From 69ce7baa9e0dd598aad80f559719fdf14928038e Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Mon, 9 Apr 2018 18:36:22 -0700 Subject: [PATCH] nir: Add floating point atomic add instrinsics Signed-off-by: Ian Romanick Reviewed-by: Caio Marcelo de Oliveira Filho --- .../glsl/gl_nir_lower_samplers_as_deref.c | 1 + src/compiler/glsl/glsl_to_nir.cpp | 18 +++++++++++++----- src/compiler/nir/nir_intrinsics.py | 4 ++++ src/compiler/nir/nir_lower_atomics_to_ssbo.c | 1 + src/compiler/nir/nir_lower_io.c | 3 +++ 5 files changed, 22 insertions(+), 5 deletions(-) diff --git a/src/compiler/glsl/gl_nir_lower_samplers_as_deref.c b/src/compiler/glsl/gl_nir_lower_samplers_as_deref.c index aae64c9f3f9..9ff5708f503 100644 --- a/src/compiler/glsl/gl_nir_lower_samplers_as_deref.c +++ b/src/compiler/glsl/gl_nir_lower_samplers_as_deref.c @@ -240,6 +240,7 @@ lower_intrinsic(nir_intrinsic_instr *instr, instr->intrinsic == nir_intrinsic_image_deref_atomic_xor || instr->intrinsic == nir_intrinsic_image_deref_atomic_exchange || instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap || + instr->intrinsic == nir_intrinsic_image_deref_atomic_fadd || instr->intrinsic == nir_intrinsic_image_deref_size) { b->cursor = nir_before_instr(&instr->instr); diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp index b1d1da89111..aba8d70a8a0 100644 --- a/src/compiler/glsl/glsl_to_nir.cpp +++ b/src/compiler/glsl/glsl_to_nir.cpp @@ -650,7 +650,9 @@ nir_visitor::visit(ir_call *ir) op = nir_intrinsic_image_deref_store; break; case ir_intrinsic_image_atomic_add: - op = nir_intrinsic_image_deref_atomic_add; + op = ir->return_deref->type->is_integer_32_64() + ? nir_intrinsic_image_deref_atomic_add + : nir_intrinsic_image_deref_atomic_fadd; break; case ir_intrinsic_image_atomic_min: op = nir_intrinsic_image_deref_atomic_min; @@ -689,7 +691,8 @@ nir_visitor::visit(ir_call *ir) op = nir_intrinsic_load_ssbo; break; case ir_intrinsic_ssbo_atomic_add: - op = nir_intrinsic_ssbo_atomic_add; + op = ir->return_deref->type->is_integer_32_64() + ? nir_intrinsic_ssbo_atomic_add : nir_intrinsic_ssbo_atomic_fadd; break; case ir_intrinsic_ssbo_atomic_and: op = nir_intrinsic_ssbo_atomic_and; @@ -755,7 +758,9 @@ nir_visitor::visit(ir_call *ir) op = nir_intrinsic_store_shared; break; case ir_intrinsic_shared_atomic_add: - op = nir_intrinsic_shared_atomic_add; + op = ir->return_deref->type->is_integer_32_64() + ? nir_intrinsic_shared_atomic_add + : nir_intrinsic_shared_atomic_fadd; break; case ir_intrinsic_shared_atomic_and: op = nir_intrinsic_shared_atomic_and; @@ -865,6 +870,7 @@ nir_visitor::visit(ir_call *ir) case nir_intrinsic_image_deref_atomic_xor: case nir_intrinsic_image_deref_atomic_exchange: case nir_intrinsic_image_deref_atomic_comp_swap: + case nir_intrinsic_image_deref_atomic_fadd: case nir_intrinsic_image_deref_samples: case nir_intrinsic_image_deref_size: { nir_ssa_undef_instr *instr_undef = @@ -1035,7 +1041,8 @@ nir_visitor::visit(ir_call *ir) case nir_intrinsic_ssbo_atomic_or: case nir_intrinsic_ssbo_atomic_xor: case nir_intrinsic_ssbo_atomic_exchange: - case nir_intrinsic_ssbo_atomic_comp_swap: { + case nir_intrinsic_ssbo_atomic_comp_swap: + case nir_intrinsic_ssbo_atomic_fadd: { int param_count = ir->actual_parameters.length(); assert(param_count == 3 || param_count == 4); @@ -1118,7 +1125,8 @@ nir_visitor::visit(ir_call *ir) case nir_intrinsic_shared_atomic_or: case nir_intrinsic_shared_atomic_xor: case nir_intrinsic_shared_atomic_exchange: - case nir_intrinsic_shared_atomic_comp_swap: { + case nir_intrinsic_shared_atomic_comp_swap: + case nir_intrinsic_shared_atomic_fadd: { int param_count = ir->actual_parameters.length(); assert(param_count == 2 || param_count == 3); diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py index d688a57181b..836cb34d74a 100644 --- a/src/compiler/nir/nir_intrinsics.py +++ b/src/compiler/nir/nir_intrinsics.py @@ -307,6 +307,7 @@ intrinsic("image_deref_atomic_or", src_comp=[1, 4, 1, 1], dest_comp=1) intrinsic("image_deref_atomic_xor", src_comp=[1, 4, 1, 1], dest_comp=1) intrinsic("image_deref_atomic_exchange", src_comp=[1, 4, 1, 1], dest_comp=1) intrinsic("image_deref_atomic_comp_swap", src_comp=[1, 4, 1, 1, 1], dest_comp=1) +intrinsic("image_deref_atomic_fadd", src_comp=[1, 4, 1, 1], dest_comp=1) intrinsic("image_deref_size", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER]) intrinsic("image_deref_samples", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER]) @@ -357,6 +358,7 @@ intrinsic("deref_atomic_or", src_comp=[1, 1], dest_comp=1) intrinsic("deref_atomic_xor", src_comp=[1, 1], dest_comp=1) intrinsic("deref_atomic_exchange", src_comp=[1, 1], dest_comp=1) intrinsic("deref_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1) +intrinsic("deref_atomic_fadd", src_comp=[1, 1], dest_comp=1) # SSBO atomic intrinsics # @@ -383,6 +385,7 @@ intrinsic("ssbo_atomic_or", src_comp=[1, 1, 1], dest_comp=1) intrinsic("ssbo_atomic_xor", src_comp=[1, 1, 1], dest_comp=1) intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1) intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1) +intrinsic("ssbo_atomic_fadd", src_comp=[1, 1, 1], dest_comp=1) # CS shared variable atomic intrinsics # @@ -408,6 +411,7 @@ intrinsic("shared_atomic_or", src_comp=[1, 1], dest_comp=1, indices=[BASE]) intrinsic("shared_atomic_xor", src_comp=[1, 1], dest_comp=1, indices=[BASE]) intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE]) intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE]) +intrinsic("shared_atomic_fadd", src_comp=[1, 1], dest_comp=1, indices=[BASE]) def system_value(name, dest_comp, indices=[]): intrinsic("load_" + name, [], dest_comp, indices, diff --git a/src/compiler/nir/nir_lower_atomics_to_ssbo.c b/src/compiler/nir/nir_lower_atomics_to_ssbo.c index 6ebd3632288..72686624329 100644 --- a/src/compiler/nir/nir_lower_atomics_to_ssbo.c +++ b/src/compiler/nir/nir_lower_atomics_to_ssbo.c @@ -57,6 +57,7 @@ lower_instr(nir_intrinsic_instr *instr, unsigned ssbo_offset, nir_builder *b) case nir_intrinsic_ssbo_atomic_xor: case nir_intrinsic_ssbo_atomic_exchange: case nir_intrinsic_ssbo_atomic_comp_swap: + case nir_intrinsic_ssbo_atomic_fadd: case nir_intrinsic_store_ssbo: case nir_intrinsic_load_ssbo: case nir_intrinsic_get_buffer_size: diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c index 9500a3091c4..3d4451e3602 100644 --- a/src/compiler/nir/nir_lower_io.c +++ b/src/compiler/nir/nir_lower_io.c @@ -280,6 +280,7 @@ lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state, OP(atomic_and) OP(atomic_or) OP(atomic_xor) + OP(atomic_fadd) #undef OP default: unreachable("Invalid atomic"); @@ -380,6 +381,7 @@ nir_lower_io_block(nir_block *block, case nir_intrinsic_deref_atomic_xor: case nir_intrinsic_deref_atomic_exchange: case nir_intrinsic_deref_atomic_comp_swap: + case nir_intrinsic_deref_atomic_fadd: /* We can lower the io for this nir instrinsic */ break; case nir_intrinsic_interp_deref_at_centroid: @@ -441,6 +443,7 @@ nir_lower_io_block(nir_block *block, case nir_intrinsic_deref_atomic_xor: case nir_intrinsic_deref_atomic_exchange: case nir_intrinsic_deref_atomic_comp_swap: + case nir_intrinsic_deref_atomic_fadd: assert(vertex_index == NULL); replacement = lower_atomic(intrin, state, var, offset); break; -- 2.30.2