bool isNot0() const {
#ifdef HAVE_SSE4
return !(_mm_test_all_zeros(this->m_, this->m_));
#else
return (this->merge() ? true : false);
#endif
}
bool square128::operator==(square128& other)
{
for (int i = 0; i < 128; i++)
{
__m128i tmp = rows[i] ^ other.rows[i];
if (not _mm_test_all_zeros(tmp, tmp))
return false;
}
return true;
}
/*****************************************************************************
* This function utilises 3 properties of the cost function lookup tables, *
* constructed in using 'cal_nmvjointsadcost' and 'cal_nmvsadcosts' in *
* vp9_encoder.c. *
* For the joint cost: *
* - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
* For the component costs: *
* - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
* (Equal costs for both components) *
* - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
* (Cost function is even) *
* If these do not hold, then this function cannot be used without *
* modification, in which case you can revert to using the C implementation, *
* which does not rely on these properties. *
*****************************************************************************/
int vp9_diamond_search_sad_avx(const MACROBLOCK *x,
const search_site_config *cfg,
MV *ref_mv, MV *best_mv, int search_param,
int sad_per_bit, int *num00,
const vp9_variance_fn_ptr_t *fn_ptr,
const MV *center_mv) {
const int_mv maxmv = pack_int_mv(x->mv_row_max, x->mv_col_max);
const __m128i v_max_mv_w = _mm_set1_epi32(maxmv.as_int);
const int_mv minmv = pack_int_mv(x->mv_row_min, x->mv_col_min);
const __m128i v_min_mv_w = _mm_set1_epi32(minmv.as_int);
const __m128i v_spb_d = _mm_set1_epi32(sad_per_bit);
const __m128i v_joint_cost_0_d = _mm_set1_epi32(x->nmvjointsadcost[0]);
const __m128i v_joint_cost_1_d = _mm_set1_epi32(x->nmvjointsadcost[1]);
// search_param determines the length of the initial step and hence the number
// of iterations.
// 0 = initial step (MAX_FIRST_STEP) pel
// 1 = (MAX_FIRST_STEP/2) pel,
// 2 = (MAX_FIRST_STEP/4) pel...
const MV *ss_mv = &cfg->ss_mv[cfg->searches_per_step * search_param];
const intptr_t *ss_os = &cfg->ss_os[cfg->searches_per_step * search_param];
const int tot_steps = cfg->total_steps - search_param;
const int_mv fcenter_mv = pack_int_mv(center_mv->row >> 3,
center_mv->col >> 3);
const __m128i vfcmv = _mm_set1_epi32(fcenter_mv.as_int);
const int ref_row = clamp(ref_mv->row, minmv.as_mv.row, maxmv.as_mv.row);
const int ref_col = clamp(ref_mv->col, minmv.as_mv.col, maxmv.as_mv.col);
int_mv bmv = pack_int_mv(ref_row, ref_col);
int_mv new_bmv = bmv;
__m128i v_bmv_w = _mm_set1_epi32(bmv.as_int);
const int what_stride = x->plane[0].src.stride;
const int in_what_stride = x->e_mbd.plane[0].pre[0].stride;
const uint8_t *const what = x->plane[0].src.buf;
const uint8_t *const in_what = x->e_mbd.plane[0].pre[0].buf +
ref_row * in_what_stride + ref_col;
// Work out the start point for the search
const uint8_t *best_address = in_what;
const uint8_t *new_best_address = best_address;
#if ARCH_X86_64
__m128i v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
#else
__m128i v_ba_d = _mm_set1_epi32((intptr_t)best_address);
#endif
unsigned int best_sad;
int i;
int j;
int step;
// Check the prerequisite cost function properties that are easy to check
// in an assert. See the function-level documentation for details on all
// prerequisites.
