[Mesa-dev] [PATCH 09/13] nir: Add floating point atomic min, max, and compare-swap instrinsics
Ian Romanick
idr at freedesktop.org
Mon Aug 20 18:37:38 UTC 2018
On 08/17/2018 05:06 PM, Caio Marcelo de Oliveira Filho wrote:
> On Fri, Jun 22, 2018 at 10:03:54PM -0700, Ian Romanick wrote:
>> From: Ian Romanick <ian.d.romanick at intel.com>
>>
>> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
>> ---
>> src/compiler/glsl/glsl_to_nir.cpp | 32 ++++++++++++++++++++++------
>> src/compiler/nir/nir_intrinsics.py | 11 +++++++++-
>> src/compiler/nir/nir_lower_atomics_to_ssbo.c | 6 +++++-
>> src/compiler/nir/nir_lower_io.c | 9 ++++++++
>> 4 files changed, 50 insertions(+), 8 deletions(-)
>>
>> diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp
>> index 90e960592a0..4d6f826f72f 100644
>> --- a/src/compiler/glsl/glsl_to_nir.cpp
>> +++ b/src/compiler/glsl/glsl_to_nir.cpp
>> @@ -735,6 +735,8 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_ssbo_atomic_imin;
>> else if (ir->return_deref->type == glsl_type::uint_type)
>> op = nir_intrinsic_ssbo_atomic_umin;
>> + else if (ir->return_deref->type == glsl_type::float_type)
>> + op = nir_intrinsic_ssbo_atomic_fmin;
>> else
>> unreachable("Invalid type");
>> break;
>> @@ -744,6 +746,8 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_ssbo_atomic_imax;
>> else if (ir->return_deref->type == glsl_type::uint_type)
>> op = nir_intrinsic_ssbo_atomic_umax;
>> + else if (ir->return_deref->type == glsl_type::float_type)
>> + op = nir_intrinsic_ssbo_atomic_fmax;
>> else
>> unreachable("Invalid type");
>> break;
>> @@ -751,7 +755,9 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_ssbo_atomic_exchange;
>> break;
>> case ir_intrinsic_ssbo_atomic_comp_swap:
>> - op = nir_intrinsic_ssbo_atomic_comp_swap;
>> + op = ir->return_deref->type->is_integer_32_64()
>> + ? nir_intrinsic_ssbo_atomic_comp_swap
>> + : nir_intrinsic_ssbo_atomic_fcomp_swap;
>
> Why not compare to glsl_type::float_type?
No particular reason. :) I guess if we ever get 64-bit float atomics,
we'd need a similar condition. *shrug*
>> break;
>> case ir_intrinsic_shader_clock:
>> op = nir_intrinsic_shader_clock;
>> @@ -803,6 +809,8 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_shared_atomic_imin;
>> else if (ir->return_deref->type == glsl_type::uint_type)
>> op = nir_intrinsic_shared_atomic_umin;
>> + else if (ir->return_deref->type == glsl_type::float_type)
>> + op = nir_intrinsic_shared_atomic_fmin;
>> else
>> unreachable("Invalid type");
>> break;
>> @@ -812,6 +820,8 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_shared_atomic_imax;
>> else if (ir->return_deref->type == glsl_type::uint_type)
>> op = nir_intrinsic_shared_atomic_umax;
>> + else if (ir->return_deref->type == glsl_type::float_type)
>> + op = nir_intrinsic_shared_atomic_fmax;
>> else
>> unreachable("Invalid type");
>> break;
>> @@ -819,7 +829,9 @@ nir_visitor::visit(ir_call *ir)
>> op = nir_intrinsic_shared_atomic_exchange;
>> break;
>> case ir_intrinsic_shared_atomic_comp_swap:
>> - op = nir_intrinsic_shared_atomic_comp_swap;
>> + op = ir->return_deref->type->is_integer_32_64()
>> + ? nir_intrinsic_shared_atomic_comp_swap
>> + : nir_intrinsic_shared_atomic_fcomp_swap;
>> break;
>> case ir_intrinsic_vote_any:
>> op = nir_intrinsic_vote_any;
>> @@ -1068,7 +1080,10 @@ nir_visitor::visit(ir_call *ir)
