void check_pull_dma(int side){
// Check left
if(md[am].held_tag[side] < 32){
mfc_write_tag_mask( 1 << md[am].held_tag[side] );
int status = mfc_read_tag_status_immediate();
if(status){
// Update idx
md[am].idx[side][HEAD] = spu_add(md[am].idx[side][HEAD], md[am].num_waiting[side]);
vector signed int buffer_size = spu_splats(mcb[am].buffer_size[side] -1);
vector unsigned int cmp_v = spu_cmpgt(md[am].idx[side][HEAD], buffer_size);
vector signed int zeros = {0,0,0,0};
buffer_size = spu_add(buffer_size,1);
zeros = spu_sel(zeros,buffer_size,cmp_v);
md[am].idx[side][HEAD] = spu_sub(md[am].idx[side][HEAD],zeros);
md[am].num_pulled[side] += md[am].num_waiting[side];
md[am].num_waiting[side] = 0;
if(md[am].num_pulled[side] == mcb[am].data_size[side]){
md[am].mm_depleted[side] = 1;
}
// Release tag
mfc_tag_release( md[am].held_tag[side] );
md[am].held_tag[side] = 32;
}
}
}
inline vector float SAHCostSIMD(vector float invarea, vector float ctravers, vector float cleft, vector float aleft, vector float cright, vector float aright)
{
vector float l = spu_mul(cleft, spu_mul(aleft, invarea));
vector float r = spu_mul(cright, spu_mul(aright, invarea));
return spu_add(ctravers, spu_add(l, r));
}
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 add()
{
int i, j, n;
n = SIZE * sizeof(float);
for (i = 0; (i + SIZE) < args.N; i += SIZE) {
mfc_get((void *)&ls1[0], (unsigned int )&args.a[i], n, TAG, 0, 0);
mfc_get((void *)&ls2[0], (unsigned int )&args.b[i], n, TAG, 0, 0);
mfc_write_tag_mask(1 << TAG);
mfc_read_tag_status_all();
for (j = 0; j < (SIZE / 4); ++j)
ls3[j] = spu_add(ls1[j], ls2[j]);
mfc_put((void *)&ls3[0], (unsigned int )&args.c[i], n, TAG, 0, 0);
}
mfc_write_tag_mask(1 << TAG);
mfc_read_tag_status_all();
if (unlikely(i < args.N)) {
/*
* args.N - i will be smaller than SIZE at this point so
* it is safe to do a DMA transfer.
* We need to make sure that size is a multiple of 16.
*/
n = ((args.N - i) * sizeof(float)) & (~127);
mfc_get((void *)&ls1[0], (unsigned int )&args.a[i], n, TAG, 0, 0);
mfc_get((void *)&ls2[0], (unsigned int )&args.b[i], n, TAG, 0, 0);
mfc_write_tag_mask(1 << TAG);
mfc_read_tag_status_all();
/* n must be divisible by 4. */
for (j = 0; j < ((args.N - i) / 4); ++j)
ls3[j] = spu_add(ls1[j], ls2[j]);
mfc_put((void *)&ls3[0], (unsigned int )&args.c[i], n, TAG, 0, 0);
mfc_write_tag_mask(1 << TAG);
mfc_read_tag_status_all();
}
/*
* At this point it may be that i is still smaller than args.N if the length
* was not divisible by the number of SPUs times 16.
*/
}
void process_data_simd (float* buf_in, float* buf_out, unsigned int size)
{
unsigned int i;
vector float *vbuf_in, *vbuf_out;
vector float v1 = (vector float) {1.0f, 1.0f, 1.0f, 1.0f};
vbuf_in = (vector float*) buf_in;
vbuf_out = (vector float*) buf_out;
for (i = 0; i < (size / 4); i++)
{
vbuf_out[i] = spu_add (vbuf_in[i], v1);
}
}
/*
* This routine fills a dma list with the appropriate effective address
* and size for the dma list element.