assert(x->nmvjointsadcost[1] == x->nmvjointsadcost[2]);
assert(x->nmvjointsadcost[1] == x->nmvjointsadcost[3]);
// Check the starting position
best_sad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
best_sad += mvsad_err_cost(x, bmv, &fcenter_mv.as_mv, sad_per_bit);
*num00 = 0;
for (i = 0, step = 0; step < tot_steps; step++) {
for (j = 0; j < cfg->searches_per_step; j += 4, i += 4) {
__m128i v_sad_d;
__m128i v_cost_d;
__m128i v_outside_d;
__m128i v_inside_d;
__m128i v_diff_mv_w;
#if ARCH_X86_64
__m128i v_blocka[2];
#else
__m128i v_blocka[1];
#endif
// Compute the candidate motion vectors
const __m128i v_ss_mv_w = _mm_loadu_si128((const __m128i*)&ss_mv[i]);
const __m128i v_these_mv_w = _mm_add_epi16(v_bmv_w, v_ss_mv_w);
// Clamp them to the search bounds
__m128i v_these_mv_clamp_w = v_these_mv_w;
v_these_mv_clamp_w = _mm_min_epi16(v_these_mv_clamp_w, v_max_mv_w);
v_these_mv_clamp_w = _mm_max_epi16(v_these_mv_clamp_w, v_min_mv_w);
// The ones that did not change are inside the search area
v_inside_d = _mm_cmpeq_epi32(v_these_mv_clamp_w, v_these_mv_w);
// If none of them are inside, then move on
if (__likely__(_mm_test_all_zeros(v_inside_d, v_inside_d))) {
continue;
}
// The inverse mask indicates which of the MVs are outside
v_outside_d = _mm_xor_si128(v_inside_d, _mm_set1_epi8(0xff));
// Shift right to keep the sign bit clear, we will use this later
// to set the cost to the maximum value.
v_outside_d = _mm_srli_epi32(v_outside_d, 1);
// Compute the difference MV
v_diff_mv_w = _mm_sub_epi16(v_these_mv_clamp_w, vfcmv);
// We utilise the fact that the cost function is even, and use the
// absolute difference. This allows us to use unsigned indexes later
// and reduces cache pressure somewhat as only a half of the table
// is ever referenced.
v_diff_mv_w = _mm_abs_epi16(v_diff_mv_w);
// Compute the SIMD pointer offsets.
{
#if ARCH_X86_64 // sizeof(intptr_t) == 8
// Load the offsets
__m128i v_bo10_q = _mm_loadu_si128((const __m128i*)&ss_os[i+0]);
__m128i v_bo32_q = _mm_loadu_si128((const __m128i*)&ss_os[i+2]);
// Set the ones falling outside to zero
v_bo10_q = _mm_and_si128(v_bo10_q,
_mm_cvtepi32_epi64(v_inside_d));
v_bo32_q = _mm_and_si128(v_bo32_q,
_mm_unpackhi_epi32(v_inside_d, v_inside_d));
// Compute the candidate addresses
v_blocka[0] = _mm_add_epi64(v_ba_q, v_bo10_q);
v_blocka[1] = _mm_add_epi64(v_ba_q, v_bo32_q);
#else // ARCH_X86 // sizeof(intptr_t) == 4
__m128i v_bo_d = _mm_loadu_si128((const __m128i*)&ss_os[i]);
v_bo_d = _mm_and_si128(v_bo_d, v_inside_d);
