KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

gallium/tgsi_exec: Reuse the atomic helper for SSBO atomics. 

I needed to refactor SSBOs to deal with a buffer overflow, and it's easier
to just delete the SSBO atomic code from the interface.

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8276>
This commit is contained in:
Eric Anholt 2020-12-30 11:59:15 -08:00 committed by Marge Bot
parent 85be1f867c
commit 89ec36457e
3 changed files with 42 additions and 236 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
src/gallium/auxiliary/tgsi/tgsi_exec.c
View File

@ -4230,67 +4230,8 @@ exec_atomop_img(struct tgsi_exec_machine *mach,
}
static void
exec_atomop_buf(struct tgsi_exec_machine *mach,
const struct tgsi_full_instruction *inst)
{
union tgsi_exec_channel r[4];
union tgsi_exec_channel value[4], value2[4];
float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE];
float rgba2[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE];
struct tgsi_buffer_params params;
int i, j;
uint unit, chan;
int kilmask = mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0];
unit = fetch_sampler_unit(mach, inst, 0);
params.execmask = mach->ExecMask & mach->NonHelperMask & ~kilmask;
params.unit = unit;
params.writemask = inst->Dst[0].Register.WriteMask;
IFETCH(&r[0], 1, TGSI_CHAN_X);
for (i = 0; i < 4; i++) {
FETCH(&value[i], 2, TGSI_CHAN_X + i);
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
FETCH(&value2[i], 3, TGSI_CHAN_X + i);
}
for (j = 0; j < TGSI_QUAD_SIZE; j++) {
rgba[0][j] = value[0].f[j];
rgba[1][j] = value[1].f[j];
rgba[2][j] = value[2].f[j];
rgba[3][j] = value[3].f[j];
}
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
for (j = 0; j < TGSI_QUAD_SIZE; j++) {
rgba2[0][j] = value2[0].f[j];
rgba2[1][j] = value2[1].f[j];
rgba2[2][j] = value2[2].f[j];
rgba2[3][j] = value2[3].f[j];
}
}
mach->Buffer->op(mach->Buffer, &params, inst->Instruction.Opcode,
r[0].i,
rgba, rgba2);
for (j = 0; j < TGSI_QUAD_SIZE; j++) {
r[0].f[j] = rgba[0][j];
r[1].f[j] = rgba[1][j];
r[2].f[j] = rgba[2][j];
r[3].f[j] = rgba[3][j];
}
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
if (inst->Dst[0].Register.WriteMask & (1 << chan)) {
store_dest(mach, &r[chan], &inst->Dst[0], inst, chan, TGSI_EXEC_DATA_FLOAT);
}
}
}
static void
exec_atomop_mem(struct tgsi_exec_machine *mach,
const struct tgsi_full_instruction *inst)
exec_atomop_membuf(struct tgsi_exec_machine *mach,
const struct tgsi_full_instruction *inst)
{
union tgsi_exec_channel offset, r0, r1;
uint chan, i;
@ -4302,11 +4243,25 @@ exec_atomop_mem(struct tgsi_exec_machine *mach,
return;
void *ptr[TGSI_QUAD_SIZE];
for (i = 0; i < TGSI_QUAD_SIZE; i++) {
if (likely(mach->LocalMemSize >= 4 && offset.u[i] <= mach->LocalMemSize - 4))
ptr[i] = (char *)mach->LocalMem + offset.u[i];
else
ptr[i] = NULL;
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
uint32_t unit = fetch_sampler_unit(mach, inst, 0);
uint32_t size;
char *buffer = mach->Buffer->lookup(mach->Buffer, unit, &size);
for (int i = 0; i < TGSI_QUAD_SIZE; i++) {
if (likely(size >= 4 && offset.u[i] <= size - 4))
ptr[i] = buffer + offset.u[i];
else
ptr[i] = NULL;
}
} else {
assert(inst->Src[0].Register.File == TGSI_FILE_MEMORY);
for (i = 0; i < TGSI_QUAD_SIZE; i++) {
if (likely(mach->LocalMemSize >= 4 && offset.u[i] <= mach->LocalMemSize - 4))
ptr[i] = (char *)mach->LocalMem + offset.u[i];
else
ptr[i] = NULL;
}
}
FETCH(&r0, 2, TGSI_CHAN_X);
@ -4381,10 +4336,8 @@ exec_atomop(struct tgsi_exec_machine *mach,
{
if (inst->Src[0].Register.File == TGSI_FILE_IMAGE)
exec_atomop_img(mach, inst);
else if (inst->Src[0].Register.File == TGSI_FILE_BUFFER)
exec_atomop_buf(mach, inst);
else if (inst->Src[0].Register.File == TGSI_FILE_MEMORY)
exec_atomop_mem(mach, inst);
else
exec_atomop_membuf(mach, inst);
}
static void

8
src/gallium/auxiliary/tgsi/tgsi_exec.h
View File

@ -165,12 +165,8 @@ struct tgsi_buffer {
const int s[TGSI_QUAD_SIZE],
float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
void (*op)(const struct tgsi_buffer *buffer,
const struct tgsi_buffer_params *params,
enum tgsi_opcode opcode,
const int s[TGSI_QUAD_SIZE],
float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE],
float rgba2[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
void *(*lookup)(const struct tgsi_buffer *buffer,
uint32_t unit, uint32_t *size);
void (*get_dims)(const struct tgsi_buffer *buffer,
const struct tgsi_buffer_params *params,

