bool SamplerJitCache::Jit_Decode5551() {
MOV(32, R(tempReg2), R(resultReg));
MOV(32, R(tempReg1), R(resultReg));
AND(32, R(tempReg2), Imm32(0x0000001F));
AND(32, R(tempReg1), Imm32(0x000003E0));
SHL(32, R(tempReg1), Imm8(3));
OR(32, R(tempReg2), R(tempReg1));
MOV(32, R(tempReg1), R(resultReg));
AND(32, R(tempReg1), Imm32(0x00007C00));
SHL(32, R(tempReg1), Imm8(6));
OR(32, R(tempReg2), R(tempReg1));
// Expand 5 -> 8. After this is just A.
MOV(32, R(tempReg1), R(tempReg2));
SHL(32, R(tempReg2), Imm8(3));
SHR(32, R(tempReg1), Imm8(2));
// Chop off the bits that were shifted out.
AND(32, R(tempReg1), Imm32(0x00070707));
OR(32, R(tempReg2), R(tempReg1));
// For A, we shift it to a single bit, and then subtract and XOR.
// That's probably the simplest way to expand it...
SHR(32, R(resultReg), Imm8(15));
// If it was 0, it's now -1, otherwise it's 0. Easy.
SUB(32, R(resultReg), Imm8(1));
XOR(32, R(resultReg), Imm32(0xFF000000));
AND(32, R(resultReg), Imm32(0xFF000000));
OR(32, R(resultReg), R(tempReg2));
return true;
}
bool SamplerJitCache::Jit_Decode5650() {
MOV(32, R(tempReg2), R(resultReg));
AND(32, R(tempReg2), Imm32(0x0000001F));
// B (we do R and B at the same time, they're both 5.)
MOV(32, R(tempReg1), R(resultReg));
AND(32, R(tempReg1), Imm32(0x0000F800));
SHL(32, R(tempReg1), Imm8(5));
OR(32, R(tempReg2), R(tempReg1));
// Expand 5 -> 8. At this point we have 00BB00RR.
MOV(32, R(tempReg1), R(tempReg2));
SHL(32, R(tempReg2), Imm8(3));
SHR(32, R(tempReg1), Imm8(2));
OR(32, R(tempReg2), R(tempReg1));
AND(32, R(tempReg2), Imm32(0x00FF00FF));
// Now's as good a time to put in A as any.
OR(32, R(tempReg2), Imm32(0xFF000000));
// Last, we need to align, extract, and expand G.
// 3 to align to G, and then 2 to expand to 8.
SHL(32, R(resultReg), Imm8(3 + 2));
AND(32, R(resultReg), Imm32(0x0000FC00));
MOV(32, R(tempReg1), R(resultReg));
// 2 to account for resultReg being preshifted, 4 for expansion.
SHR(32, R(tempReg1), Imm8(2 + 4));
OR(32, R(resultReg), R(tempReg1));
AND(32, R(resultReg), Imm32(0x0000FF00));
OR(32, R(resultReg), R(tempReg2));;
return true;
}
void
arch_store16(cpu_t *cpu, Value *val, Value *addr, BasicBlock *bb) {
Value *shift = arch_get_shift16(cpu, addr, bb);
addr = AND(addr, CONST(~3ULL));
Value *mask = XOR(SHL(CONST(65535), shift),CONST(-1ULL));
Value *old = AND(arch_load32_aligned(cpu, addr, bb), mask);
val = OR(old, SHL(AND(val, CONST(65535)), shift));
arch_store32_aligned(cpu, val, addr, bb);
}
void split_cb_shape_sign_unquant(
spx_sig_t *exc,
const void *par, /* non-overlapping codebook */
int nsf, /* number of samples in subframe */
SpeexBits *bits,
char *stack
)
{
int i,j;
int *ind, *signs;
const signed char *shape_cb;
int shape_cb_size, subvect_size, nb_subvect;
split_cb_params *params;
int have_sign;
params = (split_cb_params *) par;
subvect_size = params->subvect_size;
nb_subvect = params->nb_subvect;
shape_cb_size = 1<<params->shape_bits;
shape_cb = params->shape_cb;
have_sign = params->have_sign;
ind = PUSH(stack, nb_subvect, int);
signs = PUSH(stack, nb_subvect, int);
/* Decode codewords and gains */
for (i=0;i<nb_subvect;i++)
{
if (have_sign)
signs[i] = speex_bits_unpack_unsigned(bits, 1);
else
signs[i] = 0;
ind[i] = speex_bits_unpack_unsigned(bits, params->shape_bits);
}
/* Compute decoded excitation */
for (i=0;i<nb_subvect;i++)
{
spx_word16_t s=1;
if (signs[i])
s=-1;
#ifdef FIXED_POINT
if (s==1)
{
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]=SHL((spx_word32_t)shape_cb[ind[i]*subvect_size+j],SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]=-SHL((spx_word32_t)shape_cb[ind[i]*subvect_size+j],SIG_SHIFT-5);
}
#else
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]+=s*0.03125*shape_cb[ind[i]*subvect_size+j];
#endif
}
}
word32_t ChebyshevPolynomial(word16_t x, word32_t f[])
{
/* bk in Q15*/
word32_t bk;
word32_t bk1 = ADD32(SHL(x,1), f[1]); /* init: b4=2x+f1 */
word32_t bk2 = ONE_IN_Q15; /* init: b5=1 */
uint8_t k;
for (k=3; k>0; k--) { /* at the end of loop execution we have b1 in bk1 and b2 in bk2 */
bk = SUB32(ADD32(SHL(MULT16_32_Q15(x,bk1), 1), f[5-k]), bk2); /* bk = 2*x*bk1 − bk2 + f(5-k) all in Q15*/
bk2 = bk1;
bk1 = bk;
}
return SUB32(ADD32(MULT16_32_Q15(x,bk1), SHR(f[5],1)), bk2); /* C(x) = x*b1 - b2 + f(5)/2 */
}
static inline Value *
arch_6502_pull16(cpu_t *cpu, BasicBlock *bb)
{
Value *lo = PULL;
Value *hi = PULL;
return (OR(ZEXT16(lo), SHL(ZEXT16(hi), CONST16(8))));
}
/*** shlGetAttrType - get the type (DATA_T_xxx) of an attribute by name.
***/
int
shlGetAttrType(void* inf_v, char* attrname, pObjTrxTree* oxt)
{
pShlData inf = SHL(inf_v);
int i;
pStructInf find_inf;
pEnvVar pEV;
/** If name, it's a string **/
if (!strcmp(attrname,"name")) return DATA_T_STRING;
/** If 'content-type', it's also a string. **/
if (!strcmp(attrname,"content_type")) return DATA_T_STRING;
if (!strcmp(attrname,"inner_type")) return DATA_T_STRING;
if (!strcmp(attrname,"outer_type")) return DATA_T_STRING;
if (!strcmp(attrname,"annotation")) return DATA_T_STRING;
if (!strcmp(attrname,"status")) return DATA_T_STRING;
if (!strncmp(attrname,"arg",3)) return DATA_T_STRING;
pEV = (pEnvVar)xhLookup(&inf->envHash,attrname);
if(pEV)
{
return DATA_T_STRING;
}
mssError(1,"SHL","Could not locate requested attribute: %s",attrname);
return -1;
}
/*** shlGetNextAttr - get the next attribute name for this object.