v_blocka[0] = _mm_add_epi32(v_ba_d, v_bo_d);
#endif
}
fn_ptr->sdx4df(what, what_stride,
(const uint8_t **)&v_blocka[0], in_what_stride,
(uint32_t*)&v_sad_d);
// Look up the component cost of the residual motion vector
{
const int32_t row0 = _mm_extract_epi16(v_diff_mv_w, 0);
const int32_t col0 = _mm_extract_epi16(v_diff_mv_w, 1);
const int32_t row1 = _mm_extract_epi16(v_diff_mv_w, 2);
const int32_t col1 = _mm_extract_epi16(v_diff_mv_w, 3);
const int32_t row2 = _mm_extract_epi16(v_diff_mv_w, 4);
const int32_t col2 = _mm_extract_epi16(v_diff_mv_w, 5);
const int32_t row3 = _mm_extract_epi16(v_diff_mv_w, 6);
const int32_t col3 = _mm_extract_epi16(v_diff_mv_w, 7);
// Note: This is a use case for vpgather in AVX2
const uint32_t cost0 = x->nmvsadcost[0][row0] + x->nmvsadcost[0][col0];
const uint32_t cost1 = x->nmvsadcost[0][row1] + x->nmvsadcost[0][col1];
const uint32_t cost2 = x->nmvsadcost[0][row2] + x->nmvsadcost[0][col2];
const uint32_t cost3 = x->nmvsadcost[0][row3] + x->nmvsadcost[0][col3];
__m128i v_cost_10_d, v_cost_32_d;
v_cost_10_d = _mm_cvtsi32_si128(cost0);
v_cost_10_d = _mm_insert_epi32(v_cost_10_d, cost1, 1);
v_cost_32_d = _mm_cvtsi32_si128(cost2);
v_cost_32_d = _mm_insert_epi32(v_cost_32_d, cost3, 1);
v_cost_d = _mm_unpacklo_epi64(v_cost_10_d, v_cost_32_d);
}
// Now add in the joint cost
{
const __m128i v_sel_d = _mm_cmpeq_epi32(v_diff_mv_w,
_mm_setzero_si128());
const __m128i v_joint_cost_d = _mm_blendv_epi8(v_joint_cost_1_d,
v_joint_cost_0_d,
v_sel_d);
v_cost_d = _mm_add_epi32(v_cost_d, v_joint_cost_d);
}
// Multiply by sad_per_bit
v_cost_d = _mm_mullo_epi32(v_cost_d, v_spb_d);
// ROUND_POWER_OF_TWO(v_cost_d, 8)
v_cost_d = _mm_add_epi32(v_cost_d, _mm_set1_epi32(0x80));
v_cost_d = _mm_srai_epi32(v_cost_d, 8);
// Add the cost to the sad
v_sad_d = _mm_add_epi32(v_sad_d, v_cost_d);
// Make the motion vectors outside the search area have max cost
// by or'ing in the comparison mask, this way the minimum search won't
// pick them.
v_sad_d = _mm_or_si128(v_sad_d, v_outside_d);
// Find the minimum value and index horizontally in v_sad_d
{
// Try speculatively on 16 bits, so we can use the minpos intrinsic
const __m128i v_sad_w = _mm_packus_epi32(v_sad_d, v_sad_d);
const __m128i v_minp_w = _mm_minpos_epu16(v_sad_w);
uint32_t local_best_sad = _mm_extract_epi16(v_minp_w, 0);
uint32_t local_best_idx = _mm_extract_epi16(v_minp_w, 1);
// If the local best value is not saturated, just use it, otherwise
// find the horizontal minimum again the hard way on 32 bits.
// This is executed rarely.