*
* @param dma_list: the dma list to be filled
* @param num_elements: number of elements in list
* @param base_addr: base effective address
* @param elem size: size of each dma element
*
*/
void fill_dma_list (mfc_list_element_t* list, int num_elements, unsigned long long base_addr, unsigned int elem_size)
{
int i;
for (i = 0; i < num_elements; i++)
{
list[i].notify = 0;
list[i].size = elem_size;
list[i].eal = base_addr + i*elem_size;
}
}
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);
}
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 push(){
int avail_out = num_in_buffer(OUT);
if(!avail_out)
return;
int avail_parent = num_free_in_buffer(PARENT);
if(mcb[am].id == 0)
avail_parent = mcb[am].data_size[LEFT] + mcb[am].data_size[RIGHT];
int num_send = avail_out < avail_parent ? avail_out : avail_parent;
num_send = num_send < MAX_DMA_SIZE ? num_send : MAX_DMA_SIZE;
if(!num_send)
return;
int tag = mfc_tag_reserve();
if(tag == MFC_TAG_INVALID){
return;
} else
md[am].held_tag[OUT] = tag;
// send num_send vectors, in up to three DMA-put's
while(num_send > 0){
int parent_head = spu_extract(md[am].idx[PARENT][HEAD],0);
int free_from_head = mcb[am].buffer_size[PARENT] - parent_head;
int tail = spu_extract(md[am].idx[OUT][TAIL],0);
int avail_from_tail = mcb[am].buffer_size[OUT] - tail;
int part_send = num_send < free_from_head ? num_send : free_from_head;
part_send = part_send < avail_from_tail ? part_send : avail_from_tail;
unsigned int to = mcb[am].block_addr[OUT] + parent_head*sizeof(vector signed int);
mfc_put(&md[am].buffer[OUT][tail],
to,
part_send * sizeof(vector signed int),
md[am].held_tag[OUT],
0,0);
md[am].idx[PARENT][HEAD] = spu_add(md[am].idx[PARENT][HEAD], part_send);
parent_head = spu_extract(md[am].idx[PARENT][HEAD],0);
if(parent_head == mcb[am].buffer_size[PARENT])
md[am].idx[PARENT][HEAD] = spu_splats(0);
md[am].idx[OUT][TAIL] = spu_add(md[am].idx[OUT][TAIL], part_send);
tail = spu_extract(md[am].idx[OUT][TAIL],0);
if(tail == mcb[am].buffer_size[OUT])
md[am].idx[OUT][TAIL] = spu_splats(0);
num_send -= part_send;
}
// Inner nodes updates parent in buffer head idx
if(mcb[am].id)
mfc_putf(&md[am].idx[PARENT][HEAD],
mcb[am].idx_addr[OUT],
sizeof(vector signed int),
md[am].held_tag[OUT],
0,0);
}
void MinMaxBinFindBest3SIMD(minmaxbin_t *mmb, kdbuffer_t *result)
{
int i;
for(i=1; i < mmb->numbins; i++)
{
int j = mmb->numbins - i - 1;
vector float *min = (vector float *)mmb->minbins[i].b;
vector float *max = (vector float *)mmb->maxbins[j].b;
min[0] = spu_add(min[0], min[-1]);
max[0] = spu_add(max[0], max[1]);
}
vector float *vmax = (vector float*)result->baabb.max;
vector float *vmin = (vector float*)result->baabb.min;
vector float vwidth = spu_abs( spu_sub(*vmax, *vmin) );
vector float vnumbins = spu_splats(1/(float)mmb->numbins);
vector float vdelta = spu_mul(vwidth, vnumbins);
vector float vx = spu_add(*vmin, vdelta);
vector float vside = { vwidth[1] * vwidth[2], vwidth[0] * vwidth[2], vwidth[0] * vwidth[1], 0 };
vector float invarea = spu_splats( 1/(vwidth[0] * vside[0]));
vector float vctravers = spu_splats(2.0f);
vector float vbestcost = spu_splats(mmb->bestcost);
vector int vbesti = spu_splats(0);
vector float vbestx = vx;
for(i=0; i < mmb->numbins-1; i++)
{
vector float aleft, aright;
AreaLeftRight(*vmin, *vmax, vside, vx, &aleft, &aright);
vector float *vminbin = (vector float *)mmb->minbins[i].b;
vector float *vmaxbin = (vector float *)mmb->maxbins[i+1].b;
vector float cost = SAHCostSIMD(invarea, vctravers, *vminbin, aleft, *vmaxbin, aright);
vector unsigned int cmp = spu_cmpgt(cost, vbestcost);
vbestcost = spu_sel(cost, vbestcost, cmp);
vbesti = spu_sel(spu_splats(i), vbesti, cmp);
vbestx = spu_sel(vx, vbestx, cmp);
vx = spu_add(vx, vdelta);
}
int axis = 0;
float bestcost = vbestcost[axis];
if(vbestcost[1] < bestcost)
{
axis = 1;
bestcost = vbestcost[1];
}
if(vbestcost[2] < bestcost)
{
axis = 2;
bestcost = vbestcost[2];
}
int index = vbesti[axis];
result->plane = vbestx[axis];
result->axis = axis;
result->left_size = (int)mmb->minbins[ index ].b[axis];
result->right_size = (int)mmb->maxbins[ index+1 ].b[axis];
mmb->bestcost = vbestcost[axis];
}
void process_data_simd (float* buf_in, float* buf_out, unsigned int size)
{
unsigned int i;
vector float *vbuf_in, *vbuf_out;
vector float v1 = (vector float) {1.0f, 1.0f, 1.0f, 1.0f};
vbuf_in = (vector float*) buf_in;
vbuf_out = (vector float*) buf_out;
for (i = 0; i < (size / 4); i++)
{
vbuf_out[i] = spu_add (vbuf_in[i], v1);
}
}
int main(unsigned long long speid __attribute__ ((unused)),
unsigned long long argp,
unsigned long long envp __attribute__ ((unused)))
{
unsigned int tag;
unsigned long long in_addr, out_addr;
int i, num_chunks;
#ifdef USE_TIMER
uint64_t start, time_working;
spu_slih_register (MFC_DECREMENTER_EVENT, spu_clock_slih);
spu_clock_start();
start = spu_clock_read();
#endif /* USE_TIMER */
/* First, we reserve a MFC tag for use */
tag = mfc_tag_reserve();
if (tag == MFC_TAG_INVALID)
{
printf ("SPU ERROR, unable to reserve tag\n");
return 1;
}
/* issue DMA transfer to get the control block information from
* system memory */
mfc_get (&control_block, argp, sizeof (control_block_t), tag, 0, 0);
/* wait for the DMA to complete */
mfc_write_tag_mask (1 << tag);
mfc_read_tag_status_all ();
/* calculate the number of blocks (chunks) that this spe is assigned
* to process */
num_chunks = control_block.num_elements_per_spe/CHUNK_SIZE;
/*
* This is the main loop. We basically goes through the num_chunks
* and fetches one 'chunk' of data at a time, process it, and write
* it back to system memory until done.
*/
for (i = 0; i < num_chunks; i++)
{
/* set the in_addr and out_addr variables, we will use these for
* issuing DMA get and put commands */
in_addr = control_block.in_addr + (i* CHUNK_SIZE * sizeof(float));
out_addr = control_block.out_addr + (i * CHUNK_SIZE * sizeof(float));
/* issue a DMA get command to fetch the chunk of data from system memory */
mfc_get (local_buffer_in, in_addr, CHUNK_SIZE * sizeof(float),
tag, 0, 0);
/* wait for the DMA get to complete */
mfc_write_tag_mask (1 << tag);
mfc_read_tag_status_all ();
/* invoke process_data to work on the data that's just been moved into
* local store*/
process_data_simd (local_buffer_in, local_buffer_out, CHUNK_SIZE);
/* issue the DMA put command to transfer result from local memory to
* system memory */
mfc_put (local_buffer_out, out_addr, CHUNK_SIZE * sizeof(float),
tag, 0, 0);
/* wait for the DMA put to complete */
mfc_write_tag_mask (1 << tag);
mfc_read_tag_status_all ();
}
#ifdef USE_TIMER
time_working = (spu_clock_read() - start);
spu_clock_stop();
printf ("SPE time_working = %lld\n", time_working);
#endif /* USE_TIMER */
return 0;
}
static inline void build_blit_list(
vec_uint4* dma_list_buffer,
unsigned long eal, unsigned long stride)
{
#ifdef DEBUG_1
printf("build_blit_list: eal=%lx stride=%d\n", eal, stride);
#endif
unsigned long eal1 = eal + stride;
unsigned long stride2 = 2 * stride;
unsigned long stride8 = 8 * stride;
vec_uint4 block0 = { 128, eal, 128, eal1 };
vec_uint4 step2 = { 0, stride2, 0, stride2};
vec_uint4 step4 = spu_add(step2, step2);
vec_uint4 step6 = spu_add(step4, step2);
vec_uint4 step8 = { 0, stride8, 0, stride8};
vec_uint4 step16 = spu_add(step8, step8);
vec_uint4 block2 = spu_add(block0, step2);
vec_uint4 block4 = spu_add(block0, step4);
vec_uint4 block6 = spu_add(block0, step6);
vec_uint4 block8 = spu_add(block0, step8);
vec_uint4 block10 = spu_add(block8, step2);
vec_uint4 block12 = spu_add(block8, step4);
vec_uint4 block14 = spu_add(block8, step6);
vec_uint4 block16 = spu_add(block8, step8);
vec_uint4 block18 = spu_add(step16, block2);
vec_uint4 block20 = spu_add(step16, block4);
vec_uint4 block22 = spu_add(step16, block6);
vec_uint4 block24 = spu_add(step16, block8);
vec_uint4 block26 = spu_add(step16, block10);
vec_uint4 block28 = spu_add(step16, block12);
vec_uint4 block30 = spu_add(step16, block14);
dma_list_buffer[0] = block0;
dma_list_buffer[1] = block2;
dma_list_buffer[2] = block4;
dma_list_buffer[3] = block6;
dma_list_buffer[4] = block8;
dma_list_buffer[5] = block10;
dma_list_buffer[6] = block12;
dma_list_buffer[7] = block14;
dma_list_buffer[8] = block16;
dma_list_buffer[9] = block18;
dma_list_buffer[10] = block20;
dma_list_buffer[11] = block22;
dma_list_buffer[12] = block24;
dma_list_buffer[13] = block26;
dma_list_buffer[14] = block28;
dma_list_buffer[15] = block30;
#ifdef DEBUG_1
int i,j;
for (i=0;i<16;i++) {
printf("%lx(%x): ", &dma_list_buffer[i], i);
for (j=0; j<4; j++) {
printf("%lx ", spu_extract(dma_list_buffer[i], j));
}
printf("\n");
}
#endif
}
inline vector real_t
advec_diff_v(vector real_t cell_size,
vector real_t c2l, vector real_t w2l, vector real_t d2l,
vector real_t c1l, vector real_t w1l, vector real_t d1l,
vector real_t c, vector real_t w, vector real_t d,
vector real_t c1r, vector real_t w1r, vector real_t d1r,
vector real_t c2r, vector real_t w2r, vector real_t d2r)
{