***/
char*
shlGetNextAttr(void* inf_v, pObjTrxTree oxt)
{
pShlData inf = SHL(inf_v);
int i;
if(inf->CurAttr>=xaCount(&inf->argArray)+xaCount(&inf->envList)+1)
return NULL;
if(inf->CurAttr==xaCount(&inf->argArray)+xaCount(&inf->envList))
{
inf->CurAttr++;
return "status";
}
if(inf->CurAttr>999999 && xaCount(&inf->argArray)>=999999)
return NULL;
if(inf->CurAttr<xaCount(&inf->argArray))
{
i=snprintf(inf->sCurAttr,10,"arg%02i",inf->CurAttr++);
inf->sCurAttr[10]='\0';
return inf->sCurAttr;
}
else
{
return (char*)xaGetItem(&inf->envList,(inf->CurAttr++)-xaCount(&inf->argArray));
}
}
/* FIXME: These functions are ugly and probably introduce too much error */
void signal_mul(const spx_sig_t *x, spx_sig_t *y, spx_word32_t scale, int len)
{
int i;
for (i=0;i<len;i++)
{
y[i] = SHL(MULT16_32_Q14(SHR(x[i],7),scale),7);
}
}
static Value *
arch_get_shift16(cpu_t *cpu, Value *addr, BasicBlock *bb)
{
Value *shift = AND(LSHR(addr,CONST(1)),CONST(1));
if (!IS_LITTLE_ENDIAN(cpu))
shift = XOR(shift, CONST(1));
return SHL(shift, CONST(4));
}
/*** shlGetFirstAttr - get the first attribute name for this object.
***/
char*
shlGetFirstAttr(void* inf_v, pObjTrxTree oxt)
{
pShlData inf = SHL(inf_v);
inf->CurAttr=0;
return shlGetNextAttr(inf_v,oxt);
}
/*** shlAddAttr - add an attribute to an object. This doesn't always work
*** for all object types, and certainly makes no sense for some (like unix
*** files).
***/
int
shlAddAttr(void* inf_v, char* attrname, int type, void* val, pObjTrxTree oxt)
{
pShlData inf = SHL(inf_v);
pStructInf new_inf;
char* ptr;
return -1;
}
void lsp_interpolate(spx_lsp_t *old_lsp, spx_lsp_t *new_lsp, spx_lsp_t *interp_lsp, int len, int subframe, int nb_subframes)
{
int i;
spx_word16_t tmp = DIV32_16(SHL(1 + subframe,14),nb_subframes);
spx_word16_t tmp2 = 16384-tmp;
for (i=0;i<len;i++)
{
interp_lsp[i] = MULT16_16_P14(tmp2,old_lsp[i]) + MULT16_16_P14(tmp,new_lsp[i]);
}
}
void
vec4_gs_visitor::gs_end_primitive()
{
/* We can only do EndPrimitive() functionality when the control data
* consists of cut bits. Fortunately, the only time it isn't is when the
* output type is points, in which case EndPrimitive() is a no-op.
*/
if (gs_prog_data->control_data_format !=
GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT) {
return;
}
if (c->control_data_header_size_bits == 0)
return;
/* Cut bits use one bit per vertex. */
assert(c->control_data_bits_per_vertex == 1);
/* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
* vertex n, 0 otherwise. So all we need to do here is mark bit
* (vertex_count - 1) % 32 in the cut_bits register to indicate that
* EndPrimitive() was called after emitting vertex (vertex_count - 1);
* vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
*
* Note that if EndPrimitve() is called before emitting any vertices, this
* will cause us to set bit 31 of the control_data_bits register to 1.
* That's fine because:
*
* - If max_vertices < 32, then vertex number 31 (zero-based) will never be
* output, so the hardware will ignore cut bit 31.
*
* - If max_vertices == 32, then vertex number 31 is guaranteed to be the
* last vertex, so setting cut bit 31 has no effect (since the primitive
* is automatically ended when the GS terminates).
*
* - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
* control_data_bits register to 0 when the first vertex is emitted.
*/
/* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
src_reg one(this, glsl_type::uint_type);
emit(MOV(dst_reg(one), brw_imm_ud(1u)));
src_reg prev_count(this, glsl_type::uint_type);
emit(ADD(dst_reg(prev_count), this->vertex_count, brw_imm_ud(0xffffffffu)));
src_reg mask(this, glsl_type::uint_type);
/* Note: we're relying on the fact that the GEN SHL instruction only pays
* attention to the lower 5 bits of its second source argument, so on this
* architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
* ((vertex_count - 1) % 32).
*/
emit(SHL(dst_reg(mask), one, prev_count));
emit(OR(dst_reg(this->control_data_bits), this->control_data_bits, mask));
}
// shifts or rotates an lvalue left or right, regardless of width
Value *
arch_shiftrotate(cpu_t *cpu, Value *dst, Value *src, bool left, bool rotate, BasicBlock *bb)
{
Value *c;
Value *v = LOAD(src);
if (left) {
c = ICMP_SLT(v, CONSTs(SIZE(v), 0)); /* old MSB to carry */
v = SHL(v, CONSTs(SIZE(v), 1));
if (rotate)
v = OR(v,ZEXT(SIZE(v), LOAD(cpu->ptr_C)));
} else {
c = TRUNC1(v); /* old LSB to carry */
v = LSHR(v, CONSTs(SIZE(v), 1));
if (rotate)
v = OR(v,SHL(ZEXT(SIZE(v), LOAD(cpu->ptr_C)), CONSTs(SIZE(v), SIZE(v)-1)));
}
LET1(cpu->ptr_C, c);
return STORE(v, dst);
}
void
vec4_gs_visitor::set_stream_control_data_bits(unsigned stream_id)
{
/* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
/* Note: we are calling this *before* increasing vertex_count, so
* this->vertex_count == vertex_count - 1 in the formula above.
*/
/* Stream mode uses 2 bits per vertex */
assert(c->control_data_bits_per_vertex == 2);
/* Must be a valid stream */
assert(stream_id >= 0 && stream_id < MAX_VERTEX_STREAMS);
/* Control data bits are initialized to 0 so we don't have to set any
* bits when sending vertices to stream 0.
*/
if (stream_id == 0)
return;
/* reg::sid = stream_id */
src_reg sid(this, glsl_type::uint_type);
emit(MOV(dst_reg(sid), stream_id));
/* reg:shift_count = 2 * (vertex_count - 1) */
src_reg shift_count(this, glsl_type::uint_type);
emit(SHL(dst_reg(shift_count), this->vertex_count, 1u));
/* Note: we're relying on the fact that the GEN SHL instruction only pays
* attention to the lower 5 bits of its second source argument, so on this
* architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
* stream_id << ((2 * (vertex_count - 1)) % 32).