if (__unlikely__(local_best_sad == 0xffff)) {
__m128i v_loval_d, v_hival_d, v_loidx_d, v_hiidx_d, v_sel_d;
v_loval_d = v_sad_d;
v_loidx_d = _mm_set_epi32(3, 2, 1, 0);
v_hival_d = _mm_srli_si128(v_loval_d, 8);
v_hiidx_d = _mm_srli_si128(v_loidx_d, 8);
v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
v_hival_d = _mm_srli_si128(v_loval_d, 4);
v_hiidx_d = _mm_srli_si128(v_loidx_d, 4);
v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
local_best_sad = _mm_extract_epi32(v_loval_d, 0);
local_best_idx = _mm_extract_epi32(v_loidx_d, 0);
}
// Update the global minimum if the local minimum is smaller
if (__likely__(local_best_sad < best_sad)) {
new_bmv = ((const int_mv *)&v_these_mv_w)[local_best_idx];
new_best_address = ((const uint8_t **)v_blocka)[local_best_idx];
best_sad = local_best_sad;
}
}
}
bmv = new_bmv;
best_address = new_best_address;
v_bmv_w = _mm_set1_epi32(bmv.as_int);
#if ARCH_X86_64
v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
#else
v_ba_d = _mm_set1_epi32((intptr_t)best_address);
#endif
if (__unlikely__(best_address == in_what)) {
(*num00)++;
}
}
*best_mv = bmv.as_mv;
return best_sad;
}
// --------------------------------------------------------------
vuint32 mandelbrot_SIMD_I32(vfloat32 a, vfloat32 b, int max_iter)
// --------------------------------------------------------------
{
// version avec test de sortie en int
vuint32 iter = _mm_set1_epi32(0);
vuint32 temp = _mm_set1_epi32(0);
vuint32 un = _mm_set1_epi32(1);
vfloat32 x,y,t,t2,zero,deux,quatre;
// COMPLETER ICI
int test = 0,i = 0;
// initialisation des variables
x = _mm_set_ps(0,0,0,0);
y = _mm_set_ps(0,0,0,0);
deux = _mm_set_ps(2,2,2,2);
quatre = _mm_set_ps(4,4,4,4);
// iteration zero
t = _mm_mul_ps(x, x);
t2 = _mm_mul_ps(y, y);
y = _mm_mul_ps(x,y);
y = _mm_mul_ps(y,deux);
y = _mm_add_ps(y,b);
x = _mm_sub_ps(t,t2);
x = _mm_add_ps(x,a);
// calcul
while(i<max_iter && test ==0 ){
t = _mm_mul_ps(x, x);
t2 = _mm_mul_ps(y, y);
y = _mm_mul_ps(_mm_mul_ps(x,y),deux);
y = _mm_add_ps(y,b);
x = _mm_sub_ps(t,t2);
x = _mm_add_ps(x,a);
t2 = _mm_add_ps(t,t2);
t2 = _mm_cmple_ps(t2,quatre);
temp = _mm_and_si128(un,_mm_castps_si128(t2));
iter = _mm_add_epi32(iter,temp);
test = _mm_test_all_zeros(temp, un);
//display_vuint32(temp,"%d\t","T :: ");
//printf(" MASK::%d \n",_mm_movemask_ps(temp));
i+=1;
}
return iter;
}
//-----------------------------------------------------------------------------------------
// Rasterize the occludee AABB and depth test it against the CPU rasterized depth buffer
// If any of the rasterized AABB pixels passes the depth test exit early and mark the occludee
// as visible. If all rasterized AABB pixels are occluded then the occludee is culled
//-----------------------------------------------------------------------------------------
void TransformedAABBoxSSE::RasterizeAndDepthTestAABBox(UINT *pRenderTargetPixels)
{
// Set DAZ and FZ MXCSR bits to flush denormals to zero (i.e., make it faster)
// Denormal are zero (DAZ) is bit 6 and Flush to zero (FZ) is bit 15.
// so to enable the two to have to set bits 6 and 15 which 1000 0000 0100 0000 = 0x8040
_mm_setcsr( _mm_getcsr() | 0x8040 );
__m128i colOffset = _mm_set_epi32(0, 1, 0, 1);
__m128i rowOffset = _mm_set_epi32(0, 0, 1, 1);
__m128i fxptZero = _mm_setzero_si128();
float* pDepthBuffer = (float*)pRenderTargetPixels;
// Rasterize the AABB triangles 4 at a time
for(UINT i = 0; i < AABB_TRIANGLES; i += SSE)
{
vFloat4 xformedPos[3];
Gather(xformedPos, i);
// use fixed-point only for X and Y. Avoid work for Z and W.
