nir: Add load/store/atomic global intrinsics

These correspond roughly to reading/writing OpenCL global pointers.  The
idea is that they just take a bare address and load/store from it.  Of
course, exactly what this address means is driver-dependent.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Karol Herbst <kherbst@redhat.com>

diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py
index a5cc3f7401cea6d99101c887b1b6cb27ad315e91..90d347f7331de99f560e86b00522677f01f3aaa9 100644 (file)
--- a/src/compiler/nir/nir_intrinsics.py
+++ b/src/compiler/nir/nir_intrinsics.py
@@ -458,6 +458,34 @@ intrinsic("shared_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 intrinsic("shared_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 intrinsic("shared_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 
+# Global atomic intrinsics
+#
+# All of the shared variable atomic memory operations read a value from
+# memory, compute a new value using one of the operations below, write the
+# new value to memory, and return the original value read.
+#
+# All operations take 2 sources except CompSwap that takes 3. These
+# sources represent:
+#
+# 0: The memory address that the atomic operation will operate on.
+# 1: The data parameter to the atomic function (i.e. the value to add
+#    in shared_atomic_add, etc).
+# 2: For CompSwap only: the second data parameter.
+intrinsic("global_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
+intrinsic("global_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
+
 def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
     intrinsic("load_" + name, [], dest_comp, indices,
               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
@@ -590,6 +618,9 @@ load("shared", 1, [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }. const_index[] = { base, range }
 load("constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
+# src[] = { address }.
+# const_index[] = { access, align_mul, align_offset }
+load("global", 1, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 
 # Stores work the same way as loads, except now the first source is the value
 # to store and the second (and possibly third) source specify where to store
@@ -610,3 +641,6 @@ store("ssbo", 3, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 # src[] = { value, offset }.
 # const_index[] = { base, write_mask, align_mul, align_offset }
 store("shared", 2, [BASE, WRMASK, ALIGN_MUL, ALIGN_OFFSET])
+# src[] = { value, address }.
+# const_index[] = { write_mask, align_mul, align_offset }
+store("global", 2, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])

diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c
index 06598e054a67284a3023c6f04133123a58b3291f..f07be78aa9974cdf27973e8c08c2bfeee55c384a 100644 (file)
--- a/src/compiler/nir/nir_lower_io.c
+++ b/src/compiler/nir/nir_lower_io.c
@@ -944,6 +944,7 @@ nir_get_io_offset_src(nir_intrinsic_instr *instr)
    case nir_intrinsic_load_output:
    case nir_intrinsic_load_shared:
    case nir_intrinsic_load_uniform:
+   case nir_intrinsic_load_global:
       return &instr->src[0];
    case nir_intrinsic_load_ubo:
    case nir_intrinsic_load_ssbo:
@@ -952,6 +953,7 @@ nir_get_io_offset_src(nir_intrinsic_instr *instr)
    case nir_intrinsic_load_interpolated_input:
    case nir_intrinsic_store_output:
    case nir_intrinsic_store_shared:
+   case nir_intrinsic_store_global:
       return &instr->src[1];
    case nir_intrinsic_store_ssbo:
    case nir_intrinsic_store_per_vertex_output:

diff --git a/src/compiler/nir/nir_lower_phis_to_scalar.c b/src/compiler/nir/nir_lower_phis_to_scalar.c
index 41ae19e8391759316832d029cf7ba2d7b7bfa445..16001f736852087e79acaa3ef54dcdfef1c2dd61 100644 (file)
--- a/src/compiler/nir/nir_lower_phis_to_scalar.c
+++ b/src/compiler/nir/nir_lower_phis_to_scalar.c
@@ -88,7 +88,8 @@ is_phi_src_scalarizable(nir_phi_src *src,
          return deref->mode == nir_var_shader_in ||
                 deref->mode == nir_var_uniform ||
                 deref->mode == nir_var_mem_ubo ||
-                deref->mode == nir_var_mem_ssbo;
+                deref->mode == nir_var_mem_ssbo ||
+                deref->mode == nir_var_mem_global;
       }
 
       case nir_intrinsic_interp_deref_at_centroid:
@@ -97,6 +98,7 @@ is_phi_src_scalarizable(nir_phi_src *src,
       case nir_intrinsic_load_uniform:
       case nir_intrinsic_load_ubo:
       case nir_intrinsic_load_ssbo:
+      case nir_intrinsic_load_global:
       case nir_intrinsic_load_input:
          return true;
       default: