d_f* obj;
Fft_split_params* fftp = (Fft_split_params *)params;
unsigned int size = fftp->size;
unsigned int chunks = fftp->chunks_per_wb;
unsigned int i;
int dir = fftp->direction == fwd_fft ? CML_FFT_FWD : CML_FFT_INV;
if (iter == iter_max-1 && iter_max * chunks > fftp->chunks_per_spe)
chunks = fftp->chunks_per_spe % chunks;
assert(size >= MIN_FFT_1D_SIZE);
assert(size <= MAX_FFT_1D_SIZE);
if (iter == 0 && size != current_size)
{
#if !NDEBUG
// Check that buffer space doesn't overlap with stack
register volatile vector unsigned int get_r1 asm("1");
unsigned int stack_pointer = spu_extract(get_r1, 0);
assert(buf + 2*MAX_FFT_1D_SIZE + size + 128/4 < stack_pointer);
#endif
int rt = cml_fft1d_setup_f(&obj, CML_FFT_CC, size,
buf + 2*MAX_FFT_1D_SIZE);
assert(rt && obj != NULL);
current_size = size;
}
float* inout_re = (float*)inout + 0 * size;
float* inout_im = (float*)inout + 1*chunks * size;
for (i=0; i<chunks; ++i)
{
cml_zzfft1d_op_f(obj,
(float*)inout_re + i*size, (float*)inout_im + i*size,
(float*)inout_re + i*size, (float*)inout_im + i*size,
dir, buf);
}
if (fftp->scale != (double)1.f)
{
// Instead of regular split svmul:
// cml_core_rzsvmul1_f(fftp->scale, out_re,out_im,out_re,out_im,size);
// Take advantage of real and imag being contiguous:
cml_core_svmul1_f(fftp->scale, inout_re, inout_re, 2*size*chunks);
}
return 0;
}
void *
sbrk (ptrdiff_t increment)
{
static caddr_t heap_ptr = NULL;
caddr_t base;
vector unsigned int sp_reg, sp_delta;
vector unsigned int *sp_ptr;
caddr_t sps;
/* The stack pointer register. */
volatile register vector unsigned int sp_r1 __asm__("1");
if (heap_ptr == NULL)
heap_ptr = (caddr_t) & _end;
sps = (caddr_t) spu_extract (sp_r1, 0);
if (((int) sps - STACKSIZE - (int) heap_ptr) >= increment)
{
base = heap_ptr;
heap_ptr += increment;
sp_delta = (vector unsigned int) spu_insert (increment, spu_splats (0), 1);
/* Subtract sp_delta from the SP limit (word 1). */
sp_r1 = spu_sub (sp_r1, sp_delta);
/* Fix-up backchain. */
sp_ptr = (vector unsigned int *) spu_extract (sp_r1, 0);
do
{
sp_reg = *sp_ptr;
*sp_ptr = (vector unsigned int) spu_sub (sp_reg, sp_delta);
}
while ((sp_ptr = (vector unsigned int *) spu_extract (sp_reg, 0)));
return (base);
}
else
{
errno = ENOMEM;
return ((void *) -1);
}
}
static btVector3 convexHullSupport (const btVector3& localDirOrg, const btVector3* points, int numPoints, const btVector3& localScaling)
{
btVector3 vec = localDirOrg * localScaling;
#if defined (__CELLOS_LV2__) && defined (__SPU__)
btVector3 localDir = vec;
vec_float4 v_distMax = {-FLT_MAX,0,0,0};
vec_int4 v_idxMax = {-999,0,0,0};
int v=0;
int numverts = numPoints;
for(;v<(int)numverts-4;v+=4) {
vec_float4 p0 = vec_dot3(points[v ].get128(),localDir.get128());
vec_float4 p1 = vec_dot3(points[v+1].get128(),localDir.get128());
vec_float4 p2 = vec_dot3(points[v+2].get128(),localDir.get128());
vec_float4 p3 = vec_dot3(points[v+3].get128(),localDir.get128());
const vec_int4 i0 = {v ,0,0,0};
const vec_int4 i1 = {v+1,0,0,0};
const vec_int4 i2 = {v+2,0,0,0};
const vec_int4 i3 = {v+3,0,0,0};
vec_uint4 retGt01 = spu_cmpgt(p0,p1);
vec_float4 pmax01 = spu_sel(p1,p0,retGt01);
vec_int4 imax01 = spu_sel(i1,i0,retGt01);
vec_uint4 retGt23 = spu_cmpgt(p2,p3);
vec_float4 pmax23 = spu_sel(p3,p2,retGt23);
vec_int4 imax23 = spu_sel(i3,i2,retGt23);
vec_uint4 retGt0123 = spu_cmpgt(pmax01,pmax23);