vFxPt4 xFormedFxPtPos[3];
for(int m = 0; m < 3; m++)
{
xFormedFxPtPos[m].X = _mm_cvtps_epi32(xformedPos[m].X);
xFormedFxPtPos[m].Y = _mm_cvtps_epi32(xformedPos[m].Y);
xFormedFxPtPos[m].Z = _mm_cvtps_epi32(xformedPos[m].Z);
xFormedFxPtPos[m].W = _mm_cvtps_epi32(xformedPos[m].W);
}
// Fab(x, y) = Ax + By + C = 0
// Fab(x, y) = (ya - yb)x + (xb - xa)y + (xa * yb - xb * ya) = 0
// Compute A = (ya - yb) for the 3 line segments that make up each triangle
__m128i A0 = _mm_sub_epi32(xFormedFxPtPos[1].Y, xFormedFxPtPos[2].Y);
__m128i A1 = _mm_sub_epi32(xFormedFxPtPos[2].Y, xFormedFxPtPos[0].Y);
__m128i A2 = _mm_sub_epi32(xFormedFxPtPos[0].Y, xFormedFxPtPos[1].Y);
// Compute B = (xb - xa) for the 3 line segments that make up each triangle
__m128i B0 = _mm_sub_epi32(xFormedFxPtPos[2].X, xFormedFxPtPos[1].X);
__m128i B1 = _mm_sub_epi32(xFormedFxPtPos[0].X, xFormedFxPtPos[2].X);
__m128i B2 = _mm_sub_epi32(xFormedFxPtPos[1].X, xFormedFxPtPos[0].X);
// Compute C = (xa * yb - xb * ya) for the 3 line segments that make up each triangle
__m128i C0 = _mm_sub_epi32(_mm_mullo_epi32(xFormedFxPtPos[1].X, xFormedFxPtPos[2].Y), _mm_mullo_epi32(xFormedFxPtPos[2].X, xFormedFxPtPos[1].Y));
__m128i C1 = _mm_sub_epi32(_mm_mullo_epi32(xFormedFxPtPos[2].X, xFormedFxPtPos[0].Y), _mm_mullo_epi32(xFormedFxPtPos[0].X, xFormedFxPtPos[2].Y));
__m128i C2 = _mm_sub_epi32(_mm_mullo_epi32(xFormedFxPtPos[0].X, xFormedFxPtPos[1].Y), _mm_mullo_epi32(xFormedFxPtPos[1].X, xFormedFxPtPos[0].Y));
// Compute triangle area
__m128i triArea = _mm_mullo_epi32(A0, xFormedFxPtPos[0].X);
triArea = _mm_add_epi32(triArea, _mm_mullo_epi32(B0, xFormedFxPtPos[0].Y));
triArea = _mm_add_epi32(triArea, C0);
__m128 oneOverTriArea = _mm_div_ps(_mm_set1_ps(1.0f), _mm_cvtepi32_ps(triArea));
// Use bounding box traversal strategy to determine which pixels to rasterize
__m128i startX = _mm_and_si128(Max(Min(Min(xFormedFxPtPos[0].X, xFormedFxPtPos[1].X), xFormedFxPtPos[2].X), _mm_set1_epi32(0)), _mm_set1_epi32(0xFFFFFFFE));
__m128i endX = Min(_mm_add_epi32(Max(Max(xFormedFxPtPos[0].X, xFormedFxPtPos[1].X), xFormedFxPtPos[2].X), _mm_set1_epi32(1)), _mm_set1_epi32(SCREENW));
__m128i startY = _mm_and_si128(Max(Min(Min(xFormedFxPtPos[0].Y, xFormedFxPtPos[1].Y), xFormedFxPtPos[2].Y), _mm_set1_epi32(0)), _mm_set1_epi32(0xFFFFFFFE));
__m128i endY = Min(_mm_add_epi32(Max(Max(xFormedFxPtPos[0].Y, xFormedFxPtPos[1].Y), xFormedFxPtPos[2].Y), _mm_set1_epi32(1)), _mm_set1_epi32(SCREENH));
for(int vv = 0; vv < 3; vv++)
{
// If W (holding 1/w in our case) is not between 0 and 1,
// then vertex is behind near clip plane (1.0 in our case.