*/
src_reg mask(this, glsl_type::uint_type);
emit(SHL(dst_reg(mask), sid, shift_count));
emit(OR(dst_reg(this->control_data_bits), this->control_data_bits, mask));
}
void LPSynthesisFilter (word16_t *excitationVector, word16_t *LPCoefficients, word16_t *reconstructedSpeech)
{
int i;
/* compute excitationVector[i] - Sum0-9(LPCoefficients[j]*reconstructedSpeech[i-j]) */
for (i=0; i<L_SUBFRAME; i++) {
word32_t acc = SHL(excitationVector[i],12); /* acc get the first term of the sum, in Q12 (excitationVector is in Q0)*/
int j;
for (j=0; j<NB_LSP_COEFF; j++) {
acc = MSU16_16(acc, LPCoefficients[j], reconstructedSpeech[i-j-1]);
}
reconstructedSpeech[i] = (word16_t)SATURATE(PSHR(acc, 12), MAXINT16); /* shift right acc to get it back in Q0 and check overflow on 16 bits */
}
return;
}
/*** shlWrite - Write data to the shell's stdin
***/
int
shlWrite(void* inf_v, char* buffer, int cnt, int offset, int flags, pObjTrxTree* oxt)
{
pShlData inf = SHL(inf_v);
int i=-1;
int waitret;
int retval;
if(SHELL_DEBUG & SHELL_DEBUG_IO)
printf("%s -- %p, %p, %i, %i, %i, %p\n",__FUNCTION__,inf_v,buffer,cnt,offset,flags,oxt);
/** launch the program if it's not running already **/
if(inf->shell_pid == -1)
if(shl_internal_Launch(inf) < 0)
return -1;
/** seek is _not_ allowed (obviously) **/
if(flags & FD_U_SEEK && offset!=inf->curWrite)
return -1;
/** can't write to a dead child :) **/
if(!inf->shell_pid)
return -1;
while(i < 0)
{
i=fdWrite(inf->shell_fd,buffer,cnt,0,flags & ~FD_U_SEEK);
if(i < 0)
{
/** user doesn;t want us to block **/
if(flags & FD_U_NOBLOCK)
return -1;
/** check and make sure the child process is still alive... **/
shl_internal_UpdateStatus(inf);
if(inf->shell_pid>0)
{
/** child is alive, it just isn't ready for data yet -- wait a bit **/
thSleep(200);
}
else if(inf->shell_pid==0)
{
/** child died :( **/
return -1;
}
}
}
inf->curWrite+=i;
return i;
}
void
gen8_vec4_generator::generate_gs_prepare_channel_masks(struct brw_reg dst)
{
/* We want to left shift just DWORD 4 (the x component belonging to the
* second geometry shader invocation) by 4 bits. So generate the
* instruction:
*
* shl(1) dst.4<1>UD dst.4<0,1,0>UD 4UD { align1 WE_all }
*/
dst = suboffset(vec1(dst), 4);
default_state.access_mode = BRW_ALIGN_1;
gen8_instruction *inst = SHL(dst, dst, brw_imm_ud(4));
gen8_set_mask_control(inst, BRW_MASK_DISABLE);
default_state.access_mode = BRW_ALIGN_16;
}
bool SamplerJitCache::Jit_TransformClutIndex(const SamplerID &id, int bitsPerIndex) {
GEPaletteFormat fmt = (GEPaletteFormat)id.clutfmt;
if (!id.hasClutShift && !id.hasClutMask && !id.hasClutOffset) {
// This is simple - just mask if necessary.
if (bitsPerIndex > 8) {
AND(32, R(resultReg), Imm32(0x000000FF));
}
return true;
}
MOV(PTRBITS, R(tempReg1), ImmPtr(&gstate.clutformat));
MOV(32, R(tempReg1), MatR(tempReg1));
// Shift = (clutformat >> 2) & 0x1F
if (id.hasClutShift) {
MOV(32, R(RCX), R(tempReg1));
SHR(32, R(RCX), Imm8(2));
AND(32, R(RCX), Imm8(0x1F));
SHR(32, R(resultReg), R(RCX));
}
// Mask = (clutformat >> 8) & 0xFF
if (id.hasClutMask) {
MOV(32, R(tempReg2), R(tempReg1));
SHR(32, R(tempReg2), Imm8(8));
AND(32, R(resultReg), R(tempReg2));
}
// We need to wrap any entries beyond the first 1024 bytes.
u32 offsetMask = fmt == GE_CMODE_32BIT_ABGR8888 ? 0x00FF : 0x01FF;
// We must mask to 0xFF before ORing 0x100 in 16 bit CMODEs.
// But skip if we'll mask 0xFF after offset anyway.
if (bitsPerIndex > 8 && (!id.hasClutOffset || offsetMask != 0x00FF)) {
AND(32, R(resultReg), Imm32(0x000000FF));
}
// Offset = (clutformat >> 12) & 0x01F0
if (id.hasClutOffset) {
SHR(32, R(tempReg1), Imm8(16));
SHL(32, R(tempReg1), Imm8(4));
OR(32, R(resultReg), R(tempReg1));
AND(32, R(resultReg), Imm32(offsetMask));
}
return true;
}
void filter_mem2(const spx_sig_t *x, const spx_coef_t *num, const spx_coef_t *den, spx_sig_t *y, int N, int ord, spx_mem_t *mem)
{
int i,j;
spx_sig_t xi,yi,nyi;
for (i=0;i<N;i++)
{
int xh,xl,yh,yl;
xi=SATURATE(x[i],805306368);
yi = SATURATE(ADD32(xi, SHL(mem[0],2)),805306368);
nyi = -yi;
xh = xi>>15; xl=xi&0x00007fff; yh = yi>>15; yl=yi&0x00007fff;
for (j=0;j<ord-1;j++)
{
mem[j] = MAC16_32_Q15(MAC16_32_Q15(mem[j+1], num[j+1],xi), den[j+1],nyi);
}
mem[ord-1] = SUB32(MULT16_32_Q15(num[ord],xi), MULT16_32_Q15(den[ord],yi));
y[i] = yi;