vec_float4 pmax0123 = spu_sel(pmax23,pmax01,retGt0123);
vec_int4 imax0123 = spu_sel(imax23,imax01,retGt0123);
vec_uint4 retGtMax = spu_cmpgt(v_distMax,pmax0123);
v_distMax = spu_sel(pmax0123,v_distMax,retGtMax);
v_idxMax = spu_sel(imax0123,v_idxMax,retGtMax);
}
for(;v<(int)numverts;v++) {
vec_float4 p = vec_dot3(points[v].get128(),localDir.get128());
const vec_int4 i = {v,0,0,0};
vec_uint4 retGtMax = spu_cmpgt(v_distMax,p);
v_distMax = spu_sel(p,v_distMax,retGtMax);
v_idxMax = spu_sel(i,v_idxMax,retGtMax);
}
int ptIndex = spu_extract(v_idxMax,0);
const btVector3& supVec= points[ptIndex] * localScaling;
return supVec;
#else
btScalar maxDot;
long ptIndex = vec.maxDot( points, numPoints, maxDot);
btAssert(ptIndex >= 0);
btVector3 supVec = points[ptIndex] * localScaling;
return supVec;
#endif //__SPU__
}
int main(int argc, char **argv) {
int i, j;
vector unsigned int int_vec;
vector float float_vec;
vector double double_vec[4];
/* Get the current time to use as seed */
struct timeval time;
gettimeofday(&time, NULL);
/* The hardware number generator */
printf("\nHardware Generator:\n");
if(mc_rand_hw_init() == 0) {
int_vec = mc_rand_hw_u4();
for(i=0; i<4; i++)
printf("%u ",spu_extract(int_vec, i));
printf("\n\n");
}
else
printf("Hardware RNG is not available.\n\n");
/* The Kirkpatrick Stoll PRNG */
mc_rand_ks_init(time.tv_sec);
float_vec = mc_rand_ks_0_to_1_f4();
printf("Kirkpatrick-Stoll:\n");
for(i=0; i<4; i++)
printf("%f ", spu_extract(float_vec, i));
printf("\n\n");
/* The Mersenne Twister PRNG */
mc_rand_mt_init(time.tv_sec);
mc_rand_mt_minus1_to_1_array_d2(4, double_vec);
printf("Mersenne Twister:\n");
for(i=0; i<4; i++) {
for(j=0; j<2; j++)
printf("%g ",spu_extract(double_vec[i], j));
printf("\n");
}
return 0;
}
int main(int argc, char **argv) {
int i;
vector unsigned int all_ones = (vector unsigned int)
{0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF};
vector unsigned int all_zeroes = (vector unsigned int)
{0x00000000, 0x00000000, 0x00000000, 0x00000000};
/* These bits will form the selection mask */
unsigned short mask = 0x9;
/* Each bit in 0x9 forms a word in the mask */
vector unsigned int resultw =
spu_sel(all_zeroes, all_ones, spu_maskw(mask));
printf("resultw: ");
for (i=0; i<4; i++) {
printf("%08x", spu_extract(resultw, i));
}
/* Each bit in 0x09 forms a halfword in the mask */
vector unsigned short resulth =
spu_sel((vector unsigned short)all_zeroes,
(vector unsigned short)all_ones,
spu_maskh(mask));
printf("\nresulth: ");
for (i=0; i<8; i++) {
printf("%04x", spu_extract(resulth, i));
}
/* Each bit in 0x0009 forms a byte in the mask */
vector unsigned char resultb =
spu_sel((vector unsigned char)all_zeroes,
(vector unsigned char)all_ones,
spu_maskb(mask));
printf("\nresultb: ");
for (i=0; i<16; i++) {
printf("%02x", spu_extract(resultb, i));
}
printf("\n");
return 0;
}
void cp_buffer(int side){
int avail_out = num_free_in_buffer(OUT);
int avail_side = num_in_buffer(side);
int max = avail_out < avail_side ? avail_out : avail_side;
vector signed int *out_head;
if(mcb[am].local[OUT] < 255)
out_head = (vector signed int*) &md[ mcb[am].local[OUT] ].idx[ (mcb[am].id+1)&1 ][HEAD];
else
out_head = (vector signed int*) &md[am].idx[OUT][HEAD];
vector unsigned int cmp_v;
vector signed int from_size = spu_splats( mcb[am].buffer_size[side] );