// If W < 1, then verify 1/W > 1 (for W>0), and 1/W < 0 (for W < 0).
__m128 nearClipMask0 = _mm_cmple_ps(xformedPos[vv].W, _mm_set1_ps(0.0f));
__m128 nearClipMask1 = _mm_cmpge_ps(xformedPos[vv].W, _mm_set1_ps(1.0f));
__m128 nearClipMask = _mm_or_ps(nearClipMask0, nearClipMask1);
if(!_mm_test_all_zeros(*(__m128i*)&nearClipMask, *(__m128i*)&nearClipMask))
{
// All four vertices are behind the near plane (we're processing four triangles at a time w/ SSE)
*mVisible = true;
return;
}
}
// Now we have 4 triangles set up. Rasterize them each individually.
for(int lane=0; lane < SSE; lane++)
{
// Skip triangle if area is zero
if(triArea.m128i_i32[lane] <= 0)
{
continue;
}
// Extract this triangle's properties from the SIMD versions
__m128 zz[3], oneOverW[3];
for(int vv = 0; vv < 3; vv++)
{
zz[vv] = _mm_set1_ps(xformedPos[vv].Z.m128_f32[lane]);
oneOverW[vv] = _mm_set1_ps(xformedPos[vv].W.m128_f32[lane]);
}
__m128 oneOverTotalArea = _mm_set1_ps(oneOverTriArea.m128_f32[lane]);
zz[0] *= oneOverTotalArea;
zz[1] *= oneOverTotalArea;
zz[2] *= oneOverTotalArea;
int startXx = startX.m128i_i32[lane];
int endXx = endX.m128i_i32[lane];
int startYy = startY.m128i_i32[lane];
int endYy = endY.m128i_i32[lane];
__m128i aa0 = _mm_set1_epi32(A0.m128i_i32[lane]);
__m128i aa1 = _mm_set1_epi32(A1.m128i_i32[lane]);
__m128i aa2 = _mm_set1_epi32(A2.m128i_i32[lane]);
__m128i bb0 = _mm_set1_epi32(B0.m128i_i32[lane]);
__m128i bb1 = _mm_set1_epi32(B1.m128i_i32[lane]);
__m128i bb2 = _mm_set1_epi32(B2.m128i_i32[lane]);
__m128i cc0 = _mm_set1_epi32(C0.m128i_i32[lane]);
__m128i cc1 = _mm_set1_epi32(C1.m128i_i32[lane]);
__m128i cc2 = _mm_set1_epi32(C2.m128i_i32[lane]);
__m128i aa0Inc = _mm_slli_epi32(aa0, 1);
__m128i aa1Inc = _mm_slli_epi32(aa1, 1);
__m128i aa2Inc = _mm_slli_epi32(aa2, 1);
__m128i row, col;
int rowIdx;
// To avoid this branching, choose one method to traverse and store the pixel depth
if(gVisualizeDepthBuffer)
{
// Sequentially traverse and store pixel depths contiguously
rowIdx = (startYy * SCREENW + startXx);
}
else
{
// Tranverse pixels in 2x2 blocks and store 2x2 pixel quad depths contiguously in memory ==> 2*X
// This method provides better perfromance
rowIdx = (startYy * SCREENW + 2 * startXx);
}
col = _mm_add_epi32(colOffset, _mm_set1_epi32(startXx));
__m128i aa0Col = _mm_mullo_epi32(aa0, col);
__m128i aa1Col = _mm_mullo_epi32(aa1, col);
__m128i aa2Col = _mm_mullo_epi32(aa2, col);
row = _mm_add_epi32(rowOffset, _mm_set1_epi32(startYy));