}
}
/*** shlSetAttrValue - sets the value of an attribute. 'val' must
*** point to an appropriate data type.
***/
int
shlSetAttrValue(void* inf_v, char* attrname, int datatype, pObjData val, pObjTrxTree oxt)
{
pShlData inf = SHL(inf_v);
/** Choose the attr name **/
/** Changing name of node object? **/
if (!strcmp(attrname,"name"))
{
if (inf->Obj->Pathname->nElements == inf->Obj->SubPtr)
{
if (!strcmp(inf->Obj->Pathname->Pathbuf,".")) return -1;
if (strlen(inf->Obj->Pathname->Pathbuf) -
strlen(strrchr(inf->Obj->Pathname->Pathbuf,'/')) +
strlen(val->String) + 1 > 255)
{
mssError(1,"SHL","SetAttr 'name': name too large for internal representation");
return -1;
}
strcpy(inf->Pathname, inf->Obj->Pathname->Pathbuf);
strcpy(strrchr(inf->Pathname,'/')+1,val->String);
if (rename(inf->Obj->Pathname->Pathbuf, inf->Pathname) < 0)
{
mssError(1,"SHL","SetAttr 'name': could not rename structure file node object");
return -1;
}
strcpy(inf->Obj->Pathname->Pathbuf, inf->Pathname);
}
/** Set dirty flag **/
inf->Node->Status = SN_NS_DIRTY;
return 0;
}
/** Try to set a command parameter (env variable) **/
if (shl_internal_SetParam(inf, attrname, datatype, val) < 0)
return -1;
return -1;
}
spx_word16_t compute_rms(const spx_sig_t *x, int len)
{
int i;
spx_word32_t sum=0;
spx_sig_t max_val=1;
int sig_shift;
for (i=0;i<len;i++)
{
spx_sig_t tmp = x[i];
if (tmp<0)
tmp = -tmp;
if (tmp > max_val)
max_val = tmp;
}
sig_shift=0;
while (max_val>16383)
{
sig_shift++;
max_val >>= 1;
}
for (i=0;i<len;i+=4)
{
spx_word32_t sum2=0;
spx_word16_t tmp;
tmp = SHR(x[i],sig_shift);
sum2 += MULT16_16(tmp,tmp);
tmp = SHR(x[i+1],sig_shift);
sum2 += MULT16_16(tmp,tmp);
tmp = SHR(x[i+2],sig_shift);
sum2 += MULT16_16(tmp,tmp);
tmp = SHR(x[i+3],sig_shift);
sum2 += MULT16_16(tmp,tmp);
sum += SHR(sum2,6);
}
return SHR(SHL((spx_word32_t)spx_sqrt(1+DIV32(sum,len)),(sig_shift+3)),SIG_SHIFT);
}
INT16 q_gain_pitch ( /* Return index of quantization */
INT16 *gain /* (i) : Pitch gain to quantize */
)
{
INT16 i, index, gain_q14, err, err_min;
VPP_EFR_PROFILE_FUNCTION_ENTER(q_gain_pitch);
//gain_q14 = shl (*gain, 2);
gain_q14 = SHL(*gain, 2);
//err_min = abs_s (sub (gain_q14, qua_gain_pitch[0]));
err_min = ABS_S(SUB (gain_q14, qua_gain_pitch[0]));
index = 0;
for (i = 1; i < NB_QUA_PITCH; i++)
{
//err = abs_s (sub (gain_q14, qua_gain_pitch[i]));
err = ABS_S(SUB (gain_q14, qua_gain_pitch[i]));
//if (sub (err, err_min) < 0)
if (SUB (err, err_min) < 0)
{
err_min = err;
index = i;
}
}
//*gain = shr (qua_gain_pitch[index], 2);
*gain = SHR_D(qua_gain_pitch[index], 2);
VPP_EFR_PROFILE_FUNCTION_EXIT(q_gain_pitch);
return index;
}
void
vec4_vs_visitor::emit_prolog()
{
dst_reg sign_recovery_shift;
dst_reg normalize_factor;
dst_reg es3_normalize_factor;
for (int i = 0; i < VERT_ATTRIB_MAX; i++) {
if (vs_prog_data->inputs_read & BITFIELD64_BIT(i)) {
uint8_t wa_flags = key->gl_attrib_wa_flags[i];
dst_reg reg(ATTR, i);
dst_reg reg_d = reg;
reg_d.type = BRW_REGISTER_TYPE_D;
dst_reg reg_ud = reg;
reg_ud.type = BRW_REGISTER_TYPE_UD;
/* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes
* come in as floating point conversions of the integer values.
*/
if (wa_flags & BRW_ATTRIB_WA_COMPONENT_MASK) {
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
dst.writemask = (1 << (wa_flags & BRW_ATTRIB_WA_COMPONENT_MASK)) - 1;
emit(MUL(dst, src_reg(dst), src_reg(1.0f / 65536.0f)));
}
/* Do sign recovery for 2101010 formats if required. */
if (wa_flags & BRW_ATTRIB_WA_SIGN) {
if (sign_recovery_shift.file == BAD_FILE) {
/* shift constant: <22,22,22,30> */
sign_recovery_shift = dst_reg(this, glsl_type::uvec4_type);
emit(MOV(writemask(sign_recovery_shift, WRITEMASK_XYZ), src_reg(22u)));
emit(MOV(writemask(sign_recovery_shift, WRITEMASK_W), src_reg(30u)));
}
emit(SHL(reg_ud, src_reg(reg_ud), src_reg(sign_recovery_shift)));
emit(ASR(reg_d, src_reg(reg_d), src_reg(sign_recovery_shift)));
}
/* Apply BGRA swizzle if required. */
if (wa_flags & BRW_ATTRIB_WA_BGRA) {
src_reg temp = src_reg(reg);
temp.swizzle = BRW_SWIZZLE4(2,1,0,3);
emit(MOV(reg, temp));
}
if (wa_flags & BRW_ATTRIB_WA_NORMALIZE) {
/* ES 3.0 has different rules for converting signed normalized
* fixed-point numbers than desktop GL.
*/
if ((wa_flags & BRW_ATTRIB_WA_SIGN) && !use_legacy_snorm_formula) {
/* According to equation 2.2 of the ES 3.0 specification,
* signed normalization conversion is done by:
*
* f = c / (2^(b-1)-1)
*/
if (es3_normalize_factor.file == BAD_FILE) {
/* mul constant: 1 / (2^(b-1) - 1) */
es3_normalize_factor = dst_reg(this, glsl_type::vec4_type);
emit(MOV(writemask(es3_normalize_factor, WRITEMASK_XYZ),
src_reg(1.0f / ((1<<9) - 1))));
emit(MOV(writemask(es3_normalize_factor, WRITEMASK_W),
src_reg(1.0f / ((1<<1) - 1))));
}
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg(reg_d)));
emit(MUL(dst, src_reg(dst), src_reg(es3_normalize_factor)));
emit_minmax(BRW_CONDITIONAL_GE, dst, src_reg(dst), src_reg(-1.0f));
} else {
/* The following equations are from the OpenGL 3.2 specification:
*
* 2.1 unsigned normalization
* f = c/(2^n-1)
*
* 2.2 signed normalization
* f = (2c+1)/(2^n-1)
*
* Both of these share a common divisor, which is represented by
* "normalize_factor" in the code below.
