void float_to_int16_altivec(int16_t *dst, const float *src, int len) { int i; vector float s0, s1; vector signed int t0, t1; vector signed short d0, d1, d; vector unsigned char align; if(((long)dst)&15) //FIXME for(i=0; i<len-7; i+=8) { s0 = vec_ld(0, src+i); s1 = vec_ld(16, src+i); t0 = vec_cts(s0, 0); d0 = vec_ld(0, dst+i); t1 = vec_cts(s1, 0); d1 = vec_ld(15, dst+i); d = vec_packs(t0,t1); d1 = vec_perm(d1, d0, vec_lvsl(0,dst+i)); align = vec_lvsr(0, dst+i); d0 = vec_perm(d1, d, align); d1 = vec_perm(d, d1, align); vec_st(d0, 0, dst+i); vec_st(d1,15, dst+i); } else for(i=0; i<len-7; i+=8) { s0 = vec_ld(0, src+i); s1 = vec_ld(16, src+i); t0 = vec_cts(s0, 0); t1 = vec_cts(s1, 0); d = vec_packs(t0,t1); vec_st(d, 0, dst+i); } }
static vector signed short float_to_int16_one_altivec(const float *src) { vector float s0 = vec_ld(0, src); vector float s1 = vec_ld(16, src); vector signed int t0 = vec_cts(s0, 0); vector signed int t1 = vec_cts(s1, 0); return vec_packs(t0,t1); }
void audio_convert_float_to_s16_altivec(int16_t *out, const float *in, size_t samples) { // Unaligned loads/store is a bit expensive, so we optimize for the good path (very likely). if (((uintptr_t)out & 15) + ((uintptr_t)in & 15) == 0) { size_t i; for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8) { vector float input0 = vec_ld( 0, in); vector float input1 = vec_ld(16, in); vector signed int result0 = vec_cts(input0, 15); vector signed int result1 = vec_cts(input1, 15); vec_st(vec_packs(result0, result1), 0, out); } audio_convert_float_to_s16_C(out, in, samples - i); } else audio_convert_float_to_s16_C(out, in, samples); }
static void ProjectDlightTexture_altivec( void ) { int i, l; vec_t origin0, origin1, origin2; float texCoords0, texCoords1; vector float floatColorVec0, floatColorVec1; vector float modulateVec, colorVec, zero; vector short colorShort; vector signed int colorInt; vector unsigned char floatColorVecPerm, modulatePerm, colorChar; vector unsigned char vSel = VECCONST_UINT8(0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff); float *texCoords; byte *colors; byte clipBits[SHADER_MAX_VERTEXES]; float texCoordsArray[SHADER_MAX_VERTEXES][2]; byte colorArray[SHADER_MAX_VERTEXES][4]; unsigned hitIndexes[SHADER_MAX_INDEXES]; int numIndexes; float scale; float radius; vec3_t floatColor; float modulate = 0.0f; if ( !backEnd.refdef.num_dlights ) { return; } // There has to be a better way to do this so that floatColor // and/or modulate are already 16-byte aligned. floatColorVecPerm = vec_lvsl(0,(float *)floatColor); modulatePerm = vec_lvsl(0,(float *)&modulate); modulatePerm = (vector unsigned char)vec_splat((vector unsigned int)modulatePerm,0); zero = (vector float)vec_splat_s8(0); for ( l = 0 ; l < backEnd.refdef.num_dlights ; l++ ) { dlight_t *dl; if ( !( tess.dlightBits & ( 1 << l ) ) ) { continue; // this surface definately doesn't have any of this light } texCoords = texCoordsArray[0]; colors = colorArray[0]; dl = &backEnd.refdef.dlights[l]; origin0 = dl->transformed[0]; origin1 = dl->transformed[1]; origin2 = dl->transformed[2]; radius = dl->radius; scale = 1.0f / radius; if(r_greyscale->integer) { float luminance; luminance = LUMA(dl->color[0], dl->color[1], dl->color[2]) * 255.0f; floatColor[0] = floatColor[1] = floatColor[2] = luminance; } else if(r_greyscale->value) { float luminance; luminance = LUMA(dl->color[0], dl->color[1], dl->color[2]) * 255.0f; floatColor[0] = LERP(dl->color[0] * 255.0f, luminance, r_greyscale->value); floatColor[1] = LERP(dl->color[1] * 255.0f, luminance, r_greyscale->value); floatColor[2] = LERP(dl->color[2] * 255.0f, luminance, r_greyscale->value); } else { floatColor[0] = dl->color[0] * 255.0f; floatColor[1] = dl->color[1] * 255.0f; floatColor[2] = dl->color[2] * 255.0f; } floatColorVec0 = vec_ld(0, floatColor); floatColorVec1 = vec_ld(11, floatColor); floatColorVec0 = vec_perm(floatColorVec0,floatColorVec0,floatColorVecPerm); for ( i = 0 ; i < tess.numVertexes ; i++, texCoords += 2, colors += 4 ) { int clip = 0; vec_t dist0, dist1, dist2; dist0 = origin0 - tess.xyz[i][0]; dist1 = origin1 - tess.xyz[i][1]; dist2 = origin2 - tess.xyz[i][2]; backEnd.pc.c_dlightVertexes++; texCoords0 = 0.5f + dist0 * scale; texCoords1 = 0.5f + dist1 * scale; if( !r_dlightBacks->integer && // dist . tess.normal[i] ( dist0 * tess.normal[i][0] + dist1 * tess.normal[i][1] + dist2 * tess.normal[i][2] ) < 0.0f ) { clip = 63; } else { if ( texCoords0 < 0.0f ) { clip |= 1; } else if ( texCoords0 > 1.0f ) { clip |= 2; } if ( texCoords1 < 0.0f ) { clip |= 4; } else if ( texCoords1 > 1.0f ) { clip |= 8; } texCoords[0] = texCoords0; texCoords[1] = texCoords1; // modulate the strength based on the height and color if ( dist2 > radius ) { clip |= 16; modulate = 0.0f; } else if ( dist2 < -radius ) { clip |= 32; modulate = 0.0f; } else { dist2 = Q_fabs(dist2); if ( dist2 < radius * 0.5f ) { modulate = 1.0f; } else { modulate = 2.0f * (radius - dist2) * scale; } } } clipBits[i] = clip; modulateVec = vec_ld(0,(float *)&modulate); modulateVec = vec_perm(modulateVec,modulateVec,modulatePerm); colorVec = vec_madd(floatColorVec0,modulateVec,zero); colorInt = vec_cts(colorVec,0); // RGBx colorShort = vec_pack(colorInt,colorInt); // RGBxRGBx colorChar = vec_packsu(colorShort,colorShort); // RGBxRGBxRGBxRGBx colorChar = vec_sel(colorChar,vSel,vSel); // RGBARGBARGBARGBA replace alpha with 255 vec_ste((vector unsigned int)colorChar,0,(unsigned int *)colors); // store color } // build a list of triangles that need light numIndexes = 0; for ( i = 0 ; i < tess.numIndexes ; i += 3 ) { int a, b, c; a = tess.indexes[i]; b = tess.indexes[i+1]; c = tess.indexes[i+2]; if ( clipBits[a] & clipBits[b] & clipBits[c] ) { continue; // not lighted } hitIndexes[numIndexes] = a; hitIndexes[numIndexes+1] = b; hitIndexes[numIndexes+2] = c; numIndexes += 3; } if ( !numIndexes ) { continue; } qglEnableClientState( GL_TEXTURE_COORD_ARRAY ); qglTexCoordPointer( 2, GL_FLOAT, 0, texCoordsArray[0] ); qglEnableClientState( GL_COLOR_ARRAY ); qglColorPointer( 4, GL_UNSIGNED_BYTE, 0, colorArray ); GL_Bind( tr.dlightImage ); // include GLS_DEPTHFUNC_EQUAL so alpha tested surfaces don't add light // where they aren't rendered if ( dl->additive ) { GL_State( GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE | GLS_DEPTHFUNC_EQUAL ); } else { GL_State( GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ONE | GLS_DEPTHFUNC_EQUAL ); } R_DrawElements( numIndexes, hitIndexes ); backEnd.pc.c_totalIndexes += numIndexes; backEnd.pc.c_dlightIndexes += numIndexes; } }
