mirrors/pipewire
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pipewire/pipewire.git synced 2025-11-05 13:30:02 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

mem: align memory to requested alignment

Browse source 
Improve the allocators to always align the buffer memory to the
requested alignment
Use aligned read and writes for sse functions and check alignment,
optionally falling back to unaligned path.
Add more tests and benchmark cases
Check and warn for misaligned memory in plugins.
This commit is contained in:
Wim Taymans 2019-01-24 18:28:52 +01:00
parent dd66469570
commit 13bf70a8dd
19 changed files with 736 additions and 516 deletions
Download patch file Download diff file Expand all files Collapse all files

7
src/modules/module-audio-dsp/audio-dsp.c
View file

@ -69,7 +69,7 @@ struct port {
struct spa_handle *spa_handle;
struct spa_node *spa_node;
float empty[MAX_BUFFER_SIZE];
float empty[MAX_BUFFER_SIZE + 15];
};
struct node {
@ -101,14 +101,15 @@ static void init_buffer(struct port *port, uint32_t id)
b->datas[0].flags = 0;
b->datas[0].fd = -1;
b->datas[0].mapoffset = 0;
b->datas[0].maxsize = sizeof(port->empty);
b->datas[0].data = port->empty;
b->datas[0].maxsize = SPA_ROUND_DOWN_N(sizeof(port->empty), 16);
b->datas[0].data = SPA_PTR_ALIGN(port->empty, 16, void);
b->datas[0].chunk = b->chunk;
b->datas[0].chunk->offset = 0;
b->datas[0].chunk->size = 0;
b->datas[0].chunk->stride = 0;
port->bufs[id] = &b->buf;
memset(port->empty, 0, sizeof(port->empty));
pw_log_debug("%p %d", b->datas[0].data, b->datas[0].maxsize);
}
static void init_port(struct port *p, enum spa_direction direction)

38
src/modules/module-audio-dsp/floatmix.c
View file

@ -109,7 +109,7 @@ struct impl {
uint32_t stride;
bool started;
float empty[MAX_SAMPLES];
float empty[MAX_SAMPLES + 15];
};
#define CHECK_FREE_IN_PORT(this,d,p) ((d) == SPA_DIRECTION_INPUT && (p) < MAX_PORTS && !this->in_ports[(p)].valid)
@ -632,10 +632,12 @@ impl_node_port_use_buffers(struct spa_node *node,
if (!((d[0].type == SPA_DATA_MemPtr ||
d[0].type == SPA_DATA_MemFd ||
d[0].type == SPA_DATA_DmaBuf) && d[0].data != NULL)) {
spa_log_error(this->log, NAME " %p: invalid memory on buffer %p", this,
buffers[i]);
spa_log_error(this->log, NAME " %p: invalid memory on buffer %d", this, i);
return -EINVAL;
}
if (!SPA_IS_ALIGNED(d[0].data, 16)) {
spa_log_warn(this->log, NAME " %p: memory on buffer %d not aligned", this, i);
}
if (direction == SPA_DIRECTION_OUTPUT)
queue_buffer(this, port, b);
}
@ -717,23 +719,27 @@ impl_node_port_send_command(struct spa_node *node,
#include <xmmintrin.h>
static void mix_2(float *dst, float *src1, float *src2, int n_samples)
{
int i, unrolled;
int n, unrolled;
__m128 in[2];
unrolled = n_samples / 4;
n_samples &= 3;
if (SPA_IS_ALIGNED(src1, 16) &&
SPA_IS_ALIGNED(src2, 16) &&
SPA_IS_ALIGNED(dst, 16))
unrolled = n_samples / 4;
else
unrolled = 0;
for (i = 0; unrolled--; i += 4) {
in[0] = _mm_loadu_ps(&src1[i]),
in[1] = _mm_loadu_ps(&src2[i]),
for (n = 0; unrolled--; n += 4) {
in[0] = _mm_load_ps(&src1[n]),
in[1] = _mm_load_ps(&src2[n]),
in[0] = _mm_add_ps(in[0], in[1]);
_mm_storeu_ps(&dst[i], in[0]);
_mm_store_ps(&dst[n], in[0]);
}
for (; n_samples--; i++) {
in[0] = _mm_load_ss(&src1[i]),
in[1] = _mm_load_ss(&src2[i]),
for (; n < n_samples; n++) {
in[0] = _mm_load_ss(&src1[n]),
in[1] = _mm_load_ss(&src2[n]),
in[0] = _mm_add_ss(in[0], in[1]);
_mm_store_ss(&dst[i], in[0]);
_mm_store_ss(&dst[n], in[0]);
}
}
#else
@ -825,13 +831,13 @@ static int impl_node_process(struct spa_node *node)
outb->buffer->n_datas = 1;
outb->buffer->datas = outb->datas;
outb->datas[0].data = this->empty;
outb->datas[0].data = SPA_PTR_ALIGN(this->empty, 16, void);
outb->datas[0].chunk = outb->chunk;
outb->datas[0].chunk->offset = 0;
outb->datas[0].chunk->size = n_samples * sizeof(float);
outb->datas[0].chunk->stride = sizeof(float);
dst = this->empty;
dst = outb->datas[0].data;
if (n_buffers == 0) {
memset(dst, 0, n_samples * sizeof(float));
}