*/
if (normalize_factor.file == BAD_FILE) {
/* 1 / (2^b - 1) for b=<10,10,10,2> */
normalize_factor = dst_reg(this, glsl_type::vec4_type);
emit(MOV(writemask(normalize_factor, WRITEMASK_XYZ),
src_reg(1.0f / ((1<<10) - 1))));
emit(MOV(writemask(normalize_factor, WRITEMASK_W),
src_reg(1.0f / ((1<<2) - 1))));
}
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg((wa_flags & BRW_ATTRIB_WA_SIGN) ? reg_d : reg_ud)));
/* For signed normalization, we want the numerator to be 2c+1. */
if (wa_flags & BRW_ATTRIB_WA_SIGN) {
emit(MUL(dst, src_reg(dst), src_reg(2.0f)));
emit(ADD(dst, src_reg(dst), src_reg(1.0f)));
}
emit(MUL(dst, src_reg(dst), src_reg(normalize_factor)));
}
}
if (wa_flags & BRW_ATTRIB_WA_SCALE) {
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg((wa_flags & BRW_ATTRIB_WA_SIGN) ? reg_d : reg_ud)));
}
}
}
}
void split_cb_search_shape_sign(
spx_sig_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
const void *par, /* Codebook/search parameters*/
int p, /* number of LPC coeffs */
int nsf, /* number of samples in subframe */
spx_sig_t *exc,
spx_sig_t *r,
SpeexBits *bits,
char *stack,
int complexity
)
{
int i,j,k,m,n,q;
spx_word16_t *resp;
#ifdef _USE_SSE
__m128 *resp2;
__m128 *E;
#else
spx_word16_t *resp2;
spx_word32_t *E;
#endif
spx_word16_t *t;
spx_sig_t *e, *r2;
spx_word16_t *tmp;
spx_word32_t *ndist, *odist;
int *itmp;
spx_word16_t **ot, **nt;
int **nind, **oind;
int *ind;
const signed char *shape_cb;
int shape_cb_size, subvect_size, nb_subvect;
split_cb_params *params;
int N=2;
int *best_index;
spx_word32_t *best_dist;
int have_sign;
N=complexity;
if (N>10)
N=10;
ot=PUSH(stack, N, spx_word16_t*);
nt=PUSH(stack, N, spx_word16_t*);
oind=PUSH(stack, N, int*);
nind=PUSH(stack, N, int*);
params = (split_cb_params *) par;
subvect_size = params->subvect_size;
nb_subvect = params->nb_subvect;
shape_cb_size = 1<<params->shape_bits;
shape_cb = params->shape_cb;
have_sign = params->have_sign;
resp = PUSH(stack, shape_cb_size*subvect_size, spx_word16_t);
#ifdef _USE_SSE
resp2 = PUSH(stack, (shape_cb_size*subvect_size)>>2, __m128);
E = PUSH(stack, shape_cb_size>>2, __m128);
#else
resp2 = resp;
E = PUSH(stack, shape_cb_size, spx_word32_t);
#endif
t = PUSH(stack, nsf, spx_word16_t);
e = PUSH(stack, nsf, spx_sig_t);
r2 = PUSH(stack, nsf, spx_sig_t);
ind = PUSH(stack, nb_subvect, int);
tmp = PUSH(stack, 2*N*nsf, spx_word16_t);
for (i=0;i<N;i++)
{
ot[i]=tmp;
tmp += nsf;
nt[i]=tmp;
tmp += nsf;
}
best_index = PUSH(stack, N, int);
best_dist = PUSH(stack, N, spx_word32_t);
ndist = PUSH(stack, N, spx_word32_t);
odist = PUSH(stack, N, spx_word32_t);
itmp = PUSH(stack, 2*N*nb_subvect, int);
for (i=0;i<N;i++)
{
nind[i]=itmp;
itmp+=nb_subvect;
oind[i]=itmp;
itmp+=nb_subvect;
for (j=0;j<nb_subvect;j++)
nind[i][j]=oind[i][j]=-1;
}
/* FIXME: make that adaptive? */
for (i=0;i<nsf;i++)
t[i]=SHR(target[i],6);
for (j=0;j<N;j++)
for (i=0;i<nsf;i++)
ot[j][i]=t[i];
/*for (i=0;i<nsf;i++)
printf ("%d\n", (int)t[i]);*/
/* Pre-compute codewords response and energy */
compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
for (j=0;j<N;j++)
odist[j]=0;
/*For all subvectors*/
for (i=0;i<nb_subvect;i++)
{
/*"erase" nbest list*/
for (j=0;j<N;j++)
ndist[j]=-2;
/*For all n-bests of previous subvector*/
for (j=0;j<N;j++)
{
spx_word16_t *x=ot[j]+subvect_size*i;
/*Find new n-best based on previous n-best j*/
if (have_sign)
vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
else
vq_nbest(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
/*For all new n-bests*/
for (k=0;k<N;k++)
{
spx_word16_t *ct;
spx_word32_t err=0;
ct = ot[j];
/*update target*/
/*previous target*/
for (m=i*subvect_size;m<(i+1)*subvect_size;m++)
t[m]=ct[m];
/* New code: update only enough of the target to calculate error*/
{
int rind;
spx_word16_t *res;
spx_word16_t sign=1;
rind = best_index[k];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
res = resp+rind*subvect_size;
if (sign>0)
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] -= res[m];
else
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] += res[m];
}
/*compute error (distance)*/
err=odist[j];
for (m=i*subvect_size;m<(i+1)*subvect_size;m++)
err += t[m]*t[m];
/*update n-best list*/
if (err<ndist[N-1] || ndist[N-1]<-1)
{
/*previous target (we don't care what happened before*/
for (m=(i+1)*subvect_size;m<nsf;m++)
t[m]=ct[m];
/* New code: update the rest of the target only if it's worth it */
for (m=0;m<subvect_size;m++)
{
spx_word16_t g;
int rind;
spx_word16_t sign=1;
rind = best_index[k];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
q=subvect_size-m;
#ifdef FIXED_POINT
g=sign*shape_cb[rind*subvect_size+m];
for (n=subvect_size*(i+1);n<nsf;n++,q++)
t[n] = SUB32(t[n],MULT16_16_Q11(g,r[q]));
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
for (n=subvect_size*(i+1);n<nsf;n++,q++)
t[n] = SUB32(t[n],g*r[q]);
#endif
}
for (m=0;m<N;m++)
{
if (err < ndist[m] || ndist[m]<-1)
{
for (n=N-1;n>m;n--)
{
for (q=(i+1)*subvect_size;q<nsf;q++)
nt[n][q]=nt[n-1][q];
for (q=0;q<nb_subvect;q++)
nind[n][q]=nind[n-1][q];
ndist[n]=ndist[n-1];
}
for (q=(i+1)*subvect_size;q<nsf;q++)
nt[m][q]=t[q];
for (q=0;q<nb_subvect;q++)
nind[m][q]=oind[j][q];
nind[m][i]=best_index[k];
ndist[m]=err;
break;
}
}
}
}
if (i==0)
break;
}
/*update old-new data*/
/* just swap pointers instead of a long copy */
{
spx_word16_t **tmp2;
tmp2=ot;
ot=nt;
nt=tmp2;
}
for (j=0;j<N;j++)
for (m=0;m<nb_subvect;m++)
oind[j][m]=nind[j][m];