vector real_t acc1, acc2, acc3;
vector real_t wind, diff_term, advec_term;
vector real_t advec_term_pos, advec_term_neg;
vector real_t advec_termR, advec_termL;
acc1 = spu_add(w1l, w);
wind = spu_mul(acc1, HALF);
acc1 = spu_mul(c1l, FIVE);
acc2 = spu_mul(c, TWO);
advec_term_pos = spu_add(acc1, acc2);
advec_term_pos = spu_sub(advec_term_pos, c2l);
acc1 = spu_mul(c1l, TWO);
acc2 = spu_mul(c, FIVE);
advec_term_neg = spu_add(acc1, acc2);
advec_term_neg = spu_sub(advec_term_neg, c1r);
acc1 = (vector real_t)spu_cmpgt(wind, ZERO);
acc1 = spu_and(acc1, advec_term_pos);
acc2 = (vector real_t)spu_cmpgt(ZERO, wind);
acc2 = spu_and(acc2, advec_term_neg);
advec_termL = spu_add(acc1, acc2);
advec_termL = spu_mul(advec_termL, SIXTH);
advec_termL = spu_mul(advec_termL, wind);
acc1 = spu_add(w1r, w);
wind = spu_mul(acc1, HALF);
acc1 = spu_mul(c, FIVE);
acc2 = spu_mul(c1r, TWO);
advec_term_pos = spu_add(acc1, acc2);
advec_term_pos = spu_sub(advec_term_pos, c1l);
acc1 = spu_mul(c, TWO);
acc2 = spu_mul(c1r, FIVE);
advec_term_neg = spu_add(acc1, acc2);
advec_term_neg = spu_sub(advec_term_neg, c2r);
acc1 = (vector real_t)spu_cmpgt(wind, ZERO);
acc1 = spu_and(acc1, advec_term_pos);
acc2 = (vector real_t)spu_cmpgt(ZERO, wind);
acc2 = spu_and(acc2, advec_term_neg);
advec_termR = spu_add(acc1, acc2);
advec_termR = spu_mul(advec_termR, SIXTH);
advec_termR = spu_mul(advec_termR, wind);
acc1 = spu_sub(advec_termL, advec_termR);
advec_term = VEC_DIVIDE(acc1, cell_size);
acc1 = spu_add(d1l, d);
acc1 = spu_mul(acc1, HALF);
acc3 = spu_sub(c1l, c);
acc1 = spu_mul(acc1, acc3);
acc2 = spu_add(d, d1r);
acc2 = spu_mul(acc2, HALF);
acc3 = spu_sub(c, c1r);
acc2 = spu_mul(acc2, acc3);
acc1 = spu_sub(acc1, acc2);
acc2 = spu_mul(cell_size, cell_size);
diff_term = VEC_DIVIDE(acc1, acc2);
return spu_add(advec_term, diff_term);
}
int kernel(lwp_functions* pf,
void* params,
void* inout,
unsigned int iter,
unsigned int n)
{
Ternary_params* p = (Ternary_params*)params;
switch (p->cmd)
{
case AM:
{
int length = p->length / 4;
vector float *a = (vector float *)inout;
vector float *b = a + length;
vector float *c = a + 2 * length;
unsigned int i;
for (i = 0; i != length; ++i, ++a, ++b, ++c)
*a = spu_mul(spu_add(*a, *b), *c);
return 0;
}
case MA:
{
int length = p->length / 4;
vector float *a = (vector float *)inout;
vector float *b = a + length;
vector float *c = a + 2 * length;
unsigned int i;
for (i = 0; i != length; ++i, ++a, ++b, ++c)
*a = spu_madd(*a, *b, *c);
return 0;
}
case CAM:
{
static vector unsigned char lo =
(vector unsigned char) { 0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23};
static vector unsigned char hi =
(vector unsigned char) { 8, 9, 10, 11, 24, 25, 26, 27,
12, 13, 14, 15, 28, 29, 30, 31};
int length = p->length / 4;
float *a = (float *)inout;
float *b = a + 8 * length;
float *c = a + 16 * length;
unsigned int i;
// (a + b) * c:
// r.r = (a.r+b.r)*c.r - (a.i+b.i)*c.i
// r.i = (a.r+b.r)*c.i + (a.i+b.i)*c.r
for (i = 0; i != length; ++i, a+=8, b+=8, c+=8)
{
vector float av = {*a, *(a+2), *(a+4), *(a+6)}; // a.r
vector float bv = {*b, *(b+2), *(b+4), *(b+6)}; // b.r
vector float cv = {*c, *(c+2), *(c+4), *(c+6)}; // c.r
vector float dv = {*(a+1), *(a+3), *(a+5), *(a+7)}; // a.i
vector float ev = {*(b+1), *(b+3), *(b+5), *(b+7)}; // b.i
vector float fv = {*(c+1), *(c+3), *(c+5), *(c+7)}; // c.i
vector float trv = spu_add(av, bv); // a.r+b.r