2
src/modules/module-client-node/client-node.c
View file

@ -837,7 +837,7 @@ do_port_use_buffers(struct impl *impl,
data_size = 0;
for (j = 0; j < buffers[i]->n_metas; j++) {
data_size += buffers[i]->metas[j].size;
data_size += SPA_ROUND_UP_N(buffers[i]->metas[j].size, 8);
}
for (j = 0; j < buffers[i]->n_datas; j++) {
struct spa_data *d = buffers[i]->datas;

25
src/pipewire/link.c
View file

@ -419,6 +419,7 @@ static int alloc_buffers(struct pw_link *this,
uint32_t n_datas,
size_t *data_sizes,
ssize_t *data_strides,
size_t *data_aligns,
struct allocation *allocation)
{
int res;
@ -452,12 +453,13 @@ static int alloc_buffers(struct pw_link *this,
metas[n_metas].type = type;
metas[n_metas].size = size;
meta_size += metas[n_metas].size;
meta_size += SPA_ROUND_UP_N(metas[n_metas].size, 8);
n_metas++;
skel_size += sizeof(struct spa_meta);
}
}
data_size += meta_size;
data_size = SPA_ROUND_UP_N(data_size, data_aligns[0]);
/* data */
for (i = 0; i < n_datas; i++) {
@ -492,7 +494,7 @@ static int alloc_buffers(struct pw_link *this,
m->type = metas[j].type;
m->size = metas[j].size;
m->data = p;
p = SPA_MEMBER(p, m->size, void);
p = SPA_MEMBER(p, SPA_ROUND_UP_N(m->size, 8), void);
}
/* pointer to data structure */
b->n_datas = n_datas;
@ -509,7 +511,7 @@ static int alloc_buffers(struct pw_link *this,
d->type = SPA_DATA_MemFd;
d->flags = 0;
d->fd = m->fd;
d->mapoffset = SPA_PTRDIFF(ddp, m->ptr);
d->mapoffset = SPA_ROUND_UP_N(SPA_PTRDIFF(ddp, m->ptr), data_aligns[i]);
d->maxsize = data_sizes[j];
d->data = SPA_MEMBER(m->ptr, d->mapoffset, void);
d->chunk->offset = 0;
@ -701,9 +703,10 @@ static int do_allocation(struct pw_link *this, uint32_t in_state, uint32_t out_s
struct spa_pod_builder b = SPA_POD_BUILDER_INIT(buffer, sizeof(buffer));
uint32_t i, offset, n_params;
uint32_t max_buffers;
size_t minsize = 8192, stride = 0;
size_t minsize = 8192, stride = 0, align;
size_t data_sizes[1];
ssize_t data_strides[1];
size_t data_aligns[1];
n_params = param_filter(this, input, output, SPA_PARAM_Buffers, &b);
n_params += param_filter(this, input, output, SPA_PARAM_Meta, &b);
@ -720,25 +723,29 @@ static int do_allocation(struct pw_link *this, uint32_t in_state, uint32_t out_s
max_buffers = MAX_BUFFERS;
minsize = stride = 0;
align = 8;
param = find_param(params, n_params, SPA_TYPE_OBJECT_ParamBuffers);
if (param) {
uint32_t qmax_buffers = max_buffers,
qminsize = minsize, qstride = stride;
qminsize = minsize, qstride = stride, qalign = align;
spa_pod_parse_object(param,
SPA_TYPE_OBJECT_ParamBuffers, NULL,
SPA_PARAM_BUFFERS_buffers, SPA_POD_Int(&qmax_buffers),
SPA_PARAM_BUFFERS_size, SPA_POD_Int(&qminsize),
SPA_PARAM_BUFFERS_stride, SPA_POD_Int(&qstride));
SPA_PARAM_BUFFERS_stride, SPA_POD_Int(&qstride),
SPA_PARAM_BUFFERS_align, SPA_POD_Int(&qalign));
max_buffers =
qmax_buffers == 0 ? max_buffers : SPA_MIN(qmax_buffers,
max_buffers);
minsize = SPA_MAX(minsize, qminsize);
stride = SPA_MAX(stride, qstride);
align = SPA_MAX(align, qalign);
pw_log_debug("%d %d %d -> %zd %zd %d", qminsize, qstride, qmax_buffers,
minsize, stride, max_buffers);
pw_log_debug("%d %d %d %d -> %zd %zd %d %zd",
qminsize, qstride, qmax_buffers, qalign,
minsize, stride, max_buffers, align);
} else {
pw_log_warn("no buffers param");
minsize = 8192;
@ -754,6 +761,7 @@ static int do_allocation(struct pw_link *this, uint32_t in_state, uint32_t out_s
data_sizes[0] = minsize;
data_strides[0] = stride;
data_aligns[0] = align;
if ((res = alloc_buffers(this,
max_buffers,
@ -761,6 +769,7 @@ static int do_allocation(struct pw_link *this, uint32_t in_state, uint32_t out_s
params,
1,
data_sizes, data_strides,
data_aligns,
&allocation)) < 0) {
asprintf(&error, "error alloc buffers: %d", res);
goto error;

2
src/pipewire/remote.c
View file

@ -1075,7 +1075,7 @@ client_node_port_use_buffers(void *object,
struct spa_meta *m = &b->metas[j];
memcpy(m, &buffers[i].buffer->metas[j], sizeof(struct spa_meta));
m->data = SPA_MEMBER(bmem.map.ptr, offset, void);
offset += m->size;
offset += SPA_ROUND_UP_N(m->size, 8);
}
for (j = 0; j < b->n_datas; j++) {