for (j=0;j<N;j++)
odist[j]=ndist[j];
}
/*save indices*/
for (i=0;i<nb_subvect;i++)
{
ind[i]=nind[0][i];
speex_bits_pack(bits,ind[i],params->shape_bits+have_sign);
}
/* Put everything back together */
for (i=0;i<nb_subvect;i++)
{
int rind;
spx_word16_t sign=1;
rind = ind[i];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
#ifdef FIXED_POINT
if (sign==1)
{
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=SHL((spx_word32_t)shape_cb[rind*subvect_size+j],SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=-SHL((spx_word32_t)shape_cb[rind*subvect_size+j],SIG_SHIFT-5);
}
#else
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=sign*0.03125*shape_cb[rind*subvect_size+j];
#endif
}
/* Update excitation */
for (j=0;j<nsf;j++)
exc[j]+=e[j];
/* Update target */
syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
target[j]-=r2[j];
}
static word_type ROTL(word_type x, unsigned n) {
return SHL(x, n) | SHR(x, word_bits-n);
}
void Az_lsp (
INT16 a[], /* (i) : predictor coefficients */
INT16 lsp[], /* (o) : line spectral pairs */
INT16 old_lsp[] /* (i) : old lsp[] (in case not found 10 roots) */
)
{
INT16 i, j, nf, ip;
INT16 xlow, ylow, xhigh, yhigh, xmid, ymid, xint;
INT16 x, y, sign, exp;
INT16 *coef;
INT16 f1[M / 2 + 1], f2[M / 2 + 1];
INT32 t0=0;
VPP_EFR_PROFILE_FUNCTION_ENTER(Az_lsp);
/*-------------------------------------------------------------*
* find the sum and diff. pol. F1(z) and F2(z) *
* F1(z) <--- F1(z)/(1+z**-1) & F2(z) <--- F2(z)/(1-z**-1) *
* *
* f1[0] = 1.0; *
* f2[0] = 1.0; *
* *
* for (i = 0; i< NC; i++) *
* { *
* f1[i+1] = a[i+1] + a[M-i] - f1[i] ; *
* f2[i+1] = a[i+1] - a[M-i] + f2[i] ; *
* } *
*-------------------------------------------------------------*/
f1[0] = 1024; /* f1[0] = 1.0 */
f2[0] = 1024; /* f2[0] = 1.0 */
for (i = 0; i < NC; i++)
{
//VPP_MLX16 (t0_hi,t0_lo,a[i + 1], 8192);
//VPP_MLA16 ( t0_hi, t0_lo, a[M - i], 8192);
//t0 = VPP_SCALE64_TO_16( t0_hi, t0_lo);
//x = EXTRACT_H(t0);
t0 = (INT32) a[i + 1] + (INT32)a[M - i];
x = (INT16)(L_SHR_D(t0,2));
/* f1[i+1] = a[i+1] + a[M-i] - f1[i] */
f1[i + 1] = SUB (x, f1[i]);
//VPP_MLX16(t0_hi, t0_lo, a[i + 1], 8192);
//VPP_MLA16(t0_hi, t0_lo, a[M - i], -8192);
//x = EXTRACT_H(VPP_SCALE64_TO_16(t0_hi, t0_lo));
t0 = (INT32) a[i + 1] - (INT32)a[M - i];
x = (INT16)(L_SHR_D(t0,2));
//f2[i + 1] = add (x, f2[i]);
f2[i + 1] = ADD(x, f2[i]);
}
/*-------------------------------------------------------------*
* find the LSPs using the Chebychev pol. evaluation *
*-------------------------------------------------------------*/
nf = 0; /* number of found frequencies */
ip = 0; /* indicator for f1 or f2 */
coef = f1;
xlow = grid[0];
ylow = Chebps (xlow, coef, NC);
j = 0;
/* while ( (nf < M) && (j < grid_points) ) */
//while ((sub (nf, M) < 0) && (sub (j, grid_points) < 0))
while ((SUB (nf, M) < 0) && (SUB (j, grid_points) < 0))
{
j++;
xhigh = xlow;
yhigh = ylow;
xlow = grid[j];
ylow = Chebps (xlow, coef, NC);
//if (L_mult (ylow, yhigh) <= (INT32) 0L)
if (L_MULT(ylow, yhigh) <= (INT32) 0L)
{
/* divide 4 times the interval */
for (i = 0; i < 4; i++)
{
/* xmid = (xlow + xhigh)/2 */
// xmid = add (shr (xlow, 1), shr (xhigh, 1));
xmid = ADD ((SHR_D(xlow, 1)),(SHR_D(xhigh, 1)));
ymid = Chebps (xmid, coef, NC);
//if (L_mult (ylow, ymid) <= (INT32) 0L)
if (L_MULT(ylow, ymid) <= (INT32) 0L)
{
yhigh = ymid;
xhigh = xmid;
}
else
{
ylow = ymid;
xlow = xmid;
}
}
/*-------------------------------------------------------------*
* Linear interpolation *
* xint = xlow - ylow*(xhigh-xlow)/(yhigh-ylow); *
*-------------------------------------------------------------*/
//x = sub (xhigh, xlow);
x = SUB (xhigh, xlow);
//y = sub (yhigh, ylow);
y = SUB (yhigh, ylow);
if (y == 0)
{
xint = xlow;
}
else
{
sign = y;
//y = abs_s (y);
y = ABS_S(y);
exp = norm_s (y);
//y = shl (y, exp);
y = SHL(y, exp);
y = div_s ((INT16) 16383, y);
//t0 = L_mult (x, y);
t0 = L_MULT(x, y);
//t0 = L_shr (t0, sub (20, exp));
t0 = L_SHR_V(t0, SUB (20, exp));
//y = extract_l (t0); /* y= (xhigh-xlow)/(yhigh-ylow) */
y = EXTRACT_L(t0);
if (sign < 0)
{
//y = negate (y);
y = NEGATE(y);
}
//t0 = L_mult (ylow, y);
t0 = L_MULT(ylow, y);
//t0 = L_shr (t0, 11);
t0 = L_SHR_D(t0, 11);
//xint = sub (xlow, extract_l (t0)); /* xint = xlow - ylow*y */
xint = SUB (xlow, EXTRACT_L(t0));
}
lsp[nf] = xint;
xlow = xint;
nf++;
if (ip == 0)
{
ip = 1;
coef = f2;
}
else
{
ip = 0;
coef = f1;
}
ylow = Chebps (xlow, coef, NC);
}
}
/* Check if M roots found */
//if (sub (nf, M) < 0)
if (SUB (nf, M) < 0)
{
for (i = 0; i < M; i++)
{
lsp[i] = old_lsp[i];
}
}
VPP_EFR_PROFILE_FUNCTION_EXIT(Az_lsp);
return;
}
int crypto_aead_decrypt(
unsigned char *m,unsigned long long *mlen, // message
unsigned char *nsec, // not relavent to CLOC or SLIC
const unsigned char *c,unsigned long long clen, // ciphertext
const unsigned char *ad,unsigned long long adlen, // associated data
const unsigned char *npub, // nonce
const unsigned char *k // the master key
)
{
block estate, tstate, tmp; // encryption state, tag state, and temporary state
estate = SETZERO();
unsigned char ltag[16]; // local copy of temporary tag value
unsigned long long i, lastblocklen,j;
/* set ciphertext length */
*mlen = clen - CRYPTO_ABYTES;
/* generate round keys from master key */
AES128_KeyExpansion(k);
/* process the first (partial) block of ad */
load_partial_block(&estate, ad, (adlen>STATE_LEN)?STATE_LEN:adlen, ONE_ZERO_PADDING);
fix0(estate);
AES128_encrypt(estate, estate);