vector float tiv = spu_add(dv, ev); // a.i+b.i
vector float sv = spu_mul(trv, cv); // (a.r+b.r)*c.r
vector float tv = spu_mul(trv, fv); // (a.r+b.r)*c.i
vector float real = spu_nmsub(tiv, fv, sv); // r.r
vector float imag = spu_madd(tiv, cv, tv); // r.i
// interleave result
*(vector float *)a = spu_shuffle(real, imag, lo);
*(vector float *)(a+4) = spu_shuffle(real, imag, hi);
}
return 0;
}
case CMA:
{
static vector unsigned char lo =
(vector unsigned char) { 0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23};
static vector unsigned char hi =
(vector unsigned char) { 8, 9, 10, 11, 24, 25, 26, 27,
12, 13, 14, 15, 28, 29, 30, 31};
int length = p->length / 4;
float *a = (float *)inout;
float *b = a + 8 * length;
float *c = a + 16 * length;
unsigned int i;
// a * b + c:
// r.r = a.r*b.r + c.r - a.i*b.i
// r.i = a.r*b.i + c.i + a.i*b.r
for (i = 0; i != length; ++i, a+=8, b+=8, c+=8)
{
vector float av = {*a, *(a+2), *(a+4), *(a+6)}; // a.r
vector float bv = {*b, *(b+2), *(b+4), *(b+6)}; // b.r
vector float cv = {*c, *(c+2), *(c+4), *(c+6)}; // c.r
vector float dv = {*(a+1), *(a+3), *(a+5), *(a+7)}; // a.i
vector float ev = {*(b+1), *(b+3), *(b+5), *(b+7)}; // b.i
vector float fv = {*(c+1), *(c+3), *(c+5), *(c+7)}; // c.i
vector float real = spu_nmsub(dv, ev, spu_madd(av, bv, cv)); // r.r
vector float imag = spu_madd(dv, bv, spu_madd(av, ev, fv)); // r.i
// interleave result
*(vector float *)a = spu_shuffle(real, imag, lo);
*(vector float *)(a+4) = spu_shuffle(real, imag, hi);
}
return 0;
}
case ZAM:
{
int length = p->length / 4;
float *a_re = (float *)inout;
float *a_im = a_re + 4 * length;
float *b_re = a_re + 8 * length;
float *b_im = a_re + 12 * length;
float *c_re = a_re + 16 * length;
float *c_im = a_re + 20 * length;
unsigned int i;
// (a + b) * c:
// r.r = (a.r+b.r)*c.r - (a.i+b.i)*c.i
// r.i = (a.r+b.r)*c.i + (a.i+b.i)*c.r
for (i = 0; i != length;
++i, a_re+=4, b_re+=4, c_re+=4, a_im+=4, b_im+=4, c_im+=4)
{
vector float *av = (vector float *)a_re;
vector float *bv = (vector float *)b_re;
vector float *cv = (vector float *)c_re;
vector float *dv = (vector float *)a_im;
vector float *ev = (vector float *)b_im;
vector float *fv = (vector float *)c_im;
vector float trv = spu_add(*av, *bv); // a.r+b.r
vector float tiv = spu_add(*dv, *ev); // a.i+b.i
vector float sv = spu_mul(trv, *cv); // (a.r+b.r)*c.r
vector float tv = spu_mul(trv, *fv); // (a.r+b.r)*c.i
*av = spu_nmsub(tiv, *fv, sv); // r.r
*dv = spu_madd(tiv, *cv, tv); // r.i
}
return 0;
}
case ZMA:
{
int length = p->length / 4;
float *a_re = (float *)inout;
float *a_im = a_re + 4 * length;
float *b_re = a_re + 8 * length;
float *b_im = a_re + 12 * length;
float *c_re = a_re + 16 * length;
float *c_im = a_re + 20 * length;
unsigned int i;
// a * b + c:
// r.r = a.r*b.r + c.r - a.i*b.i
// r.i = a.r*b.i + c.i + a.i*b.r
for (i = 0; i != length;
++i, a_re+=4, b_re+=4, c_re+=4, a_im+=4, b_im+=4, c_im+=4)
{
vector float *av = (vector float *)a_re;
vector float *bv = (vector float *)b_re;
vector float *cv = (vector float *)c_re;
vector float *dv = (vector float *)a_im;
vector float *ev = (vector float *)b_im;
vector float *fv = (vector float *)c_im;
vector float tmp = spu_nmsub(*dv, *ev, spu_madd(*av, *bv, *cv));
*dv = spu_madd(*dv, *bv, spu_madd(*av, *ev, *fv));
*av = tmp;
}
return 0;
}
}
return 1;
}
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);