int main () { vector float fa = {1.0, 2.0, 3.0, -4.0}; vector float fb = {-2.0, -3.0, -4.0, -5.0}; vector float fc = vec_cpsgn (fa, fb); vector long long la = {5L, 14L}; vector long long lb = {3L, 86L}; vector long long lc = vec_and (la, lb); vector bool long long ld = {0, -1}; vector long long le = vec_and (la, ld); vector long long lf = vec_and (ld, lb); vector unsigned long long ua = {5L, 14L}; vector unsigned long long ub = {3L, 86L}; vector unsigned long long uc = vec_and (ua, ub); vector bool long long ud = {0, -1}; vector unsigned long long ue = vec_and (ua, ud); vector unsigned long long uf = vec_and (ud, ub); vector long long lg = vec_andc (la, lb); vector long long lh = vec_andc (la, ld); vector long long li = vec_andc (ld, lb); vector unsigned long long ug = vec_andc (ua, ub); vector unsigned long long uh = vec_andc (ua, ud); vector unsigned long long ui = vec_andc (ud, ub); vector double da = {1.0, -4.0}; vector double db = {-2.0, 5.0}; vector double dc = vec_cpsgn (da, db); vector long long lj = vec_mergeh (la, lb); vector long long lk = vec_mergeh (la, ld); vector long long ll = vec_mergeh (ld, la); vector unsigned long long uj = vec_mergeh (ua, ub); vector unsigned long long uk = vec_mergeh (ua, ud); vector unsigned long long ul = vec_mergeh (ud, ua); vector long long lm = vec_mergel (la, lb); vector long long ln = vec_mergel (la, ld); vector long long lo = vec_mergel (ld, la); vector unsigned long long um = vec_mergel (ua, ub); vector unsigned long long un = vec_mergel (ua, ud); vector unsigned long long uo = vec_mergel (ud, ua); vector long long lp = vec_nor (la, lb); vector long long lq = vec_nor (la, ld); vector long long lr = vec_nor (ld, la); vector unsigned long long up = vec_nor (ua, ub); vector unsigned long long uq = vec_nor (ua, ud); vector unsigned long long ur = vec_nor (ud, ua); vector long long ls = vec_or (la, lb); vector long long lt = vec_or (la, ld); vector long long lu = vec_or (ld, la); vector unsigned long long us = vec_or (ua, ub); vector unsigned long long ut = vec_or (ua, ud); vector unsigned long long uu = vec_or (ud, ua); vector unsigned char ca = {0,4,8,1,5,9,2,6,10,3,7,11,15,12,14,13}; vector long long lv = vec_perm (la, lb, ca); vector unsigned long long uv = vec_perm (ua, ub, ca); vector long long lw = vec_sel (la, lb, lc); vector long long lx = vec_sel (la, lb, uc); vector long long ly = vec_sel (la, lb, ld); vector unsigned long long uw = vec_sel (ua, ub, lc); vector unsigned long long ux = vec_sel (ua, ub, uc); vector unsigned long long uy = vec_sel (ua, ub, ld); vector long long lz = vec_xor (la, lb); vector long long l0 = vec_xor (la, ld); vector long long l1 = vec_xor (ld, la); vector unsigned long long uz = vec_xor (ua, ub); vector unsigned long long u0 = vec_xor (ua, ud); vector unsigned long long u1 = vec_xor (ud, ua); int ia = vec_all_eq (ua, ub); int ib = vec_all_ge (ua, ub); int ic = vec_all_gt (ua, ub); int id = vec_all_le (ua, ub); int ie = vec_all_lt (ua, ub); int ig = vec_all_ne (ua, ub); int ih = vec_any_eq (ua, ub); int ii = vec_any_ge (ua, ub); int ij = vec_any_gt (ua, ub); int ik = vec_any_le (ua, ub); int il = vec_any_lt (ua, ub); int im = vec_any_ne (ua, ub); vector int sia = {9, 16, 25, 36}; vector int sib = {-8, -27, -64, -125}; vector int sic = vec_mergee (sia, sib); vector int sid = vec_mergeo (sia, sib); vector unsigned int uia = {9, 16, 25, 36}; vector unsigned int uib = {8, 27, 64, 125}; vector unsigned int uic = vec_mergee (uia, uib); vector unsigned int uid = vec_mergeo (uia, uib); vector bool int bia = {0, -1, -1, 0}; vector bool int bib = {-1, -1, 0, -1}; vector bool int bic = vec_mergee (bia, bib); vector bool int bid = vec_mergeo (bia, bib); vector unsigned int uie = vec_packsu (ua, ub); vector long long l2 = vec_cntlz (la); vector unsigned long long u2 = vec_cntlz (ua); vector int sie = vec_cntlz (sia); vector unsigned int uif = vec_cntlz (uia); vector short ssa = {20, -40, -60, 80, 100, -120, -140, 160}; vector short ssb = vec_cntlz (ssa); vector unsigned short usa = {81, 72, 63, 54, 45, 36, 27, 18}; vector unsigned short usb = vec_cntlz (usa); vector signed char sca = {-4, 3, -9, 15, -31, 31, 0, 0, 1, 117, -36, 99, 98, 97, 96, 95}; vector signed char scb = vec_cntlz (sca); vector unsigned char cb = vec_cntlz (ca); vector double dd = vec_xl (0, &y); vec_xst (dd, 0, &z); vector double de = vec_round (dd); vector double df = vec_splat (de, 0); vector double dg = vec_splat (de, 1); vector long long l3 = vec_splat (l2, 0); vector long long l4 = vec_splat (l2, 1); vector unsigned long long u3 = vec_splat (u2, 0); vector unsigned long long u4 = vec_splat (u2, 1); vector bool long long l5 = vec_splat (ld, 0); vector bool long long l6 = vec_splat (ld, 1); vector long long l7 = vec_div (l3, l4); vector unsigned long long u5 = vec_div (u3, u4); vector long long l8 = vec_mul (l3, l4); vector unsigned long long u6 = vec_mul (u3, u4); vector double dh = vec_ctf (la, -2); vector double di = vec_ctf (ua, 2); vector long long l9 = vec_cts (dh, -2); vector unsigned long long u7 = vec_ctu (di, 2); return 0; }
void test1() { // CHECK-LABEL: define void @test1 // CHECK-LE-LABEL: define void @test1 res_vf = vec_abs(vf); // CHECK: call <4 x float> @llvm.fabs.v4f32(<4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x float> @llvm.fabs.v4f32(<4 x float> %{{[0-9]*}}) dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_add(vd, vd); // CHECK: fadd <2 x double> // CHECK-LE: fadd <2 x double> res_vd = vec_and(vbll, vd); // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> // CHECK-LE: and <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> res_vd = vec_and(vd, vbll); // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> // CHECK-LE: and <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> res_vd = vec_and(vd, vd); // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> // CHECK-LE: and <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_andc(vbll, vd); // CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64> // CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1> // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> // CHECK-LE: bitcast <2 x double> %{{[0-9]*}} to <2 x i64> // CHECK-LE: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1> // CHECK-LE: and <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_andc(vd, vbll); // CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64> // CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1> // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> // CHECK-LE: bitcast <2 x double> %{{[0-9]*}} to <2 x i64> // CHECK-LE: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1> // CHECK-LE: and <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> dummy(); // CHECK: call void @dummy() res_vd = vec_andc(vd, vd); // CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64> // CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1> // CHECK: and <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_ceil(vd); // CHECK: call <2 x double> @llvm.ceil.v2f64(<2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x double> @llvm.ceil.v2f64(<2 x double> %{{[0-9]*}}) res_vf = vec_ceil(vf); // CHECK: call <4 x float> @llvm.ceil.v4f32(<4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x float> @llvm.ceil.v4f32(<4 x float> %{{[0-9]*}}) res_vbll = vec_cmpeq(vd, vd); // CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpeqdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpeqdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) res_vbi = vec_cmpeq(vf, vf); // CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpeqsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpeqsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) res_vbll = vec_cmpge(vd, vd); // CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) res_vbi = vec_cmpge(vf, vf); // CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) res_vbll = vec_cmpgt(vd, vd); // CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) res_vbi = vec_cmpgt(vf, vf); // CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) res_vbll = vec_cmple(vd, vd); // CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) res_vbi = vec_cmple(vf, vf); // CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) res_vbll = vec_cmplt(vd, vd); // CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) // CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}}) res_vbi = vec_cmplt(vf, vf); // CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) // CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}}) /* vec_cpsgn */ res_vf = vec_cpsgn(vf, vf); // CHECK: call <4 x float> @llvm.copysign.v4f32(<4 x float> %{{.+}}, <4 x float> %{{.+}}) // CHECK-LE: call <4 x float> @llvm.copysign.v4f32(<4 x float> %{{.+}}, <4 x float> %{{.+}}) res_vd = vec_cpsgn(vd, vd); // CHECK: call <2 x double> @llvm.copysign.v2f64(<2 x double> %{{.