__m128i bb0Row = _mm_add_epi32(_mm_mullo_epi32(bb0, row), cc0);
__m128i bb1Row = _mm_add_epi32(_mm_mullo_epi32(bb1, row), cc1);
__m128i bb2Row = _mm_add_epi32(_mm_mullo_epi32(bb2, row), cc2);
__m128i bb0Inc = _mm_slli_epi32(bb0, 1);
__m128i bb1Inc = _mm_slli_epi32(bb1, 1);
__m128i bb2Inc = _mm_slli_epi32(bb2, 1);
// Incrementally compute Fab(x, y) for all the pixels inside the bounding box formed by (startX, endX) and (startY, endY)
for(int r = startYy; r < endYy; r += 2,
row = _mm_add_epi32(row, _mm_set1_epi32(2)),
rowIdx = rowIdx + 2 * SCREENW,
bb0Row = _mm_add_epi32(bb0Row, bb0Inc),
bb1Row = _mm_add_epi32(bb1Row, bb1Inc),
bb2Row = _mm_add_epi32(bb2Row, bb2Inc))
{
// Compute barycentric coordinates
int idx = rowIdx;
__m128i alpha = _mm_add_epi32(aa0Col, bb0Row);
__m128i beta = _mm_add_epi32(aa1Col, bb1Row);
__m128i gama = _mm_add_epi32(aa2Col, bb2Row);
int idxIncr;
if(gVisualizeDepthBuffer)
{
idxIncr = 2;
}
else
{
idxIncr = 4;
}
for(int c = startXx; c < endXx; c += 2,
idx = idx + idxIncr,
alpha = _mm_add_epi32(alpha, aa0Inc),
beta = _mm_add_epi32(beta, aa1Inc),
gama = _mm_add_epi32(gama, aa2Inc))
{
//Test Pixel inside triangle
__m128i mask = _mm_cmplt_epi32(fxptZero, _mm_or_si128(_mm_or_si128(alpha, beta), gama));
// Early out if all of this quad's pixels are outside the triangle.
if(_mm_test_all_zeros(mask, mask))
{
continue;
}
// Compute barycentric-interpolated depth
__m128 depth = _mm_mul_ps(_mm_cvtepi32_ps(alpha), zz[0]);
depth = _mm_add_ps(depth, _mm_mul_ps(_mm_cvtepi32_ps(beta), zz[1]));
depth = _mm_add_ps(depth, _mm_mul_ps(_mm_cvtepi32_ps(gama), zz[2]));
__m128 previousDepthValue;
if(gVisualizeDepthBuffer)
{
previousDepthValue = _mm_set_ps(pDepthBuffer[idx], pDepthBuffer[idx + 1], pDepthBuffer[idx + SCREENW], pDepthBuffer[idx + SCREENW + 1]);
}
else
{
previousDepthValue = *(__m128*)&pDepthBuffer[idx];
}
__m128 depthMask = _mm_cmpge_ps( depth, previousDepthValue);
__m128i finalMask = _mm_and_si128( mask, _mm_castps_si128(depthMask));
if(!_mm_test_all_zeros(finalMask, finalMask))
{
*mVisible = true;
return; //early exit
}
}//for each column
}// for each row
}// for each triangle
}// for each set of SIMD# triangles
}
int test_mm_test_all_zeros(__m128i x, __m128i y) {
// CHECK-LABEL: test_mm_test_all_zeros
// CHECK: call i32 @llvm.x86.sse41.ptestz
// CHECK-ASM: ptest %xmm{{.*}}, %xmm{{.*}}
return _mm_test_all_zeros(x, y);
}
int test_mm_test_all_zeros(__m128i x, __m128i y) {
// CHECK-LABEL: test_mm_test_all_zeros
// CHECK: call i32 @llvm.x86.sse41.ptestz(<2 x i64> %{{.*}}, <2 x i64> %{{.*}})
return _mm_test_all_zeros(x, y);
}