177
src/gallium/drivers/softpipe/sp_buffer.c
View File

@ -142,169 +142,26 @@ sp_tgsi_store(const struct tgsi_buffer *buffer,
}
}
/*
* Implement atomic operations on unsigned integers.
*/
static void
handle_op_atomic(const struct pipe_shader_buffer *bview,
bool just_read,
unsigned char *data_ptr,
uint qi,
enum tgsi_opcode opcode,
unsigned writemask,
float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE],
float rgba2[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE])
{
uint c = 0; /* SSBO atomics are always on the .x channel. */
unsigned sdata;
memcpy(&sdata, data_ptr + (c * 4), 4);
if (just_read) {
((uint32_t *)rgba[c])[qi] = sdata;
return;
}
switch (opcode) {
case TGSI_OPCODE_ATOMUADD: {
unsigned temp = sdata;
sdata += ((uint32_t *)rgba[c])[qi];
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMXCHG: {
unsigned temp = sdata;
sdata = ((uint32_t *)rgba[c])[qi];
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMCAS: {
unsigned dst_x = sdata;
unsigned cmp_x = ((uint32_t *)rgba[c])[qi];
unsigned src_x = ((uint32_t *)rgba2[c])[qi];
unsigned temp = sdata;
sdata = (dst_x == cmp_x) ? src_x : dst_x;
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMAND: {
unsigned temp = sdata;
sdata &= ((uint32_t *)rgba[c])[qi];
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMOR: {
unsigned temp = sdata;
sdata |= ((uint32_t *)rgba[c])[qi];
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMXOR: {
unsigned temp = sdata;
sdata ^= ((uint32_t *)rgba[c])[qi];
((uint32_t *)rgba[c])[qi] = temp;
break;
}
case TGSI_OPCODE_ATOMUMIN: {
unsigned dst_x = sdata;
unsigned src_x = ((uint32_t *)rgba[c])[qi];
sdata = MIN2(dst_x, src_x);
((uint32_t *)rgba[c])[qi] = dst_x;
break;
}
case TGSI_OPCODE_ATOMUMAX: {
unsigned dst_x = sdata;
unsigned src_x = ((uint32_t *)rgba[c])[qi];
sdata = MAX2(dst_x, src_x);
((uint32_t *)rgba[c])[qi] = dst_x;
break;
}
case TGSI_OPCODE_ATOMIMIN: {
int dst_x = sdata;
int src_x = ((uint32_t *)rgba[c])[qi];
sdata = MIN2(dst_x, src_x);
((uint32_t *)rgba[c])[qi] = dst_x;
break;
}
case TGSI_OPCODE_ATOMIMAX: {
int dst_x = sdata;
int src_x = ((uint32_t *)rgba[c])[qi];
sdata = MAX2(dst_x, src_x);
((uint32_t *)rgba[c])[qi] = dst_x;
break;
}
case TGSI_OPCODE_ATOMFADD: {
float temp = uif(sdata);
sdata = fui(temp + rgba[c][qi]);
rgba[c][qi] = temp;
break;
}
default:
assert(!"Unexpected TGSI opcode in sp_tgsi_op");
break;
}
if (writemask & TGSI_WRITEMASK_X) {
memcpy(data_ptr + (c * 4), &sdata, 4);
}
}
/*
* Implement atomic buffer operations.
*/
static void
sp_tgsi_op(const struct tgsi_buffer *buffer,
const struct tgsi_buffer_params *params,
enum tgsi_opcode opcode,
const int s[TGSI_QUAD_SIZE],
float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE],
float rgba2[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE])
static void *
sp_tgsi_ssbo_lookup(const struct tgsi_buffer *buffer,
uint32_t unit,
uint32_t *size)
{
struct sp_tgsi_buffer *sp_buf = (struct sp_tgsi_buffer *)buffer;
struct pipe_shader_buffer *bview;
struct softpipe_resource *spr;
unsigned width;
int j, c;
unsigned char *data_ptr;
if (params->unit >= PIPE_MAX_SHADER_BUFFERS)
return;
bview = &sp_buf->sp_bview[params->unit];
spr = softpipe_resource(bview->buffer);
if (!spr)
goto fail_write_all_zero;
if (!get_dimensions(bview, spr, &width))
goto fail_write_all_zero;
for (j = 0; j < TGSI_QUAD_SIZE; j++) {
int s_coord;
bool just_read = false;
s_coord = s[j];
if (s_coord >= width) {
for (c = 0; c < 4; c++) {
rgba[c][j] = 0;
}
continue;
}
/* just readback the value for atomic if execmask isn't set */
if (!(params->execmask & (1 << j))) {
just_read = true;
}
data_ptr = (unsigned char *)spr->data + bview->buffer_offset + s_coord;
/* we should see atomic operations on r32 formats */
handle_op_atomic(bview, just_read, data_ptr, j,
opcode, params->writemask, rgba, rgba2);
if (unit >= PIPE_MAX_SHADER_BUFFERS) {
*size = 0;
return NULL;
}
return;
fail_write_all_zero:
memset(rgba, 0, TGSI_NUM_CHANNELS * TGSI_QUAD_SIZE * 4);
return;
struct pipe_shader_buffer *bview = &sp_buf->sp_bview[unit];
struct softpipe_resource *spr = softpipe_resource(bview->buffer);
if (!spr || !get_dimensions(bview, spr, size)) {
*size = 0;
return NULL;
}
return (char *)spr->data + bview->buffer_offset;
}
/*
@ -339,7 +196,7 @@ sp_create_tgsi_buffer(void)
buf->base.load = sp_tgsi_load;
buf->base.store = sp_tgsi_store;
buf->base.op = sp_tgsi_op;
buf->base.lookup = sp_tgsi_ssbo_lookup;
buf->base.get_dims = sp_tgsi_get_dims;
return buf;
};