if((ad[0] & 0x80) || (adlen == 0)){
// appy h
h(estate);
}
else{
// do nothing
}
if(adlen > STATE_LEN){ // ad is of moer than one block
i = STATE_LEN;
/* process the middle ad blocks, excluding the first and last (partial) block */
while((i+STATE_LEN) < adlen)
{
tmp = LOAD(ad+i);
estate = XOR(estate, tmp);
AES128_encrypt(estate, estate);
i += STATE_LEN;
}
/* process the last (partial) ad block */
load_partial_block(&tmp, ad+i, adlen - i, ONE_ZERO_PADDING);
estate = XOR(estate, tmp);
AES128_encrypt(estate, estate);
}
/* process the nonce */
load_partial_block(&tmp, npub, CRYPTO_NPUBBYTES, PARAM_OZP);
estate = XOR(estate, tmp);
if((adlen % STATE_LEN) || (adlen == 0)){
/* apply f2 */
f2(estate);
}
else{
/* apply f1 */
f1(estate);
}
/* process ciphertext */
tstate = estate;
AES128_encrypt(estate, estate);
if(*mlen){
/* apply g2 to tag state */
g2(tstate);
}
else{
/* apply g1 to tag state */
g1(tstate);
}
AES128_encrypt(tstate, tstate);
i = 0;
/* process all the message except for the last message/ciphertext block */
while((i + STATE_LEN) < (*mlen)){
tmp = LOAD(c+i);
estate = XOR(estate, tmp);
STORE(m+i, estate);
tstate = XOR(tmp, tstate);
AES128_encrypt(tstate, tstate);
fix1(tmp);
print_state("after applying fix1\n", estate);
AES128_encrypt(tmp, estate);
i += STATE_LEN;
}
/* process the last block of the message/ciphetext */
lastblocklen = (*mlen) - i;
if(lastblocklen > 0){
load_partial_block(&tmp, c+i, lastblocklen, ZERO_APPEND);
estate = XOR(estate, tmp);
print_state("after xoring last partial message block\n", estate);
store_partial_block(m+i, estate, lastblocklen);
unsigned char shift_bytes = (STATE_LEN - (unsigned char)lastblocklen);
tmp = AND(SHR(_mm_set1_epi8(0xff), shift_bytes), tmp);
tstate = XOR(tstate, tmp);
/* add the one zero padding */
tstate = XOR(tstate, SHL(_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0x80), lastblocklen));
if((*mlen) % STATE_LEN){
/* apply f2 */
f2(tstate);
}
else{
/* apply f1 */
f1(tstate);
}
AES128_encrypt(tstate, tstate);
}
/* compare tag and output message */
STORE(ltag, tstate);
for(j = 0; j < CRYPTO_ABYTES; j++){
if(ltag[j] != c[clen - CRYPTO_ABYTES + j])
return RETURN_TAG_NO_MATCH;
}
return RETURN_SUCCESS;
}
void gainQuantization(bcg729EncoderChannelContextStruct *encoderChannelContext, word16_t targetSignal[], word16_t filteredAdaptativeCodebookVector[], word16_t convolvedFixedCodebookVector[], word16_t fixedCodebookVector[], word64_t xy64, word64_t yy64,
word16_t *quantizedAdaptativeCodebookGain, word16_t *quantizedFixedCodebookGain, uint16_t *gainCodebookStage1, uint16_t *gainCodebookStage2)
{
int i,j;
word64_t xz64=0, yz64=0, zz64=0;
word32_t xy;
word32_t yy;
word32_t xz;
word32_t yz;
word32_t zz;
uint16_t minNormalization = 31;
uint16_t currentNormalization;
word32_t bestAdaptativeCodebookGain, bestFixedCodebookGain;
word64_t denominator;
word16_t predictedFixedCodebookGain;
uint16_t indexBaseGa=0;
uint16_t indexBaseGb=0;
uint16_t indexGa=0, indexGb=0;
word64_t distanceMin = MAXINT64;
/*** compute spec 3.9 eq63 terms first on 64 bits and then scale them if needed to fit on 32 ***/
/* Xy64 and Yy64 already computed during adaptativeCodebookGain computation */
for (i=0; i<L_SUBFRAME; i++) {
xz64 = MAC64(xz64, targetSignal[i], convolvedFixedCodebookVector[i]); /* in Q12 */
yz64 = MAC64(yz64, filteredAdaptativeCodebookVector[i], convolvedFixedCodebookVector[i]); /* in Q12 */
zz64 = MAC64(zz64, convolvedFixedCodebookVector[i], convolvedFixedCodebookVector[i]); /* in Q24 */
}
/* now scale this terms to have them fit on 32 bits - terms Xy, Xz and Yz shall fit on 31 bits because used in eq63 with a factor 2 */
xy = SHR64(((xy64<0)?-xy64:xy64),30);
yy = SHR64(yy64,31);
xz = SHR64(((xz64<0)?-xz64:xz64),30);
yz = SHR64(((yz64<0)?-yz64:yz64),30);
zz = SHR64(zz64,31);
currentNormalization = countLeadingZeros(xy);
if (currentNormalization<minNormalization) {
minNormalization = currentNormalization;
}
currentNormalization = countLeadingZeros(xz);
if (currentNormalization<minNormalization) {
minNormalization = currentNormalization;
}
currentNormalization = countLeadingZeros(yz);
if (currentNormalization<minNormalization) {
minNormalization = currentNormalization;
}
currentNormalization = countLeadingZeros(yy);
if (currentNormalization<minNormalization) {
minNormalization = currentNormalization;
}
currentNormalization = countLeadingZeros(zz);
if (currentNormalization<minNormalization) {
minNormalization = currentNormalization;
}
if (minNormalization<31) { /* we shall normalise, values are over 32 bits */
minNormalization = 31 - minNormalization;
xy = (word32_t)SHR64(xy64, minNormalization);
yy = (word32_t)SHR64(yy64, minNormalization);
xz = (word32_t)SHR64(xz64, minNormalization);
yz = (word32_t)SHR64(yz64, minNormalization);
zz = (word32_t)SHR64(zz64, minNormalization);
} else { /* no need to normalise, values already fit on 32 bits, just cast them */
xy = (word32_t)xy64; /* in Q0 */
yy = (word32_t)yy64; /* in Q0 */
xz = (word32_t)xz64; /* in Q12 */
yz = (word32_t)yz64; /* in Q12 */
zz = (word32_t)zz64; /* in Q24 */
}
/*** compute the best gains minimizinq eq63 ***/
/* Note this bestgain computation is not at all described in the spec, got it from ITU code */
/* bestAdaptativeCodebookGain = (zz.xy - xz.yz) / (yy*zz) - yz^2) */
/* bestfixedCodebookGain = (yy*xz - xy*yz) / (yy*zz) - yz^2) */
/* best gain are computed in Q9 and Q2 and fits on 16 bits */
