tgsi: add ATOMFADD operation

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

Signed-off-by: Ilia Mirkin <[email protected]>
Reviewed-by: Marek Olšák <[email protected]>

6 files changed, 23 insertions, 2 deletions

diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index 5db515a0753..99edf33c22e 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -4253,6 +4253,9 @@ exec_atomop_mem(struct tgsi_exec_machine *mach,
       if (val == value[0].u[0])
          val = value2[0].u[0];
       break;
+   case TGSI_OPCODE_ATOMFADD:
+      val = fui(r[0].f[0] + value[0].f[0]);
+      break;
    default:
       break;
    }
@@ -5933,6 +5936,7 @@ exec_instruction(
    case TGSI_OPCODE_ATOMUMAX:
    case TGSI_OPCODE_ATOMIMIN:
    case TGSI_OPCODE_ATOMIMAX:
+   case TGSI_OPCODE_ATOMFADD:
       exec_atomop(mach, inst);
       break;
 
diff --git a/src/gallium/auxiliary/tgsi/tgsi_info_opcodes.h b/src/gallium/auxiliary/tgsi/tgsi_info_opcodes.h
index c3787c2fbb2..f391b0cea6b 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_info_opcodes.h
+++ b/src/gallium/auxiliary/tgsi/tgsi_info_opcodes.h
@@ -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, ENDLOOP, .is_branch = 1, .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, RESQ)
 OPCODE(1, 1, COMP, READ_FIRST)
diff --git a/src/gallium/auxiliary/tgsi/tgsi_scan.c b/src/gallium/auxiliary/tgsi/tgsi_scan.c
index 4ca84902dd4..d776fc7bef1 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_scan.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_scan.c
@@ -391,6 +391,7 @@ scan_instruction(struct tgsi_shader_info *info,
    case TGSI_OPCODE_ATOMUMAX:
    case TGSI_OPCODE_ATOMIMIN:
    case TGSI_OPCODE_ATOMIMAX:
+   case TGSI_OPCODE_ATOMFADD:
       if (tgsi_is_bindless_image_file(fullinst->Src[0].Register.File)) {
          info->uses_bindless_images = true;
 
diff --git a/src/gallium/auxiliary/tgsi/tgsi_util.c b/src/gallium/auxiliary/tgsi/tgsi_util.c
index ebbd561f7d0..1e5582ba273 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_util.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_util.c
@@ -385,6 +385,7 @@ tgsi_util_get_inst_usage_mask(const struct tgsi_full_instruction *inst,
    case TGSI_OPCODE_ATOMUMAX:
    case TGSI_OPCODE_ATOMIMIN:
    case TGSI_OPCODE_ATOMIMAX:
+   case TGSI_OPCODE_ATOMFADD:
       if (src_idx == 0) {
          read_mask = TGSI_WRITEMASK_XY; /* bindless handle possible */
       } else if (src_idx == 1) {
diff --git a/src/gallium/docs/source/tgsi.rst b/src/gallium/docs/source/tgsi.rst
index 7d4ebb62674..277f25ca41b 100644
--- a/src/gallium/docs/source/tgsi.rst
+++ b/src/gallium/docs/source/tgsi.rst
@@ -2684,6 +2684,21 @@ These atomic operations may only be used with 32-bit integer image formats.
   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
 
   Syntax: ``ATOMXCHG dst, resource, offset, src``
diff --git a/src/gallium/include/pipe/p_shader_tokens.h b/src/gallium/include/pipe/p_shader_tokens.h
index bef826f23b5..8f290615c3f 100644
--- a/src/gallium/include/pipe/p_shader_tokens.h
+++ b/src/gallium/include/pipe/p_shader_tokens.h
@@ -442,7 +442,7 @@ enum tgsi_opcode {
    TGSI_OPCODE_BGNSUB             = 100,
    TGSI_OPCODE_ENDLOOP            = 101,
    TGSI_OPCODE_ENDSUB             = 102,
-   /* gap */
+   TGSI_OPCODE_ATOMFADD           = 103,
    TGSI_OPCODE_TXQS               = 104,
    TGSI_OPCODE_RESQ               = 105,
    TGSI_OPCODE_READ_FIRST         = 106,