}
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;
}
void merge_buffers(){
vector unsigned int cmp_v, cmp_v2;
const vector signed int one_at_0 = {1,0,0,0};
const vector signed int one_at_1 = {0,1,0,0};
const vector signed int one_at_2 = {0,0,1,0};
const vector signed int ones = {1,1,1,1};
const vector signed int zeros = {0,0,0,0};
const vector unsigned char cmp_v_shuffle_mask = {31,31,31,31,
31,31,31,31,
31,31,31,31,
31,31,31,31};
vector unsigned char rev_mask;
const vector unsigned char rev_left = {12,13,14,15,
8,9,10,11,
4,5,6,7,
0,1,2,3};
const vector unsigned char rev_right = {28,29,30,31,
24,25,26,27,
20,21,22,23,
16,17,18,19};
vector signed int *out_head_idx;
if(mcb[am].local[OUT] < 255){
int parent_idx = mcb[am].local[OUT];
int side = (mcb[am].id+1)&1;
out_head_idx = (vector signed int*) &md[parent_idx].idx[side][HEAD];
} else {
out_head_idx = (vector signed int*) &md[am].idx[OUT][HEAD];
}
vector signed int *left_tail_idx = (vector signed int*) &md[am].idx[LEFT][TAIL];
vector signed int *right_tail_idx = (vector signed int*) &md[am].idx[RIGHT][TAIL];
vector signed int size_v = {mcb[am].buffer_size[LEFT], mcb[am].buffer_size[RIGHT], mcb[am].buffer_size[OUT], 0};
vector signed int avail_v = {num_in_buffer(LEFT), num_in_buffer(RIGHT), num_free_in_buffer(OUT), 1};
vector signed int avail_before = { spu_extract(avail_v, 0), spu_extract(avail_v, 1), 0, 0 };
vector unsigned int avail = spu_gather( spu_cmpgt(avail_v, zeros) ); // avail = 0x0F if all avail_v > zeros
vector signed int *left, *right, *out;
left = (vector signed int*) &md[am].buffer[LEFT][ spu_extract(*left_tail_idx,0) ];
right = (vector signed int*) &md[am].buffer[RIGHT][ spu_extract(*right_tail_idx,0) ];
out = (vector signed int*) &md[am].buffer[OUT][ spu_extract(*out_head_idx,0) ];
#ifdef TRACE_TIME
dec_val2 = spu_read_decrementer();
#endif
while(spu_extract(avail,0) == 0x0F){
// cmp left and right to determine who gets eaten
cmp_v = spu_cmpgt(*left,*right);
cmp_v = spu_shuffle(cmp_v, cmp_v, cmp_v_shuffle_mask);
// cmp_v = {FFFF,FFFF,FFFF,FFFF} if left[3] > right[3]
*out = spu_sel(*left,*right,cmp_v);
rev_mask = spu_sel(rev_right,rev_left,(vector unsigned char)cmp_v);
*left = spu_shuffle(*left,*right,rev_mask);
// data to be sorted is now in out and left, left in descending order
sort_vectors(out,left);
// update index of the used side
if( spu_extract(cmp_v,0) ){
// left[3] > right[3]
*right_tail_idx = spu_add(*right_tail_idx,ones);
avail_v = spu_sub(avail_v, one_at_1);
right++;
// modulus hack
cmp_v2 = spu_cmpeq(*right_tail_idx, size_v);
if( __builtin_expect( spu_extract(cmp_v2,0) ,0) ){
*right_tail_idx = zeros;
right = (vector signed int*) &md[am].buffer[RIGHT][0];
}
} else {
*right = *left;
*left_tail_idx = spu_add(*left_tail_idx,ones);
avail_v = spu_sub(avail_v, one_at_0);
left++;
// modulus hack
cmp_v2 = spu_cmpeq(*left_tail_idx, size_v);
if( __builtin_expect( spu_extract(cmp_v2,0) ,0) ){
*left_tail_idx = zeros;
left = (vector signed int*) &md[am].buffer[LEFT][0];
}
}
// update out head idx
*out_head_idx = spu_add(*out_head_idx,ones);
avail_v = spu_sub(avail_v, one_at_2);
out++;
// modulus hack
cmp_v2 = spu_cmpeq(*out_head_idx, size_v);
if( __builtin_expect(spu_extract(cmp_v2,0),0) ){
out = (vector signed int*) &md[am].buffer[OUT][0];
*out_head_idx = zeros;
}
// is there data still available?
