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_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); }
/** * 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)); } }
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); }
int allposinf_double2( vec_double2 x ) { vec_ullong2 posinf = spu_andc( isinfd2 ( x ), signbitd2 ( x ) ); return ( spu_extract(posinf,0) != 0 && spu_extract(posinf,1) != 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; }
/* 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)); }