denominator = MAC64(MULT32_32(yy, zz), -yz, yz); /* (yy*zz) - yz^2) in Q24 (always >= 0)*/
/* avoid division by zero */
if (denominator==0) { /* consider it to be one */
bestAdaptativeCodebookGain = (word32_t)(SHR64(MAC64(MULT32_32(zz, xy), -xz, yz), 15)); /* MAC in Q24 -> Q9 */
bestFixedCodebookGain = (word32_t)(SHR64(MAC64(MULT32_32(yy, xz), -xy, yz), 10)); /* MAC in Q12 -> Q2 */
} else {
/* bestAdaptativeCodebookGain in Q9 */
uint16_t numeratorNorm;
word64_t numerator = MAC64(MULT32_32(zz, xy), -xz, yz); /* in Q24 */
/* check if we can shift it by 9 without overflow as the bestAdaptativeCodebookGain in computed in Q9 */
word32_t numeratorH = (word32_t)(SHR64(numerator,32));
numeratorH = (numeratorH>0)?numeratorH:-numeratorH;
numeratorNorm = countLeadingZeros(numeratorH);
if (numeratorNorm >= 9) {
bestAdaptativeCodebookGain = (word32_t)(DIV64(SHL64(numerator,9), denominator)); /* bestAdaptativeCodebookGain in Q9 */
} else {
word64_t shiftedDenominator = SHR64(denominator, 9-numeratorNorm);
if (shiftedDenominator>0) { /* can't shift left by 9 the numerator, can we shift right by 9-numeratorNorm the denominator without hiting 0 */
bestAdaptativeCodebookGain = (word32_t)(DIV64(SHL64(numerator, numeratorNorm),shiftedDenominator)); /* bestAdaptativeCodebookGain in Q9 */
} else {
bestAdaptativeCodebookGain = SHL((word32_t)(DIV64(SHL64(numerator, numeratorNorm), denominator)), 9-numeratorNorm); /* shift left the division result to reach Q9 */
}
}
numerator = MAC64(MULT32_32(yy, xz), -xy, yz); /* in Q12 */
/* check if we can shift it by 14(it's in Q12 and denominator in Q24) without overflow as the bestFixedCodebookGain in computed in Q2 */
numeratorH = (word32_t)(SHR64(numerator,32));
numeratorH = (numeratorH>0)?numeratorH:-numeratorH;
numeratorNorm = countLeadingZeros(numeratorH);
if (numeratorNorm >= 14) {
bestFixedCodebookGain = (word32_t)(DIV64(SHL64(numerator,14), denominator));
} else {
word64_t shiftedDenominator = SHR64(denominator, 14-numeratorNorm); /* bestFixedCodebookGain in Q14 */
if (shiftedDenominator>0) { /* can't shift left by 9 the numerator, can we shift right by 9-numeratorNorm the denominator without hiting 0 */
bestFixedCodebookGain = (word32_t)(DIV64(SHL64(numerator, numeratorNorm),shiftedDenominator)); /* bestFixedCodebookGain in Q14 */
} else {
bestFixedCodebookGain = SHL((word32_t)(DIV64(SHL64(numerator, numeratorNorm), denominator)), 14-numeratorNorm); /* shift left the division result to reach Q14 */
}
}
}
/*** Compute the predicted gain as in spec 3.9.1 eq71 in Q6 ***/
predictedFixedCodebookGain = (word16_t)(SHR32(MACodeGainPrediction(encoderChannelContext->previousGainPredictionError, fixedCodebookVector), 12)); /* in Q16 -> Q4 range [3,1830] */
/*** preselection spec 3.9.2 ***/
/* Note: spec just says to select the best 50% of each vector, ITU code go through magical constant computation to select the begining of a continuous range */
/* much more simple here : vector are ordened in growing order so just select 2 (4 for Gb) indexes before the first value to be superior to the best gain previously computed */
while (indexBaseGa<6 && bestFixedCodebookGain>(MULT16_16_Q14(GACodebook[indexBaseGa][1],predictedFixedCodebookGain))) { /* bestFixedCodebookGain> in Q2, GACodebook in Q12 *predictedFixedCodebookGain in Q4 -> Q16-14 */
indexBaseGa++;
}
if (indexBaseGa>0) indexBaseGa--;
if (indexBaseGa>0) indexBaseGa--;
while (indexBaseGb<12 && bestAdaptativeCodebookGain>(SHR(GBCodebook[indexBaseGb][0],5))) {
indexBaseGb++;
}
if (indexBaseGb>0) indexBaseGb--;
if (indexBaseGb>0) indexBaseGb--;
if (indexBaseGb>0) indexBaseGb--;
if (indexBaseGb>0) indexBaseGb--;
/*** test all possibilities of Ga and Gb indexes and select the best one ***/
xy = -SHL(xy,1); /* xy term is always used with a -2 factor */
xz = -SHL(xz,1); /* xz term is always used with a -2 factor */
yz = SHL(yz,1); /* yz term is always used with a 2 factor */
for (i=0; i<4; i++) {
for (j=0; j<8; j++) {
/* compute gamma->gc and gp */
word16_t gp = ADD16(GACodebook[i+indexBaseGa][0], GBCodebook[j+indexBaseGb][0]); /* result in Q14 */
word16_t gamma = ADD16(GACodebook[i+indexBaseGa][1], GBCodebook[j+indexBaseGb][1]); /* result in Q3.12 (range [0.185, 5.05])*/
word32_t gc = MULT16_16_Q14(gamma, predictedFixedCodebookGain); /* gamma in Q12, predictedFixedCodebookGain in Q4 -> Q16 -14 -> Q2 */
/* compute E as in eq63 (first term excluded) */
word64_t acc = MULT32_32(MULT16_16(gp, gp), yy); /* acc = gp^2*yy gp in Q14, yy in Q0 -> acc in Q28 */
acc = MAC64(acc, MULT16_16(gc, gc), zz); /* gc in Q2, zz in Q24 -> acc in Q28, note gc is on 32 bits but in a range making gc^2 fitting on 32 bits */
acc = MAC64(acc, SHL32((word32_t)gp, 14), xy); /* gp in Q14 shifted to Q28, xy in Q0 -> acc in Q28 */
acc = MAC64(acc, SHL32(gc, 14), xz); /* gc in Q2 shifted to Q16, xz in Q12 -> acc in Q28 */
acc = MAC64(acc, MULT16_16(gp,gc), yz); /* gp in Q14, gc in Q2 yz in Q12 -> acc in Q28 */
if (acc<distanceMin) {
distanceMin = acc;
indexGa = i+indexBaseGa;
indexGb = j+indexBaseGb;
*quantizedAdaptativeCodebookGain = gp;
*quantizedFixedCodebookGain = (word16_t)SHR(gc, 1);
}
}
}
/* update the previous gain prediction error */
computeGainPredictionError(ADD16(GACodebook[indexGa][1], GBCodebook[indexGb][1]), encoderChannelContext->previousGainPredictionError);
/* mapping of indexes */
*gainCodebookStage1 = indexMappingGA[indexGa];
*gainCodebookStage2 = indexMappingGB[indexGb];
return;
}