avail = spu_gather(spu_cmpgt(avail_v, zeros));
}
#ifdef TRACE_TIME
merge_loop_ticks += -(spu_read_decrementer() - dec_val2);
#endif
// how much got produced?
vector signed int consumed = spu_sub(avail_before, avail_v);
int consumed_left = spu_extract(consumed, 0);
int consumed_right = spu_extract(consumed, 1);
if(consumed_left)
update_tail(LEFT);
if(consumed_right)
update_tail(RIGHT);
md[am].consumed[LEFT] += consumed_left;
md[am].consumed[RIGHT] += consumed_right;
if(md[am].consumed[LEFT] == mcb[am].data_size[LEFT])
md[am].depleted[LEFT] = 1;
if(md[am].consumed[RIGHT] == mcb[am].data_size[RIGHT])
md[am].depleted[RIGHT] = 1;
if(mcb[am].local[OUT] < 255 && md[am].depleted[LEFT] && md[am].depleted[RIGHT]){
md[am].done = 1;
--num_active_mergers;
}
}
/* Scans the string pointed to by s for the character c and
* returns a pointer to the last occurance of c. If
* c is not found, then NULL is returned.
*/
char * strrchr(const char *s, int c)
{
int nskip;
vec_uchar16 *ptr, data, vc;
vec_uint4 cmp_c, cmp_0, cmp;
vec_uint4 res_ptr, res_cmp;
vec_uint4 mask, result;
vec_uint4 one = spu_splats(0xffffU);
/* Scan memory array a quadword at a time. Skip leading
* mis-aligned bytes.
*/
ptr = (vec_uchar16 *)s;
nskip = -((unsigned int)(ptr) & 15);
mask = spu_rlmask(one, nskip);
vc = spu_splats((unsigned char)(c));
data = *ptr++;
ptr = (vec_uchar16 *)((unsigned int)ptr & ~15);
cmp_c = spu_and(spu_gather(spu_cmpeq(data, vc)), mask);
cmp_0 = spu_and(spu_gather(spu_cmpeq(data, 0)), mask);
res_ptr = spu_splats(0U);
res_cmp = spu_splats(0U);
while (spu_extract(cmp_0, 0) == 0) {
cmp = spu_cmpeq(cmp_c, 0);
res_ptr = spu_sel(spu_promote((unsigned int)(ptr), 0), res_ptr, cmp);
res_cmp = spu_sel(cmp_c, res_cmp, cmp);
data = *ptr++;
cmp_c = spu_gather(spu_cmpeq(data, vc));
cmp_0 = spu_gather(spu_cmpeq(data, 0));
cmp = spu_cmpeq(cmp_c, 0);
}
/* Compute the location of the last character before termination
* character.
*
* First mask off compare results following the first termination character.
*/
mask = spu_sl(one, 31 - spu_extract(spu_cntlz(cmp_0), 0));
cmp_c = spu_and(cmp_c, mask);
/* Conditionally update res_ptr and res_cmd if a match was found in the last
* quadword.
*/
cmp = spu_cmpeq(cmp_c, 0);
res_ptr = spu_sel(spu_promote((unsigned int)(ptr), 0), res_ptr, cmp);
res_cmp = spu_sel(cmp_c, res_cmp, cmp);
/* Bit reserve res_cmp for locating last occurance.
*/
mask = spu_cmpeq(res_cmp, 0);
res_cmp = (vec_uint4)spu_maskb(spu_extract(res_cmp, 0));
res_cmp = spu_gather((vec_uchar16)spu_shuffle(res_cmp, res_cmp,
VEC_LITERAL(vec_uchar16,
15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0)));
/* Compute the location (ptr) of the last occurance of c. If no
* occurance was found (ie, element 0 of res_cmp == 0, then return
* NULL.