vector signed int out_size = spu_splats( mcb[ mcb[am].local[OUT] ].buffer_size[ (mcb[am].id+1)&1 ] );
vector signed int ones = {1,1,1,1};
vector signed int zeros = {0,0,0,0};
int i;
for(i = 0; i < max; i++){
md[am].buffer[OUT][spu_extract( *out_head,0)] = md[am].buffer[side][spu_extract(md[am].idx[side][TAIL],0)];
// update idx
md[am].idx[side][TAIL] = spu_add(md[am].idx[side][TAIL], ones);
cmp_v = spu_cmpeq(md[am].idx[side][TAIL],from_size);
md[am].idx[side][TAIL] = spu_sel(md[am].idx[side][TAIL], zeros, cmp_v);
*out_head = spu_add(*out_head,ones);
cmp_v = spu_cmpeq(*out_head, out_size);
*out_head = spu_sel(*out_head,zeros,cmp_v);
}
update_tail(side);
md[am].consumed[side] += max;
if(mcb[am].local[OUT] < 255 && md[am].consumed[side] == mcb[am].data_size[side]){
md[am].depleted[side] = 1;
md[am].done = 1;
--num_active_mergers;
}
}
void R2O_CutHoles(u8 *outcodes, i32 width, i32 height, u32 lod)
{
// get window extents
f32 x_min = spu_extract(origin_world, 0);
f32 z_min = spu_extract(origin_world, 2);
f32 x_max = x_min + (f32)width * step;
f32 z_max = z_min + (f32)height * step;
// make sure we cut the right water object
vu32 coords_u32 = *(vu32 *)&g_WaterObject.m_origin;
u32 tag = spu_extract(coords_u32, 0) ^ spu_extract(coords_u32, 2);
R2OHole *p_hole = g_Holes;
for (u32 i=0; i<g_R2OCon.m_num_holes; i++, p_hole++)
{
if (tag==p_hole->m_tag && lod==p_hole->m_lod)
{
// set start of hole
f32 x = p_hole->m_xcoord;
f32 z = p_hole->m_zcoord;
// set coords deltas
f32 dx = p_hole->m_dir ? 0.0f : 2.0f*step;
f32 dz = p_hole->m_dir ? 2.0f*step : 0.0f;
// loop over length of hole
for (u32 l=0; l<p_hole->m_cnt; l++)
{
// test for overlap with the current window
if (x>=x_min && x<x_max && z>=z_min && z<z_max)
{
// translate coords to col/row values
i32 c = (i32)((x-x_min) * inv_step);
i32 r = (i32)((z-z_min) * inv_step);
// punch out the hole
outcodes[r*width + c] |= 0x80;
// test camera against hole
f32 x0 = x - step;
f32 x1 = x + step;
f32 z0 = z - step;
f32 z1 = z + step;
f32 x_cam = spu_extract(g_pViewData->m_camera_position, 0);
f32 z_cam = spu_extract(g_pViewData->m_camera_position, 2);
if (x_cam>=x0 && x_cam<x1 && z_cam>=z0 && z_cam<z1)
{
g_RenderData.m_b_camera_over_water = false;
}
}
// step coords
x += dx;
z += dz;
}
}
}
}
int
main(unsigned long long id) {
vector unsigned int x = get_vector_param_3();
vector unsigned int count = (vector unsigned int){0,0,0,0};
vector unsigned int result = (vector unsigned int){0,0,0,0};
spu_ready();
count = popc(x);
result = reduce_word(count);
spu_write_out_mbox(spu_extract(result, 0));
return SPU_SUCCESS;
}
int
strncmp_ea (__ea void *s1, __ea const void *s2, size_ea_t n3)
{
__ea void *curr_s1 = (__ea void *) s1;
__ea void *curr_s2 = (__ea void *) s2;
void *l_s1;
void *l_s2;
int min;
size_ea_t s2_n;
size_ea_t s1_n;
int ret;
vec_uint4 end_v;
ret = 0; /* in case n3 is 0 */
while (n3)
{
l_s2 = __cache_fetch (curr_s2);
l_s1 = __cache_fetch (curr_s1);
/*
* Use the smaller of the size left to compare (n3), the space left in
* s2 cacheline (s2_n), or the space left in the s1 cacheline (s1_n)
*/
s2_n = ROUND_UP_NEXT_128 ((size_ea_t) curr_s2) - (size_ea_t) curr_s2;
s1_n = ROUND_UP_NEXT_128 ((size_ea_t) curr_s1) - (size_ea_t) curr_s1;
min = three_way_min (s2_n, s1_n, n3);
ret = _strncmp_internal (l_s1, l_s2, min, &end_v, 1);
/*
* Only the first slot of end_v is set.