+}}, <2 x double> %{{.+}}) // CHECK-LE: call <2 x double> @llvm.copysign.v2f64(<2 x double> %{{.+}}, <2 x double> %{{.+}}) /* vec_div */ res_vsll = vec_div(vsll, vsll); // CHECK: sdiv <2 x i64> // CHECK-LE: sdiv <2 x i64> res_vull = vec_div(vull, vull); // CHECK: udiv <2 x i64> // CHECK-LE: udiv <2 x i64> res_vf = vec_div(vf, vf); // CHECK: fdiv <4 x float> // CHECK-LE: fdiv <4 x float> res_vd = vec_div(vd, vd); // CHECK: fdiv <2 x double> // CHECK-LE: fdiv <2 x double> /* vec_max */ res_vf = vec_max(vf, vf); // CHECK: @llvm.ppc.vsx.xvmaxsp // CHECK-LE: @llvm.ppc.vsx.xvmaxsp res_vd = vec_max(vd, vd); // CHECK: @llvm.ppc.vsx.xvmaxdp // CHECK-LE: @llvm.ppc.vsx.xvmaxdp res_vf = vec_vmaxfp(vf, vf); // CHECK: @llvm.ppc.vsx.xvmaxsp // CHECK-LE: @llvm.ppc.vsx.xvmaxsp /* vec_min */ res_vf = vec_min(vf, vf); // CHECK: @llvm.ppc.vsx.xvminsp // CHECK-LE: @llvm.ppc.vsx.xvminsp res_vd = vec_min(vd, vd); // CHECK: @llvm.ppc.vsx.xvmindp // CHECK-LE: @llvm.ppc.vsx.xvmindp res_vf = vec_vminfp(vf, vf); // CHECK: @llvm.ppc.vsx.xvminsp // CHECK-LE: @llvm.ppc.vsx.xvminsp res_d = __builtin_vsx_xsmaxdp(d, d); // CHECK: @llvm.ppc.vsx.xsmaxdp // CHECK-LE: @llvm.ppc.vsx.xsmaxdp res_d = __builtin_vsx_xsmindp(d, d); // CHECK: @llvm.ppc.vsx.xsmindp // CHECK-LE: @llvm.ppc.vsx.xsmindp /* vec_perm */ res_vsll = vec_perm(vsll, vsll, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_perm(vull, vull, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vbll = vec_perm(vbll, vbll, vuc); // CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> // CHECK-LE: xor <16 x i8> // CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> res_vf = vec_round(vf); // CHECK: call <4 x float> @llvm.round.v4f32(<4 x float> // CHECK-LE: call <4 x float> @llvm.round.v4f32(<4 x float> res_vd = vec_round(vd); // CHECK: call <2 x double> @llvm.round.v2f64(<2 x double> // CHECK-LE: call <2 x double> @llvm.round.v2f64(<2 x double> res_vd = vec_perm(vd, vd, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vd = vec_splat(vd, 1); // CHECK: [[T1:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32> // CHECK: [[T2:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32> // CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> // CHECK-LE: xor <16 x i8> // CHECK-LE: [[T1:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32> // CHECK-LE: [[T2:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32> // CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> res_vbll = vec_splat(vbll, 1); // CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> // CHECK-LE: xor <16 x i8> // CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> res_vsll = vec_splat(vsll, 1); // CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> // CHECK-LE: xor <16 x i8> // CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> res_vull = vec_splat(vull, 1); // CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> // CHECK-LE: xor <16 x i8> // CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32> // CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8> res_vsi = vec_pack(vsll, vsll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vui = vec_pack(vull, vull); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vbi = vec_pack(vbll, vbll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vsll = vec_vperm(vsll, vsll, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_vperm(vull, vull, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vd = vec_vperm(vd, vd, vuc); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm /* vec_vsx_ld */ res_vsi = vec_vsx_ld(0, &vsi); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vui = vec_vsx_ld(0, &vui); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vf = vec_vsx_ld (0, &vf); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vsll = vec_vsx_ld(0, &vsll); // CHECK: @llvm.ppc.vsx.lxvd2x // CHECK-LE: @llvm.ppc.vsx.lxvd2x res_vull = vec_vsx_ld(0, &vull); // CHECK: @llvm.ppc.vsx.lxvd2x // CHECK-LE: @llvm.ppc.vsx.lxvd2x res_vd = vec_vsx_ld(0, &vd); // CHECK: @llvm.ppc.vsx.lxvd2x // CHECK-LE: @llvm.ppc.vsx.lxvd2x res_vull = vec_vsx_ld(0, &vull); // CHECK: @llvm.ppc.vsx.lxvd2x // CHECK-LE: @llvm.ppc.vsx.lxvd2x res_vd = vec_vsx_ld(0, &vd); // CHECK: @llvm.ppc.vsx.lxvd2x // CHECK-LE: @llvm.ppc.vsx.lxvd2x res_vss = vec_vsx_ld(0, &vss); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vss = vec_vsx_ld(0, &ss); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vus = vec_vsx_ld(0, &vus); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vus = vec_vsx_ld(0, &us); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vbc = vec_vsx_ld(0, &vbc); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vsc = vec_vsx_ld(0, &vsc); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vuc = vec_vsx_ld(0, &vuc); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vsc = vec_vsx_ld(0, &sc); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x res_vuc = vec_vsx_ld(0, &uc); // CHECK: @llvm.ppc.vsx.lxvw4x // CHECK-LE: @llvm.ppc.vsx.lxvw4x /* vec_vsx_st */ vec_vsx_st(vsi, 0, &res_vsi); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vsi, 0, &res_si); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vui, 0, &res_vui); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vui, 0, &res_ui); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vf, 0, &res_vf); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vsll, 0, &res_vsll); // CHECK: @llvm.ppc.vsx.stxvd2x // CHECK-LE: @llvm.ppc.vsx.stxvd2x vec_vsx_st(vull, 0, &res_vull); // CHECK: @llvm.ppc.vsx.stxvd2x // CHECK-LE: @llvm.ppc.vsx.stxvd2x vec_vsx_st(vd, 0, &res_vd); // CHECK: @llvm.ppc.vsx.stxvd2x // CHECK-LE: @llvm.ppc.vsx.stxvd2x vec_vsx_st(vss, 0, &res_vss); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vss, 0, &res_ss); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vus, 0, &res_vus); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vus, 0, &res_us); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vsc, 0, &res_vsc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vsc, 0, &res_sc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vuc, 0, &res_vuc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vuc, 0, &res_uc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vbc, 0, &res_vbc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vbc, 0, &res_sc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x vec_vsx_st(vbc, 0, &res_uc); // CHECK: @llvm.ppc.vsx.stxvw4x // CHECK-LE: @llvm.ppc.vsx.stxvw4x /* vec_and */ res_vsll = vec_and(vsll, vsll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_and(vbll, vsll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_and(vsll, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_and(vull, vull); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_and(vbll, vull); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_and(vull, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vbll = vec_and(vbll, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> /* vec_vand */ res_vsll = vec_vand(vsll, vsll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_vand(vbll, vsll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_vand(vsll, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_vand(vull, vull); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_vand(vbll, vull); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_vand(vull, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> res_vbll = vec_vand(vbll, vbll); // CHECK: and <2 x i64> // CHECK-LE: and <2 x i64> /* vec_andc */ res_vsll = vec_andc(vsll, vsll); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_andc(vbll, vsll); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vsll = vec_andc(vsll, vbll); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_andc(vull, vull); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_andc(vbll, vull); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vull = vec_andc(vull, vbll); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vbll = vec_andc(vbll, vbll); // CHECK: xor <2 x i64> // CHECK: and <2 x i64> // CHECK-LE: xor <2 x i64> // CHECK-LE: and <2 x i64> res_vf = vec_floor(vf); // CHECK: call <4 x float> @llvm.floor.v4f32(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.floor.v4f32(<4 x float> %{{[0-9]+}}) res_vd = vec_floor(vd); // CHECK: call <2 x double> @llvm.floor.v2f64(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.floor.v2f64(<2 x double> %{{[0-9]+}}) res_vf = vec_madd(vf, vf, vf); // CHECK: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}) res_vd = vec_madd(vd, vd, vd); // CHECK: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}) /* vec_mergeh */ res_vsll = vec_mergeh(vsll, vsll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vsll = vec_mergeh(vsll, vbll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vsll = vec_mergeh(vbll, vsll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergeh(vull, vull); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergeh(vull, vbll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergeh(vbll, vull); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm /* vec_mergel */ res_vsll = vec_mergel(vsll, vsll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vsll = vec_mergel(vsll, vbll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vsll = vec_mergel(vbll, vsll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergel(vull, vull); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergel(vull, vbll); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm res_vull = vec_mergel(vbll, vull); // CHECK: @llvm.ppc.altivec.vperm // CHECK-LE: @llvm.ppc.altivec.vperm /* vec_msub */ res_vf = vec_msub(vf, vf, vf); // CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} // CHECK-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> // CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} // CHECK-LE-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> res_vd = vec_msub(vd, vd, vd); // CHECK: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}} // CHECK-NEXT: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> // CHECK-LE: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}} // CHECK-LE-NEXT: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> res_vsll = vec_mul(vsll, vsll); // CHECK: mul <2 x i64> // CHECK-LE: mul <2 x i64> res_vull = vec_mul(vull, vull); // CHECK: mul <2 x i64> // CHECK-LE: mul <2 x i64> res_vf = vec_mul(vf, vf); // CHECK: fmul <4 x float> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: fmul <4 x float> %{{[0-9]+}}, %{{[0-9]+}} res_vd = vec_mul(vd, vd); // CHECK: fmul <2 x double> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: fmul <2 x double> %{{[0-9]+}}, %{{[0-9]+}} res_vf = vec_nearbyint(vf); // CHECK: call <4 x float> @llvm.round.v4f32(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.round.v4f32(<4 x float> %{{[0-9]+}}) res_vd = vec_nearbyint(vd); // CHECK: call <2 x double> @llvm.round.v2f64(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.round.v2f64(<2 x double> %{{[0-9]+}}) res_vf = vec_nmadd(vf, vf, vf); // CHECK: [[FM:[0-9]+]] = call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}) // CHECK-NEXT: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %[[FM]] // CHECK-LE: [[FM:[0-9]+]] = call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}) // CHECK-LE-NEXT: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %[[FM]] res_vd = vec_nmadd(vd, vd, vd); // CHECK: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}) // CHECK-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]] // CHECK-LE: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}) // CHECK-LE-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]] res_vf = vec_nmsub(vf, vf, vf); // CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} // CHECK-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> // CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} // CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} // CHECK-LE-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> // CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}} res_vd = vec_nmsub(vd, vd, vd); // CHECK: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}} // CHECK-NEXT: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> // CHECK-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]] // CHECK-LE: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}} // CHECK-LE-NEXT: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> // CHECK-LE-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]] /* vec_nor */ res_vsll = vec_nor(vsll, vsll); // CHECK: or <2 x i64> // CHECK: xor <2 x i64> // CHECK-LE: or <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_nor(vull, vull); // CHECK: or <2 x i64> // CHECK: xor <2 x i64> // CHECK-LE: or <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_nor(vbll, vbll); // CHECK: or <2 x i64> // CHECK: xor <2 x i64> // CHECK-LE: or <2 x i64> // CHECK-LE: xor <2 x i64> res_vd = vec_nor(vd, vd); // CHECK: bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK: [[OR:%.+]] = or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-NEXT: xor <2 x i64> [[OR]], <i64 -1, i64 -1> // CHECK-LE: bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK-LE: [[OR:%.+]] = or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE-NEXT: xor <2 x i64> [[OR]], <i64 -1, i64 -1> /* vec_or */ res_vsll = vec_or(vsll, vsll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vsll = vec_or(vbll, vsll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vsll = vec_or(vsll, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_or(vull, vull); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_or(vbll, vull); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_or(vull, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vbll = vec_or(vbll, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vd = vec_or(vd, vd); // CHECK: bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK: or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK-LE: or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} res_vd = vec_or(vbll, vd); // CHECK: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK: [[T2:%.+]] = or <2 x i64> %{{[0-9]+}}, [[T1]] // CHECK: bitcast <2 x i64> [[T2]] to <2 x double> // CHECK-LE: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK-LE: [[T2:%.+]] = or <2 x i64> %{{[0-9]+}}, [[T1]] // CHECK-LE: bitcast <2 x i64> [[T2]] to <2 x double> res_vd = vec_or(vd, vbll); // CHECK: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK: [[T2:%.+]] = or <2 x i64> [[T1]], %{{[0-9]+}} // CHECK: bitcast <2 x i64> [[T2]] to <2 x double> // CHECK-LE: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64> // CHECK-LE: [[T2:%.