*/
result = spu_sub(spu_add(res_ptr, 15), spu_cntlz(res_cmp));
result = spu_andc(result, mask);
return ((char *)spu_extract(result, 0));
}
Triangle* getTriangleBuffer(Context* context)
{
// if we've already allocated a triangle buffer (and we're in the same context)
if (context == _currentTriangleContext && _currentTriangle)
return _currentTriangle;
// trash the default values
_currentTriangleContext = context;
_currentTriangle = NULL;
// read the current renderable cache line to ensure there is room for the triangle data
// in the cache line buffer; we do this by comparing against all 16 cache line blocks
// to make sure that extending the write pointer wouldn't clobber the data
unsigned long long cache_ea = context->renderableCacheLine;
if (cache_ea == 0)
return NULL;
char cachebuffer[128+127];
RenderableCacheLine* cache = (RenderableCacheLine*) ( ((unsigned int)cachebuffer+127) & ~127 );
// printf("GTB: reading to %x from %x:%x\n", cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea));
spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_GETLLAR_CMD);
spu_readch(MFC_RdAtomicStat);
// extendvalid = ( read<=write && test<end ) || ( read>write && test<read )
// extendvalid = ( read>write && read>test ) || ( read<=write && end>test )
// simplifies to extendvalid = selb(end, read, read>write) > test
// or extendvalid = selb(end>test, read>test, read>write)
// rewind = next >= end
// rewindvalid = read != 0
// valid = extendvalid && (!rewind || rewindvalid)
// = extendvalid && (!rewind || !rewindinvalid)
// = extendvalid && !(rewind && rewindinvalid)
// invalid = ! (extendvalid && !(rewind && rewindinvalid))
// = (!extendvalid || (rewind && rewindinvalid))
vec_ushort8 v_writeptr = spu_splats( cache->endTriangle );
vec_ushort8 v_readptr0 = cache->chunkTriangle[0];
vec_ushort8 v_readptr1 = cache->chunkTriangle[1];
vec_ushort8 v_testptr = spu_add(v_writeptr, TRIANGLE_MAX_SIZE);
vec_ushort8 v_nextptr = spu_add(v_writeptr, 2*TRIANGLE_MAX_SIZE);
vec_ushort8 v_endptr = spu_splats( (unsigned short)TRIANGLE_BUFFER_SIZE);
vec_ushort8 v_zero = spu_splats( (unsigned short) 0 );
vec_uchar16 v_merger = (vec_uchar16) { 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31 };
vec_ushort8 v_max0_test = spu_sel( v_endptr, v_readptr0, spu_cmpgt( v_readptr0, v_writeptr ) );
vec_ushort8 v_max1_test = spu_sel( v_endptr, v_readptr1, spu_cmpgt( v_readptr1, v_writeptr ) );
vec_ushort8 v_extend0_valid = spu_cmpgt( v_max0_test, v_testptr );
vec_ushort8 v_extend1_valid = spu_cmpgt( v_max1_test, v_testptr );
vec_ushort8 v_rewind0_invalid = spu_cmpeq( v_readptr0, v_zero );
vec_ushort8 v_rewind1_invalid = spu_cmpeq( v_readptr1, v_zero );
vec_ushort8 v_rewind8 = spu_cmpgt( v_nextptr, v_endptr );
vec_uchar16 v_extend_valid = (vec_uchar16) spu_shuffle( v_extend0_valid, v_extend1_valid, v_merger );
vec_uchar16 v_rewind_invalid = (vec_uchar16) spu_shuffle( v_rewind0_invalid, v_rewind1_invalid, v_merger );
vec_uchar16 v_rewind = (vec_uchar16) v_rewind8;
vec_uchar16 v_valid_rhs = spu_and( v_rewind_invalid, v_rewind );
vec_uchar16 v_invalid = spu_orc( v_valid_rhs, v_extend_valid );
// check to see if the chunk is being processed
vec_uint4 v_free = spu_gather(
spu_cmpeq( spu_splats( (unsigned char) CHUNKNEXT_FREE_BLOCK ), cache->chunkNext ) );
vec_uint4 v_invalid_bits = spu_andc( spu_gather( v_invalid ), (vec_uint4) v_free );
// if any of the bits are invalid, then no can do
if ( spu_extract(v_invalid_bits, 0) ) {
return NULL;
}
// fetch in the data before this triangle in the cache buffer
unsigned int offset = cache->endTriangle;
_currentTriangleBufferExtra = offset & 127;
unsigned long long trianglebuffer_ea = cache_ea + TRIANGLE_OFFSET_FROM_CACHE_LINE + (offset & ~127);
if (_currentTriangleBufferExtra) {
spu_mfcdma64(_currentTriangleBuffer, mfc_ea2h(trianglebuffer_ea), mfc_ea2l(trianglebuffer_ea), 128, 0, MFC_GET_CMD);
// ensure DMA did actually complete
mfc_write_tag_mask(1<<0);
mfc_read_tag_status_all();
}
// final bit of initialisation
_currentTriangle = (Triangle*) (_currentTriangleBuffer+_currentTriangleBufferExtra);
_currentTriangleOffset = offset;
_currentTriangleRewind = v_rewind8;
_currentTriangleCacheEndTriangleEAL = mfc_ea2l(cache_ea) + (((char*)&cache->endTriangle) - ((char*)cache));
_currentTriangleCacheEndTriangleEAH = mfc_ea2h(cache_ea);
_currentTriangleBufferEA = trianglebuffer_ea;
// printf("Allocated new triangle buffer: %x\n", offset);
// and return the buffer ready to go
return _currentTriangle;
}