*/
/* if (ret || spu_extract(spu_cmpeq(end_v, 0), 0)) { */
/* if (ret || spu_extract(spu_gather(spu_cmpeq(end_v, 0)), 0)) { */
if (ret || spu_extract (end_v, 0))
/*
* If any NUL values were seen (end_v values of zero) we still have
* to return ret, as it might not be zero.
*/
return ret;
curr_s2 += min;
curr_s1 += min;
n3 -= min;
}
return ret;
}
unsigned int
__mfc_tag_reserve (void)
{
vector unsigned int mask = (vector unsigned int)
{ 0x80000000, 0x80000000, 0x80000000, 0x80000000 };
vector unsigned int count_zeros, is_valid;
vector signed int count_neg;
count_zeros = spu_cntlz (__mfc_tag_table);
count_neg = spu_sub (0, (vector signed int) count_zeros);
mask = spu_rlmask (mask, (vector signed int) count_neg);
__mfc_tag_table = spu_andc (__mfc_tag_table, mask);
is_valid = spu_cmpeq (count_zeros, 32);
count_zeros = spu_sel (count_zeros, is_valid, is_valid);
return spu_extract (count_zeros, 0);
}
void pull(int side){
int avail_in = num_free_in_buffer(side);
int avail_mm = mcb[am].data_size[side] - md[am].num_pulled[side];
int num_pull = avail_in < avail_mm ? avail_in : avail_mm;
num_pull = num_pull < MAX_DMA_SIZE ? num_pull : MAX_DMA_SIZE;
int head = spu_extract(md[am].idx[side][HEAD],0);
int avail_from_head = mcb[am].buffer_size[side] - head;
int first_pull = num_pull < avail_from_head ? num_pull : avail_from_head;
if(!first_pull)
return;
// pull #first_pull
unsigned int to_ea = (unsigned int) &md[am].buffer[side][head];
int tag = mfc_tag_reserve();
if(tag == MFC_TAG_INVALID){
return;
} else {
md[am].held_tag[side] = tag;
}
mfc_get((void*)to_ea,
mcb[am].block_addr[side],
first_pull * sizeof(vector signed int),
md[am].held_tag[side],
0,0);
mcb[am].block_addr[side] += first_pull * sizeof(vector signed int);
if(first_pull < num_pull){
to_ea = (unsigned int) &md[am].buffer[side][0];
int second_pull = num_pull - first_pull;
mfc_get((void*)to_ea,
mcb[am].block_addr[side],
second_pull * sizeof(vector signed int),
md[am].held_tag[side],
0,0);
mcb[am].block_addr[side] += second_pull * sizeof(vector signed int);
}
md[am].num_waiting[side] = num_pull;
}
unsigned int
__mfc_multi_tag_release (unsigned int first_tag, unsigned int number_of_tags)
{
vector unsigned int table_copy, tmp, tmp1;
vector unsigned int one = (vector unsigned int)
{ 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF };
vector unsigned int is_invalid;
unsigned int last_tag;
vector unsigned int has_been_reserved;
last_tag = first_tag + number_of_tags;
table_copy = spu_sl (one, number_of_tags);
table_copy = spu_rl (table_copy, -last_tag);
table_copy = spu_xor (table_copy, -1);
/* Make sure the tags are in range and valid. */
tmp = spu_cmpgt (spu_promote(last_tag, 0), 32);
tmp1 = spu_cmpgt (spu_promote(number_of_tags, 0), 32);
is_invalid = spu_cmpgt (spu_promote(first_tag, 0), 31);
/* All bits are set to 1 if invalid, 0 if valid. */
is_invalid = spu_or (tmp, is_invalid);
is_invalid = spu_or (tmp1, is_invalid);
/* check whether these tags have been reserved */
tmp = spu_rlmask (one, (int)-number_of_tags);
tmp1 = spu_sl (__mfc_tag_table, first_tag);
has_been_reserved = spu_cmpgt(tmp1, tmp);
is_invalid = spu_or (has_been_reserved, is_invalid);
table_copy = spu_sel (__mfc_tag_table, table_copy, table_copy);
__mfc_tag_table = spu_sel (table_copy, __mfc_tag_table, is_invalid);
return spu_extract (is_invalid, 0);
}
/**
* 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 ClipToRectangle(vf32 clip_min, vf32 clip_max)
{