+]] = or <2 x i64> [[T1]], %{{[0-9]+}} // CHECK-LE: bitcast <2 x i64> [[T2]] to <2 x double> res_vf = vec_re(vf); // CHECK: call <4 x float> @llvm.ppc.vsx.xvresp(<4 x float> // CHECK-LE: call <4 x float> @llvm.ppc.vsx.xvresp(<4 x float> res_vd = vec_re(vd); // CHECK: call <2 x double> @llvm.ppc.vsx.xvredp(<2 x double> // CHECK-LE: call <2 x double> @llvm.ppc.vsx.xvredp(<2 x double> res_vf = vec_rint(vf); // CHECK: call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %{{[0-9]+}}) res_vd = vec_rint(vd); // CHECK: call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %{{[0-9]+}}) res_vf = vec_rsqrte(vf); // CHECK: call <4 x float> @llvm.ppc.vsx.xvrsqrtesp(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.ppc.vsx.xvrsqrtesp(<4 x float> %{{[0-9]+}}) res_vd = vec_rsqrte(vd); // CHECK: call <2 x double> @llvm.ppc.vsx.xvrsqrtedp(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.ppc.vsx.xvrsqrtedp(<2 x double> %{{[0-9]+}}) dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vf = vec_sel(vd, vd, vbll); // CHECK: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1> // CHECK: and <2 x i64> %{{[0-9]+}}, // CHECK: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK: or <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]+}} to <2 x double> // CHECK-LE: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1> // CHECK-LE: and <2 x i64> %{{[0-9]+}}, // CHECK-LE: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: or <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]+}} to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_sel(vd, vd, vull); // CHECK: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1> // CHECK: and <2 x i64> %{{[0-9]+}}, // CHECK: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK: or <2 x i64> // CHECK: bitcast <2 x i64> %{{[0-9]+}} to <2 x double> // CHECK-LE: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1> // CHECK-LE: and <2 x i64> %{{[0-9]+}}, // CHECK-LE: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: or <2 x i64> // CHECK-LE: bitcast <2 x i64> %{{[0-9]+}} to <2 x double> res_vf = vec_sqrt(vf); // CHECK: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{[0-9]+}}) res_vd = vec_sqrt(vd); // CHECK: call <2 x double> @llvm.sqrt.v2f64(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.sqrt.v2f64(<2 x double> %{{[0-9]+}}) res_vd = vec_sub(vd, vd); // CHECK: fsub <2 x double> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: fsub <2 x double> %{{[0-9]+}}, %{{[0-9]+}} res_vf = vec_trunc(vf); // CHECK: call <4 x float> @llvm.trunc.v4f32(<4 x float> %{{[0-9]+}}) // CHECK-LE: call <4 x float> @llvm.trunc.v4f32(<4 x float> %{{[0-9]+}}) res_vd = vec_trunc(vd); // CHECK: call <2 x double> @llvm.trunc.v2f64(<2 x double> %{{[0-9]+}}) // CHECK-LE: call <2 x double> @llvm.trunc.v2f64(<2 x double> %{{[0-9]+}}) /* vec_vor */ res_vsll = vec_vor(vsll, vsll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vsll = vec_vor(vbll, vsll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vsll = vec_vor(vsll, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_vor(vull, vull); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_vor(vbll, vull); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vull = vec_vor(vull, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> res_vbll = vec_vor(vbll, vbll); // CHECK: or <2 x i64> // CHECK-LE: or <2 x i64> /* vec_xor */ res_vsll = vec_xor(vsll, vsll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vsll = vec_xor(vbll, vsll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vsll = vec_xor(vsll, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_xor(vull, vull); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_xor(vbll, vull); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_xor(vull, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vbll = vec_xor(vbll, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_xor(vd, vd); // CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK: bitcast <2 x i64> [[X1]] to <2 x double> // CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_xor(vd, vbll); // CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK: bitcast <2 x i64> [[X1]] to <2 x double> // CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double> dummy(); // CHECK: call void @dummy() // CHECK-LE: call void @dummy() res_vd = vec_xor(vbll, vd); // CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK: bitcast <2 x i64> [[X1]] to <2 x double> // CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}} // CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double> /* vec_vxor */ res_vsll = vec_vxor(vsll, vsll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vsll = vec_vxor(vbll, vsll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vsll = vec_vxor(vsll, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_vxor(vull, vull); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_vxor(vbll, vull); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vull = vec_vxor(vull, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vbll = vec_vxor(vbll, vbll); // CHECK: xor <2 x i64> // CHECK-LE: xor <2 x i64> res_vsll = vec_cts(vd, 0); // CHECK: fmul <2 x double> // CHECK: fptosi <2 x double> %{{.*}} to <2 x i64> // CHECK-LE: fmul <2 x double> // CHECK-LE: fptosi <2 x double> %{{.*}} to <2 x i64> res_vsll = vec_cts(vd, 31); // CHECK: fmul <2 x double> // CHECK: fptosi <2 x double> %{{.*}} to <2 x i64> // CHECK-LE: fmul <2 x double> // CHECK-LE: fptosi <2 x double> %{{.*}} to <2 x i64> res_vsll = vec_ctu(vd, 0); // CHECK: fmul <2 x double> // CHECK: fptoui <2 x double> %{{.*}} to <2 x i64> // CHECK-LE: fmul <2 x double> // CHECK-LE: fptoui <2 x double> %{{.*}} to <2 x i64> res_vsll = vec_ctu(vd, 31); // CHECK: fmul <2 x double> // CHECK: fptoui <2 x double> %{{.*}} to <2 x i64> // CHECK-LE: fmul <2 x double> // CHECK-LE: fptoui <2 x double> %{{.*}} to <2 x i64> res_vd = vec_ctf(vsll, 0); // CHECK: sitofp <2 x i64> %{{.*}} to <2 x double> // CHECK: fmul <2 x double> // CHECK-LE: sitofp <2 x i64> %{{.*}} to <2 x double> // CHECK-LE: fmul <2 x double> res_vd = vec_ctf(vsll, 31); // CHECK: sitofp <2 x i64> %{{.*}} to <2 x double> // CHECK: fmul <2 x double> // CHECK-LE: sitofp <2 x i64> %{{.*}} to <2 x double> // CHECK-LE: fmul <2 x double> res_vd = vec_ctf(vull, 0); // CHECK: uitofp <2 x i64> %{{.*}} to <2 x double> // CHECK: fmul <2 x double> // CHECK-LE: uitofp <2 x i64> %{{.*}} to <2 x double> // CHECK-LE: fmul <2 x double> res_vd = vec_ctf(vull, 31); // CHECK: uitofp <2 x i64> %{{.*}} to <2 x double> // CHECK: fmul <2 x double> // CHECK-LE: uitofp <2 x i64> %{{.*}} to <2 x double> // CHECK-LE: fmul <2 x double> }
static int dct_quantize_altivec(MpegEncContext* s, DCTELEM* data, int n, int qscale, int* overflow) { int lastNonZero; vector float row0, row1, row2, row3, row4, row5, row6, row7; vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; const vector float zero = (const vector float)FOUROF(0.); // used after quantize step int oldBaseValue = 0; // Load the data into the row/alt vectors { vector signed short data0, data1, data2, data3, data4, data5, data6, data7; data0 = vec_ld(0, data); data1 = vec_ld(16, data); data2 = vec_ld(32, data); data3 = vec_ld(48, data); data4 = vec_ld(64, data); data5 = vec_ld(80, data); data6 = vec_ld(96, data); data7 = vec_ld(112, data); // Transpose the data before we start TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); // load the data into floating point vectors. We load // the high half of each row into the main row vectors // and the low half into the alt vectors. row0 = vec_ctf(vec_unpackh(data0), 0); alt0 = vec_ctf(vec_unpackl(data0), 0); row1 = vec_ctf(vec_unpackh(data1), 0); alt1 = vec_ctf(vec_unpackl(data1), 0); row2 = vec_ctf(vec_unpackh(data2), 0); alt2 = vec_ctf(vec_unpackl(data2), 0); row3 = vec_ctf(vec_unpackh(data3), 0); alt3 = vec_ctf(vec_unpackl(data3), 0); row4 = vec_ctf(vec_unpackh(data4), 0); alt4 = vec_ctf(vec_unpackl(data4), 0); row5 = vec_ctf(vec_unpackh(data5), 0); alt5 = vec_ctf(vec_unpackl(data5), 0); row6 = vec_ctf(vec_unpackh(data6), 0); alt6 = vec_ctf(vec_unpackl(data6), 0); row7 = vec_ctf(vec_unpackh(data7), 0); alt7 = vec_ctf(vec_unpackl(data7), 0); } // The following block could exist as a separate an altivec dct // function. However, if we put it inline, the DCT data can remain // in the vector local variables, as floats, which we'll use during the // quantize step... { const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f); const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f); const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f); const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f); const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f); const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f); const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f); const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f); const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f); const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f); const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f); const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f); int whichPass, whichHalf; for(whichPass = 1; whichPass<=2; whichPass++) { for(whichHalf = 1; whichHalf<=2; whichHalf++) { vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; vector float tmp10, tmp11, tmp12, tmp13; vector float z1, z2, z3, z4, z5; tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7]; tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7]; tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4]; tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4]; tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6]; tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6]; tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5]; tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5]; tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3; tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3; tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2; tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2; // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); row0 = vec_add(tmp10, tmp11); // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); row4 = vec_sub(tmp10, tmp11); // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero); // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), // CONST_BITS-PASS1_BITS); row2 = vec_madd(tmp13, vec_0_765366865, z1); // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), // CONST_BITS-PASS1_BITS); row6 = vec_madd(tmp12, vec_1_847759065, z1); z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7; z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6; z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6; z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7; // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero); // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ z3 = vec_madd(z3, vec_1_961570560, z5); // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ z4 = vec_madd(z4, vec_0_390180644, z5); // The following adds are rolled into the multiplies above // z3 = vec_add(z3, z5); // z3 += z5; // z4 = vec_add(z4, z5); // z4 += z5; // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ // Wow! It's actually more efficient to roll this multiply // into the adds below, even thought the multiply gets done twice! // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero); // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ // Same with this one... // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero); // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3)); // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4)); // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3)); // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4)); // Swap the row values with the alts. If this is the first half, // this sets up the low values to be acted on in the second half. // If this is the second half, it puts the high values back in // the row values where they are expected to be when we're done. SWAP(row0, alt0); SWAP(row1, alt1); SWAP(row2, alt2); SWAP(row3, alt3); SWAP(row4, alt4); SWAP(row5, alt5); SWAP(row6, alt6); SWAP(row7, alt7); } if (whichPass == 1) { // transpose the data for the second pass // First, block transpose the upper right with lower left. SWAP(row4, alt0); SWAP(row5, alt1); SWAP(row6, alt2); SWAP(row7, alt3); // Now, transpose each block of four TRANSPOSE4(row0, row1, row2, row3); TRANSPOSE4(row4, row5, row6, row7); TRANSPOSE4(alt0, alt1, alt2, alt3); TRANSPOSE4(alt4, alt5, alt6, alt7); } } } // perform the quantize step, using the floating point data // still in the row/alt registers { const int* biasAddr; const vector signed int* qmat; vector float bias, negBias; if (s->mb_intra) { vector signed int baseVector; // We must cache element 0 in the intra case // (it needs special handling). baseVector = vec_cts(vec_splat(row0, 0), 0); vec_ste(baseVector, 0, &oldBaseValue); qmat = (vector signed int*)s->q_intra_matrix[qscale]; biasAddr = &(s->intra_quant_bias); } else { qmat = (vector signed int*)s->q_inter_matrix[qscale]; biasAddr = &(s->inter_quant_bias); } // Load the bias vector (We add 0.5 to the bias so that we're // rounding when we convert to int, instead of flooring.) { vector signed int biasInt; const vector float negOneFloat = (vector float)FOUROF(-1.0f); LOAD4(biasInt, biasAddr); bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); negBias = vec_madd(bias, negOneFloat, zero); } { vector float q0, q1, q2, q3, q4, q5, q6, q7; q0 = vec_ctf(qmat[0], QMAT_SHIFT); q1 = vec_ctf(qmat[2], QMAT_SHIFT); q2 = vec_ctf(qmat[4], QMAT_SHIFT); q3 = vec_ctf(qmat[6], QMAT_SHIFT); q4 = vec_ctf(qmat[8], QMAT_SHIFT); q5 = vec_ctf(qmat[10], QMAT_SHIFT); q6 = vec_ctf(qmat[12], QMAT_SHIFT); q7 = vec_ctf(qmat[14], QMAT_SHIFT); row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias), vec_cmpgt(row0, zero)); row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias), vec_cmpgt(row1, zero)); row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias), vec_cmpgt(row2, zero)); row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias), vec_cmpgt(row3, zero)); row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias), vec_cmpgt(row4, zero)); row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias), vec_cmpgt(row5, zero)); row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias), vec_cmpgt(row6, zero)); row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias), vec_cmpgt(row7, zero)); q0 = vec_ctf(qmat[1], QMAT_SHIFT); q1 = vec_ctf(qmat[3], QMAT_SHIFT); q2 = vec_ctf(qmat[5], QMAT_SHIFT); q3 = vec_ctf(qmat[7], QMAT_SHIFT); q4 = vec_ctf(qmat[9], QMAT_SHIFT); q5 = vec_ctf(qmat[11], QMAT_SHIFT); q6 = vec_ctf(qmat[13], QMAT_SHIFT); q7 = vec_ctf(qmat[15], QMAT_SHIFT); alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias), vec_cmpgt(alt0, zero)); alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias), vec_cmpgt(alt1, zero)); alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias), vec_cmpgt(alt2, zero)); alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias), vec_cmpgt(alt3, zero)); alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias), vec_cmpgt(alt4, zero)); alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias), vec_cmpgt(alt5, zero)); alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias), vec_cmpgt(alt6, zero)); alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias), vec_cmpgt(alt7, zero)); } } // Store the data back into the original block { vector signed short data0, data1, data2, data3, data4, data5, data6, data7; data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0)); data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0)); data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0)); data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0)); data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0)); data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0)); data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0)); data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0)); { // Clamp