tgsi: add ATOMFADD operation

This is supported by at least NVIDIA hardware, and exposeable via GL
extensions.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This commit is contained in:
Ilia Mirkin 2018-04-18 21:13:22 -04:00
parent bac8534267
commit 5574414edc
6 changed files with 23 additions and 2 deletions

View File

@ -4253,6 +4253,9 @@ exec_atomop_mem(struct tgsi_exec_machine *mach,
if (val == value[0].u[0]) if (val == value[0].u[0])
val = value2[0].u[0]; val = value2[0].u[0];
break; break;
case TGSI_OPCODE_ATOMFADD:
val = fui(r[0].f[0] + value[0].f[0]);
break;
default: default:
break; break;
} }
@ -5933,6 +5936,7 @@ exec_instruction(
case TGSI_OPCODE_ATOMUMAX: case TGSI_OPCODE_ATOMUMAX:
case TGSI_OPCODE_ATOMIMIN: case TGSI_OPCODE_ATOMIMIN:
case TGSI_OPCODE_ATOMIMAX: case TGSI_OPCODE_ATOMIMAX:
case TGSI_OPCODE_ATOMFADD:
exec_atomop(mach, inst); exec_atomop(mach, inst);
break; break;

View File

@ -101,7 +101,7 @@ OPCODE(0, 0, NONE, BGNLOOP, .is_branch = 1, .post_indent = 1)
OPCODE(0, 0, NONE, BGNSUB, .post_indent = 1) OPCODE(0, 0, NONE, BGNSUB, .post_indent = 1)
OPCODE(0, 0, NONE, ENDLOOP, .is_branch = 1, .pre_dedent = 1) OPCODE(0, 0, NONE, ENDLOOP, .is_branch = 1, .pre_dedent = 1)
OPCODE(0, 0, NONE, ENDSUB, .pre_dedent = 1) OPCODE(0, 0, NONE, ENDSUB, .pre_dedent = 1)
OPCODE_GAP(103) /* removed */ OPCODE(1, 3, OTHR, ATOMFADD, .is_store = 1)
OPCODE(1, 1, OTHR, TXQS, .is_tex = 1) OPCODE(1, 1, OTHR, TXQS, .is_tex = 1)
OPCODE(1, 1, OTHR, RESQ) OPCODE(1, 1, OTHR, RESQ)
OPCODE(1, 1, COMP, READ_FIRST) OPCODE(1, 1, COMP, READ_FIRST)

View File

@ -391,6 +391,7 @@ scan_instruction(struct tgsi_shader_info *info,
case TGSI_OPCODE_ATOMUMAX: case TGSI_OPCODE_ATOMUMAX:
case TGSI_OPCODE_ATOMIMIN: case TGSI_OPCODE_ATOMIMIN:
case TGSI_OPCODE_ATOMIMAX: case TGSI_OPCODE_ATOMIMAX:
case TGSI_OPCODE_ATOMFADD:
if (tgsi_is_bindless_image_file(fullinst->Src[0].Register.File)) { if (tgsi_is_bindless_image_file(fullinst->Src[0].Register.File)) {
info->uses_bindless_images = true; info->uses_bindless_images = true;

View File

@ -385,6 +385,7 @@ tgsi_util_get_inst_usage_mask(const struct tgsi_full_instruction *inst,
case TGSI_OPCODE_ATOMUMAX: case TGSI_OPCODE_ATOMUMAX:
case TGSI_OPCODE_ATOMIMIN: case TGSI_OPCODE_ATOMIMIN:
case TGSI_OPCODE_ATOMIMAX: case TGSI_OPCODE_ATOMIMAX:
case TGSI_OPCODE_ATOMFADD:
if (src_idx == 0) { if (src_idx == 0) {
read_mask = TGSI_WRITEMASK_XY; /* bindless handle possible */ read_mask = TGSI_WRITEMASK_XY; /* bindless handle possible */
} else if (src_idx == 1) { } else if (src_idx == 1) {

View File

@ -2684,6 +2684,21 @@ These atomic operations may only be used with 32-bit integer image formats.
resource[offset] = dst_x + src_x resource[offset] = dst_x + src_x
.. opcode:: ATOMFADD - Atomic floating point addition
Syntax: ``ATOMFADD dst, resource, offset, src``
Example: ``ATOMFADD TEMP[0], BUFFER[0], TEMP[1], TEMP[2]``
The following operation is performed atomically:
.. math::
dst_x = resource[offset]
resource[offset] = dst_x + src_x
.. opcode:: ATOMXCHG - Atomic exchange .. opcode:: ATOMXCHG - Atomic exchange
Syntax: ``ATOMXCHG dst, resource, offset, src`` Syntax: ``ATOMXCHG dst, resource, offset, src``

View File

@ -442,7 +442,7 @@ enum tgsi_opcode {
TGSI_OPCODE_BGNSUB = 100, TGSI_OPCODE_BGNSUB = 100,
TGSI_OPCODE_ENDLOOP = 101, TGSI_OPCODE_ENDLOOP = 101,
TGSI_OPCODE_ENDSUB = 102, TGSI_OPCODE_ENDSUB = 102,
/* gap */ TGSI_OPCODE_ATOMFADD = 103,
TGSI_OPCODE_TXQS = 104, TGSI_OPCODE_TXQS = 104,
TGSI_OPCODE_RESQ = 105, TGSI_OPCODE_RESQ = 105,
TGSI_OPCODE_READ_FIRST = 106, TGSI_OPCODE_READ_FIRST = 106,