// convert from world coords to integer rows & cols
vf32 norm_min = spu_mul(clip_min - origin_world, spu_splats(inv_step));
vf32 norm_max = spu_mul(clip_max - origin_world, spu_splats(inv_step));
vi32 int_min = VecFloor4(norm_min);
vi32 int_max = VecCeil4 (norm_max);
// expand rectangle by 1 gridpoint because quads incident to the verts we're about to cull out will also be culled out
i32 c_min = spu_extract(int_min, 0) - 1;
i32 c_max = spu_extract(int_max, 0) + 1;
i32 r_min = spu_extract(int_min, 2) - 1;
i32 r_max = spu_extract(int_max, 2) + 1;
// trim loop bounds to rectangle so we don't splat memory
i32 c0 = c_min >= 0 ? c_min : 0;
i32 c1 = c_max < nc ? c_max+1 : nc;
i32 r0 = r_min >= 0 ? r_min : 0;
i32 r1 = r_max < nr ? r_max+1 : nr;
// cull left points
if (c_min>=0 && c_min<nc)
{
u8 *p = &g_Outcodes[r0*nc+c_min];
for (i32 r=r0; r<r1; r++,p+=nc)
{
*p |= 0x80;
}
}
// cull right points
if (c_max>=0 && c_max<nc)
{
u8 *p = &g_Outcodes[r0*nc+c_max];
for (i32 r=r0; r<r1; r++,p+=nc)
{
*p |= 0x80;
}
}
// cull upper points
if (r_min>=0 && r_min<nr)
{
u8 *p = &g_Outcodes[r_min*nc+c0];
for (i32 c=c0; c<c1; c++,p++)
{
*p |= 0x80;
}
}
// cull lower points
if (r_max>=0 && r_max<nr)
{
u8 *p = &g_Outcodes[r_max*nc+c0];
for (i32 c=c0; c<c1; c++,p++)
{
*p |= 0x80;
}
}
}
void InitBasisEtc()
{
// Use a fixed initial step size for now; 128m for lod 0.
// This yields an fft tile size of 32 x 128m = 4096m for lod 0
// and maximum dimensions of 8192m x 8192m
step = g_R2OCon.m_step;
vf32 step_vec = (vf32){step, 0, step, 0};
// get inverse-step using float magic (since taking the reciprocal of a power of 2 yields a 1-bit error)
qword q_step = si_from_float(step);
qword q_magic = si_ilhu(0x7F00);
inv_step = si_to_float(si_sf(q_step, q_magic));
// set clip window
clip_min = g_WaterObject.m_origin;
clip_max = g_WaterObject.m_origin + g_WaterObject.m_dimensions;
// set origin at gridpoint below clip min
f32 magic_float = 1.5f * 8388608.0f * step;
vf32 magic_vf32 = (vf32){magic_float, 0, magic_float, 0};
origin_world = (clip_min + magic_vf32) - magic_vf32;
// compute gridpoint above clip max
vf32 max_corner = (clip_max + magic_vf32) - magic_vf32;
max_corner += step_vec;
// offset both corners by the necessary amount of padding
origin_world -= step_vec * spu_splats(8.0f);
max_corner += step_vec * spu_splats(8.0f);
// set num cols & num rows
vf32 dims = max_corner - origin_world;
nc = (i32)(spu_extract(dims,0) * inv_step) + 1;
nr = (i32)(spu_extract(dims,2) * inv_step) + 1;
// record true nc, nr
true_nc = nc - 16;
true_nr = nr - 16;
// alignment requirements (ooh, that's a bit strict)
nc = (nc + 7) & -8;
nr = (nr + 7) & -8;
// deal with large grids
if (nc > 80)
{
nc = 80;
true_nc = 64;
dims = spu_insert((nc-1)*step, dims, 0);
}
if (nr > 80)
{
nr = 80;
true_nr = 64;
dims = spu_insert((nr-1)*step, dims, 2);
}
max_corner = origin_world + dims;
even_step = step;
even_inv_step = inv_step;
even_basis_col = (vf32){1.0f, 0.0f, 0.0f, 0.0f};
even_basis_row = (vf32){0.0f, 0.0f, 1.0f, 0.0f};
const f32 r = 0.707106781187f;
odd_step = even_step * r;
odd_inv_step= even_inv_step * r * 2.0f;
odd_basis_col = (vf32){ r, 0.0f, r, 0.0f};
odd_basis_row = (vf32){-r, 0.0f, r, 0.0f};
basis_col = even_basis_col;
basis_row = even_basis_row;
dvc_world = spu_splats(step) * basis_col;
dvr_world = spu_splats(step) * basis_row;
// set base lod origin
g_RenderData.m_origins[0] = origin_world;
g_RenderData.m_cols_rows[0] = nc<<8 | nr;
c0_amb = 0;
r0_amb = 0;
SetBasisEtc(0,0);
}