for overflow vector signed int max_q_int, min_q_int; vector signed short max_q, min_q; LOAD4(max_q_int, &(s->max_qcoeff)); LOAD4(min_q_int, &(s->min_qcoeff)); max_q = vec_pack(max_q_int, max_q_int); min_q = vec_pack(min_q_int, min_q_int); data0 = vec_max(vec_min(data0, max_q), min_q); data1 = vec_max(vec_min(data1, max_q), min_q); data2 = vec_max(vec_min(data2, max_q), min_q); data4 = vec_max(vec_min(data4, max_q), min_q); data5 = vec_max(vec_min(data5, max_q), min_q); data6 = vec_max(vec_min(data6, max_q), min_q); data7 = vec_max(vec_min(data7, max_q), min_q); } { vector bool char zero_01, zero_23, zero_45, zero_67; vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67; vector signed char negOne = vec_splat_s8(-1); vector signed char* scanPtr = (vector signed char*)(s->intra_scantable.inverse); signed char lastNonZeroChar; // Determine the largest non-zero index. zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero), vec_cmpeq(data1, (vector signed short)zero)); zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero), vec_cmpeq(data3, (vector signed short)zero)); zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero), vec_cmpeq(data5, (vector signed short)zero)); zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero), vec_cmpeq(data7, (vector signed short)zero)); // 64 biggest values scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01); scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23); scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45); scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67); // 32 largest values scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23); scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67); // 16 largest values scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45); // 8 largest values scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), vec_mergel(scanIndexes_01, negOne)); // 4 largest values scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), vec_mergel(scanIndexes_01, negOne)); // 2 largest values scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), vec_mergel(scanIndexes_01, negOne)); // largest value scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), vec_mergel(scanIndexes_01, negOne)); scanIndexes_01 = vec_splat(scanIndexes_01, 0); vec_ste(scanIndexes_01, 0, &lastNonZeroChar); lastNonZero = lastNonZeroChar; // While the data is still in vectors we check for the transpose IDCT permute // and handle it using the vector unit if we can. This is the permute used // by the altivec idct, so it is common when using the altivec dct. if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) { TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); } vec_st(data0, 0, data); vec_st(data1, 16, data); vec_st(data2, 32, data); vec_st(data3, 48, data); vec_st(data4, 64, data); vec_st(data5, 80, data); vec_st(data6, 96, data); vec_st(data7, 112, data); } } // special handling of block[0] if (s->mb_intra) { if (!s->h263_aic) { if (n < 4) oldBaseValue /= s->y_dc_scale; else oldBaseValue /= s->c_dc_scale; } // Divide by 8, rounding the result data[0] = (oldBaseValue + 4) >> 3; } // We handled the transpose permutation above and we don't // need to permute the "no" permutation case. if ((lastNonZero > 0) && (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) { ff_block_permute(data, s->dsp.idct_permutation, s->intra_scantable.scantable, lastNonZero); } return lastNonZero; }
inline v_int32x4 v_trunc(const v_float32x4& a) { return v_int32x4(vec_cts(a.val)); }
inline v_int32x4 v_ceil(const v_float32x4& a) { return v_int32x4(vec_cts(vec_ceil(a.val))); }
/** Rounding **/ inline v_int32x4 v_round(const v_float32x4& a) { return v_int32x4(vec_cts(vec_round(a.val))); }
static void ProjectDlightTexture_altivec( void ) { int i, l; vec_t origin0, origin1, origin2; float texCoords0, texCoords1; vector float floatColorVec0, floatColorVec1; vector float modulateVec, colorVec, zero; vector short colorShort; vector signed int colorInt; vector unsigned char floatColorVecPerm, modulatePerm, colorChar; vector unsigned char vSel = VECCONST_UINT8(0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff); float *texCoords; byte *colors; int *intColors; byte clipBits[SHADER_MAX_VERTEXES]; float texCoordsArray[SHADER_MAX_VERTEXES][2]; byte colorArray[SHADER_MAX_VERTEXES][4]; glIndex_t hitIndexes[SHADER_MAX_INDEXES]; int numIndexes; float scale; float radius; float radiusInverseCubed; float intensity, remainder; vec3_t floatColor; float modulate = 0.0f; qboolean vertexLight; if ( !backEnd.refdef.num_dlights ) { return; } // There has to be a better way to do this so that floatColor // and/or modulate are already 16-byte aligned. floatColorVecPerm = vec_lvsl(0,(float *)floatColor); modulatePerm = vec_lvsl(0,(float *)&modulate); modulatePerm = (vector unsigned char)vec_splat((vector unsigned int)modulatePerm,0); zero = (vector float)vec_splat_s8(0); for ( l = 0 ; l < backEnd.refdef.num_dlights ; l++ ) { dlight_t *dl; if ( !( tess.dlightBits & ( 1 << l ) ) ) { continue; // this surface definately doesn't have any of this light } // clear colors Com_Memset( colorArray, 0, sizeof( colorArray ) ); texCoords = texCoordsArray[0]; colors = colorArray[0]; dl = &backEnd.refdef.dlights[l]; origin0 = dl->transformed[0]; origin1 = dl->transformed[1]; origin2 = dl->transformed[2]; radius = dl->radius; scale = 1.0f / radius; radiusInverseCubed = dl->radiusInverseCubed; intensity = dl->intensity; vertexLight = ( ( dl->flags & REF_DIRECTED_DLIGHT ) || ( dl->flags & REF_VERTEX_DLIGHT ) ); // directional lights have max intensity and washout remainder intensity if ( dl->flags & REF_DIRECTED_DLIGHT ) { remainder = intensity * 0.125; } else { remainder = 0.0f; } if(r_greyscale->integer) { float luminance; luminance = LUMA(dl->color[0], dl->color[1], dl->color[2]) * 255.0f; floatColor[0] = floatColor[1] = floatColor[2] = luminance; } else if(r_greyscale->value) { float luminance; luminance = LUMA(dl->color[0], dl->color[1], dl->color[2]) * 255.0f; floatColor[0] = LERP(dl->color[0] * 255.0f, luminance, r_greyscale->value); floatColor[1] = LERP(dl->color[1] * 255.0f, luminance, r_greyscale->value); floatColor[2] = LERP(dl->color[2] * 255.0f, luminance, r_greyscale->value); } else { floatColor[0] = dl->color[0] * 255.0f; floatColor[1] = dl->color[1] * 255.0f; floatColor[2] = dl->color[2] * 255.0f; } floatColorVec0 = vec_ld(0, floatColor); floatColorVec1 = vec_ld(11, floatColor); floatColorVec0 = vec_perm(floatColorVec0,floatColorVec0,floatColorVecPerm); for ( i = 0 ; i < tess.numVertexes ; i++, texCoords += 2, colors += 4 ) { int clip = 0; vec_t dist0, dist1, dist2; dist0 = origin0 - tess.xyz[i][0]; dist1 = origin1 - tess.xyz[i][1]; dist2 = origin2 - tess.xyz[i][2]; backEnd.pc.c_dlightVertexes++; // directional dlight, origin is a directional normal if ( dl->flags & REF_DIRECTED_DLIGHT ) { // twosided surfaces use absolute value of the calculated lighting modulate = intensity * DotProduct( dl->origin, tess.normal[ i ] ); if ( tess.shader->cullType == CT_TWO_SIDED ) { modulate = fabs( modulate ); } modulate += remainder; } // spherical vertex lit dlight else if ( dl->flags & REF_VERTEX_DLIGHT ) { vec3_t dir; dir[ 0 ] = radius - fabs( dist0 ); if ( dir[ 0 ] <= 0.0f ) { continue; } dir[ 1 ] = radius - fabs( dist1 ); if ( dir[ 1 ] <= 0.0f ) { continue; } dir[ 2 ] = radius - fabs( dist2 ); if ( dir[ 2 ] <= 0.0f ) { continue; } modulate = intensity * dir[ 0 ] * dir[ 1 ] * dir[ 2 ] * radiusInverseCubed; } // vertical cylinder dlight else { texCoords0 = 0.5f + dist0 * scale; texCoords1 = 0.5f + dist1 * scale; if( !r_dlightBacks->integer && // dist . tess.normal[i] ( dist0 * tess.normal[i][0] + dist1 * tess.normal[i][1] + dist2 * tess.normal[i][2] ) < 0.0f ) { clip = 63; } else { if ( texCoords0 < 0.0f ) { clip |= 1; } else if ( texCoords0 > 1.0f ) { clip |= 2; } if ( texCoords1 < 0.0f ) { clip |= 4; } else if ( texCoords1 > 1.0f ) { clip |= 8; } texCoords[0] = texCoords0; texCoords[1] = texCoords1; // modulate the strength based on the height and color if ( dist2 > radius ) { clip |= 16; modulate = 0.0f; } else if ( dist2 < -radius ) { clip |= 32; modulate = 0.0f; } else { dist2 = Q_fabs(dist2); if ( dist2 < radius * 0.5f ) { modulate = intensity; } else { modulate = intensity * 2.0f * (radius - dist2) * scale; } } } } clipBits[i] = clip; // optimizations if ( vertexLight && modulate < ( 1.0f / 128.0f ) ) { continue; } else if ( modulate > 1.0f ) { modulate = 1.0f; } // ZTM: FIXME: should probably clamp to 0-255 range before converting to char, // but I don't know how to do altvec stuff or if it's even used anymore modulateVec = vec_ld(0,(float *)&modulate); modulateVec = vec_perm(modulateVec,modulateVec,modulatePerm); colorVec = vec_madd(floatColorVec0,modulateVec,zero); colorInt = vec_cts(colorVec,0); // RGBx colorShort = vec_pack(colorInt,colorInt); // RGBxRGBx colorChar = vec_packsu(colorShort,colorShort); // RGBxRGBxRGBxRGBx colorChar = vec_sel(colorChar,vSel,vSel); // RGBARGBARGBARGBA replace alpha with 255 vec_ste((vector unsigned int)colorChar,0,(unsigned int *)colors); // store color } // build a list of triangles that need light intColors = (int*) colorArray; numIndexes = 0; for ( i = 0 ; i < tess.numIndexes ; i += 3 ) { int a, b, c; a = tess.indexes[i]; b = tess.indexes[i+1]; c = tess.indexes[i+2]; if ( vertexLight ) { if ( !( intColors[ a ] | intColors[ b ] | intColors[ c ] ) ) { continue; } } else { if ( clipBits[a] & clipBits[b] & clipBits[c] ) { continue; // not lighted } } hitIndexes[numIndexes] = a; hitIndexes[numIndexes+1] = b; hitIndexes[numIndexes+2] = c; numIndexes += 3; } if ( !numIndexes ) { continue; } if ( !vertexLight ) { qglEnableClientState( GL_TEXTURE_COORD_ARRAY ); qglTexCoordPointer( 2, GL_FLOAT, 0, texCoordsArray[0] ); } else { qglDisableClientState( GL_TEXTURE_COORD_ARRAY ); } qglEnableClientState( GL_COLOR_ARRAY ); qglColorPointer( 4, GL_UNSIGNED_BYTE, 0, colorArray ); if ( dl->dlshader ) { shader_t *dls = dl->dlshader; for ( i = 0; i < dls->numUnfoggedPasses; i++ ) { shaderStage_t *stage = dls->stages[i]; R_BindAnimatedImage( &dls->stages[i]->bundle[0] ); GL_State( stage->stateBits | GLS_DEPTHFUNC_EQUAL ); R_DrawElements( numIndexes, hitIndexes ); backEnd.pc.c_totalIndexes += numIndexes; backEnd.pc.c_dlightIndexes += numIndexes; } } else { R_FogOff(); if ( !vertexLight ) { GL_Bind( tr.dlightImage ); } else { GL_Bind( tr.whiteImage ); } // include GLS_DEPTHFUNC_EQUAL so alpha tested surfaces don't add light // where they aren't rendered if ( dl->flags & REF_ADDITIVE_DLIGHT ) { GL_State( GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE | GLS_DEPTHFUNC_EQUAL ); } else { GL_State( GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ONE | GLS_DEPTHFUNC_EQUAL ); } R_DrawElements( numIndexes, hitIndexes ); backEnd.pc.c_totalIndexes += numIndexes; backEnd.pc.c_dlightIndexes += numIndexes; RB_FogOn(); } } }
/* ************************************************************************* NAME: test_float2char USAGE: test_float2char(); returns: void DESCRIPTION: convert floats to ints to shorts to chars to test rounding/truncating REFERENCES: Ian Ollmann's Altivec Tutorial LIMITATIONS: GLOBAL VARIABLES: accessed: none modified: none FUNCTIONS CALLED: fprintf vec_st - store a vector to memory vec_cts - convert to fixed point vec_pack - narrow 2 vectors and pack them into one vector vec_packsu - saturating pack, unsigned result REVISION HISTORY: STR Description of Revision Author 07-Feb-2011 Based off of gpk's condense_float_rgbas kaj ************************************************************************* */ void test_float2char(void) { vector float floatVec1 = {0.0, -50.14455, 21.7790, 100.0 }; vector float floatVec2 = {-100.4567, 250.14455, -210.7790, 170.0 }; vector float floatVec3 = {0.09876, -150.55, 1.7790, -120.0 }; vector float floatVec4 = {-90.234, -30.9455, -205.90, 199.9 }; vector signed int intVec1, intVec2, intVec3, intVec4; vector signed short shortVec1, shortVec2; vector unsigned char charVec; unsigned char printchar[CHAR_ARRAYSIZE] __attribute__ ((aligned (16))); float printfloat[FLOAT_ARRAYSIZE] __attribute__ ((aligned (16))); signed short printshort[SHORT_ARRAYSIZE] __attribute__ ((aligned (16))); signed int printint[INT_ARRAYSIZE] __attribute__ ((aligned (16))); int i; fprintf(stderr, "--- function %s ------\n", __FUNCTION__); /* print out floats */ vec_st(floatVec1,0,printfloat); printFloats(printfloat); vec_st(floatVec2,0,printfloat); printFloats(printfloat); vec_st(floatVec3,0,printfloat); printFloats(printfloat); vec_st(floatVec4,0,printfloat); printFloats(printfloat); //floatVec1 = vec_trunc(floatVec1); vec_trunc(floatVec1); /* convert from float to signed int and print */ intVec1 = vec_cts(floatVec1,0); vec_st(intVec1,0,printint); printInts(printint); intVec2 = vec_cts(floatVec2,0); vec_st(intVec2,0,printint); printInts(printint); intVec3 = vec_cts(floatVec3,0); vec_st(intVec3,0,printint); printInts(printint); intVec4 = vec_cts(floatVec4,0); vec_st(intVec4,0,printint); printInts(printint); /* convert from signed int to signed short and print */ shortVec1 = vec_pack(intVec1,intVec2); vec_st(shortVec1,0,printshort); printShorts(printshort); shortVec2 = vec_pack(intVec3,intVec4); vec_st(shortVec2,0,printshort); printShorts(printshort); /* convert from signed short to unsigned char and print */ charVec = vec_packsu(shortVec1, shortVec2); vec_st(charVec,0,printchar); printChars(printchar); } /* test_float2char */
void mandel_altivec(unsigned char *image, const struct spec *s) { vector float xmin, ymin, xscale, yscale, iter_scale, depth_scale; vector float threshold = VF_ALL(4.0); vector float one = VF_ALL(1.0); vector float zero = VF_ALL(0.0); xmin = VF_ALL(s->xlim[0]); ymin = VF_ALL(s->ylim[0]); xscale = VF_ALL((s->xlim[1] - s->xlim[0]) / s->width); yscale = VF_ALL((s->ylim[1] - s->ylim[0]) / s->height); iter_scale = VF_ALL(1.0f / s->iterations); depth_scale = VF_ALL(s->depth - 1); #pragma omp parallel for schedule(dynamic, 1) for (int y = 0; y < s->height; y++) { for (int x = 0; x < s->width; x += 4) { vector float mx = (vector float) { x, x + 1, x + 2, x + 3 }; vector float my = VF_ALL(y); vector float cr = vec_madd(mx, xscale, xmin); vector float ci = vec_madd(my, yscale, ymin); vector float zr = cr; vector float zi = ci; int k = 1; vector float mk = VF_ALL(1); while (++k < s->iterations) { /* Compute z1 from z0 */ vector float zr2cr = vec_madd(zr, zr, cr); vector float zi2 = vec_madd(zi, zi, zero); vector float zrzi = vec_madd(zr, zi, zero); /* zr1 = zr0 * zr0 - zi0 * zi0 + cr */ /* zi1 = zr0 * zi0 + zr0 * zi0 + ci */ zr = vec_sub(zr2cr, zi2); zi = vec_add(vec_add(zrzi, zrzi), ci); /* Increment k */ vector float zr2 = vec_madd(zr, zr, zero); vector float mag2 = vec_madd(zi, zi, zr2); vector bool int mask = vec_cmplt(mag2, threshold); mk = vec_add(mk, vec_and(one, mask)); if(vec_all_ge(mag2, threshold)) break; } mk = vec_madd(mk, iter_scale, zero); mk = vec_madd(vec_rsqrte(mk), mk, zero); mk = vec_madd(mk, depth_scale, zero); vector int pixels = vec_cts(mk, 0); unsigned char *dst = image + y * s->width * 3 + x * 3; unsigned char *src = (unsigned char *)&pixels; for (int i = 0; i < 4; i++) { dst[i * 3 + 0] = src[(i * 4) + 3]; dst[i * 3 + 1] = src[(i * 4) + 3]; dst[i * 3 + 2] = src[(i * 4) + 3]; } } } }