claunia/cuetools.net
1
0
Fork 0
mirror of https://github.com/claunia/cuetools.net.git synced 2025-12-16 18:14:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

FLACCL: was producing broken files when using -11 --fast-gpu

Browse Source
This commit is contained in:
Grigory Chudov
2013-10-15 00:22:41 -04:00
parent 1aeed2e7a8
commit 0855fbfaae
1 changed files with 34 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
CUETools.Codecs.FLACCL/flac.cl
View File

@@ -796,7 +796,7 @@ void clQuantizeLPC(
int cbits = min(51 - 2 * clz(shared.task.blocksize), shared.task.abits) - minprecision + (i - ((i >> precisions) << precisions)); int cbits = min(51 - 2 * clz(shared.task.blocksize), shared.task.abits) - minprecision + (i - ((i >> precisions) << precisions));
#if BITS_PER_SAMPLE <= 16 #if BITS_PER_SAMPLE <= 16
// Limit cbits so that 32-bit arithmetics will be enough when calculating residual // Limit cbits so that 32-bit arithmetics will be enough when calculating residual
cbits = min(cbits, clz(order) + 1 - shared.task.obits); cbits = min(cbits, clz(order + 1) + 1 - shared.task.obits);
#endif #endif
cbits = clamp(cbits, 3, 15); cbits = clamp(cbits, 3, 15);
@@ -1584,7 +1584,7 @@ void clCalcPartition32(
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
// we must ensure that psize * (t >> k) doesn't overflow; // we must ensure that psize * (t >> k) doesn't overflow;
uint4 lim = 0x07ffffffU; uint4 lim = 0x03ffffffU;
int x = tid >> 5; int x = tid >> 5;
__local uint * chunk = &res[x << 5]; __local uint * chunk = &res[x << 5];
// calc number of unary bits for each group of 32 residual samples // calc number of unary bits for each group of 32 residual samples
@@ -1984,7 +1984,6 @@ void clCalcOutputOffsets(
int firstFrame int firstFrame
) )
{ {
const int channels = 2;
__local FLACCLSubframeData ltasks[MAX_CHANNELS]; __local FLACCLSubframeData ltasks[MAX_CHANNELS];
__local volatile int mypos[MAX_CHANNELS]; __local volatile int mypos[MAX_CHANNELS];
int offset = 0; int offset = 0;
@@ -2153,15 +2152,17 @@ void clRiceEncoding(
{ {
int offs = pos + tid; int offs = pos + tid;
int iv = residual[task.residualOffs + offs]; int iv = residual[task.residualOffs + offs];
int part = offs / plen; // >> plenoffs; int part = offs / plen;
//int part = offs >> plenoffs;
#if 0 #if 0
int k = brp[part]; int k = brp[part];
#else #else
int k = best_rice_parameters[(get_group_id(0) << max_porder) + part]; int k = best_rice_parameters[(get_group_id(0) << max_porder) + part];
#endif #endif
int pstart = offs == task.residualOrder || offs == part * plen; int pstart = offs == part * plen;
//int pstart = offs == part << plenoffs;
uint v = (iv << 1) ^ (iv >> 31); uint v = (iv << 1) ^ (iv >> 31);
int mylen = select(0, (int)(v >> k) + 1 + k + select(0, RICE_PARAM_BITS, pstart), offs >= task.residualOrder && offs < bs); int mylen = select(0, (int)(v >> k) + 1 + k, offs >= task.residualOrder && offs < bs) + select(0, RICE_PARAM_BITS, pstart);
mypos[tid] = mylen; mypos[tid] = mylen;
// Inclusive scan(+) // Inclusive scan(+)
@@ -2187,8 +2188,6 @@ void clRiceEncoding(
// printf("Oops: %d\n", mypos[tid]); // printf("Oops: %d\n", mypos[tid]);
data[tid] = select(0U, remainder, tid == 0); data[tid] = select(0U, remainder, tid == 0);
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (mylen)
{
if (pstart) if (pstart)
{ {
int kpos = mp - mylen; int kpos = mp - mylen;
@@ -2200,6 +2199,8 @@ void clRiceEncoding(
if (kval0) atomic_or(&data[kpos0], kval0); if (kval0) atomic_or(&data[kpos0], kval0);
if (kpos1 && kval1) atomic_or(&data[kpos0 + 1], kval1); if (kpos1 && kval1) atomic_or(&data[kpos0 + 1], kval1);
} }
if (offs >= task.residualOrder && offs < bs)
{
int qpos = mp - k - 1; int qpos = mp - k - 1;
int qpos0 = (qpos >> 5) - start32; int qpos0 = (qpos >> 5) - start32;
int qpos1 = qpos & 31; int qpos1 = qpos & 31;
@@ -2221,9 +2222,9 @@ void clRiceEncoding(
int part = offs / plen; // >> plenoffs; int part = offs / plen; // >> plenoffs;
//int k = brp[min(255, part)]; //int k = brp[min(255, part)];
int k = offs < bs ? best_rice_parameters[(get_group_id(0) << max_porder) + part] : 0; int k = offs < bs ? best_rice_parameters[(get_group_id(0) << max_porder) + part] : 0;
int pstart = offs == task.residualOrder || offs == part * plen; int pstart = offs == part * plen;
uint v = (iv << 1) ^ (iv >> 31); uint v = (iv << 1) ^ (iv >> 31);
int mylen = select(0, (int)(v >> k) + 1 + k + select(0, RICE_PARAM_BITS, pstart), offs >= task.residualOrder && offs < bs); int mylen = select(0, (int)(v >> k) + 1 + k, offs >= task.residualOrder && offs < bs) + select(0, RICE_PARAM_BITS, pstart);
mypos[tid] = mylen; mypos[tid] = mylen;
// Inclusive scan(+) // Inclusive scan(+)
@@ -2248,8 +2249,6 @@ void clRiceEncoding(
// printf("Oops: %d\n", mypos[tid]); // printf("Oops: %d\n", mypos[tid]);
data[tid] = select(0U, remainder, tid == 0); data[tid] = select(0U, remainder, tid == 0);
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (mylen)
{
if (pstart) if (pstart)
{ {
int kpos = mp - mylen; int kpos = mp - mylen;
@@ -2261,6 +2260,8 @@ void clRiceEncoding(
if (kval0) atomic_or(&data[kpos0], kval0); if (kval0) atomic_or(&data[kpos0], kval0);
if (kpos1 && kval1) atomic_or(&data[kpos0 + 1], kval1); if (kpos1 && kval1) atomic_or(&data[kpos0 + 1], kval1);
} }
if (offs >= task.residualOrder && offs < bs)
{
int qpos = mp - k - 1; int qpos = mp - k - 1;
int qpos0 = (qpos >> 5) - start32; int qpos0 = (qpos >> 5) - start32;
int qpos1 = qpos & 31; int qpos1 = qpos & 31;