>> 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_ssbo_atomic_fadd:
>> + case nir_intrinsic_ssbo_atomic_fmin:
>> + case nir_intrinsic_ssbo_atomic_fmax:
>> + case nir_intrinsic_ssbo_atomic_fcomp_swap: {
>> int param_count = ir->actual_parameters.length();
>> assert(param_count == 3 || param_count == 4);
>>
>> @@ -1089,7 +1104,8 @@ nir_visitor::visit(ir_call *ir)
>>
>> /* data2 parameter (only with atomic_comp_swap) */
>> if (param_count == 4) {
>> - assert(op == nir_intrinsic_ssbo_atomic_comp_swap);
>> + assert(op == nir_intrinsic_ssbo_atomic_comp_swap ||
>> + op == nir_intrinsic_ssbo_atomic_fcomp_swap);
>> param = param->get_next();
>> inst = (ir_instruction *) param;
>> instr->src[3] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue()));
>> @@ -1152,7 +1168,10 @@ nir_visitor::visit(ir_call *ir)
>> case nir_intrinsic_shared_atomic_xor:
>> case nir_intrinsic_shared_atomic_exchange:
>> case nir_intrinsic_shared_atomic_comp_swap:
>> - case nir_intrinsic_shared_atomic_fadd: {
>> + case nir_intrinsic_shared_atomic_fadd:
>> + case nir_intrinsic_shared_atomic_fmin:
>> + case nir_intrinsic_shared_atomic_fmax:
>> + case nir_intrinsic_shared_atomic_fcomp_swap: {
>> int param_count = ir->actual_parameters.length();
>> assert(param_count == 2 || param_count == 3);
>>
>> @@ -1168,7 +1187,8 @@ nir_visitor::visit(ir_call *ir)
>>
>> /* data2 parameter (only with atomic_comp_swap) */
>> if (param_count == 3) {
>> - assert(op == nir_intrinsic_shared_atomic_comp_swap);
>> + assert(op == nir_intrinsic_shared_atomic_comp_swap ||
>> + op == nir_intrinsic_shared_atomic_fcomp_swap);
>> param = param->get_next();
>> inst = (ir_instruction *) param;
>> instr->src[2] =
>> diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py
>> index f1726e54fb0..e492b96346f 100644
>> --- a/src/compiler/nir/nir_intrinsics.py
>> +++ b/src/compiler/nir/nir_intrinsics.py
>> @@ -357,6 +357,9 @@ intrinsic("var_atomic_xor", src_comp=[1], dest_comp=1, num_vars=1)
>> intrinsic("var_atomic_exchange", src_comp=[1], dest_comp=1, num_vars=1)
>> intrinsic("var_atomic_comp_swap", src_comp=[1, 1], dest_comp=1, num_vars=1)
>> intrinsic("var_atomic_fadd", src_comp=[1], dest_comp=1, num_vars=1)
>> +intrinsic("var_atomic_fmin", src_comp=[1], dest_comp=1, num_vars=1)
>> +intrinsic("var_atomic_fmax", src_comp=[1], dest_comp=1, num_vars=1)
>> +intrinsic("var_atomic_fcomp_swap", src_comp=[1, 1], dest_comp=1, num_vars=1)
>>
>> # SSBO atomic intrinsics
>> #
>> @@ -383,7 +386,10 @@ 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)
>> +intrinsic("ssbo_atomic_fadd", src_comp=[1, 1, 1], dest_comp=1)
>> +intrinsic("ssbo_atomic_fmin", src_comp=[1, 1, 1], dest_comp=1)
>> +intrinsic("ssbo_atomic_fmax", src_comp=[1, 1, 1], dest_comp=1)
>> +intrinsic("ssbo_atomic_fcomp_swap", src_comp=[1, 1, 1, 1], dest_comp=1)
>>
>> # CS shared variable atomic intrinsics
>> #
>> @@ -410,6 +416,9 @@ 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])
>> +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])
>>
>> def system_value(name, dest_comp, indices=[]):
>> intrinsic("load_" + name, [], dest_comp, 0, indices,
>> diff --git a/src/compiler/nir/nir_lower_atomics_to_ssbo.c b/src/compiler/nir/nir_lower_atomics_to_ssbo.c
>> index 72af914fb47..3af9acafef2 100644
>> --- a/src/compiler/nir/nir_lower_atomics_to_ssbo.c
>> +++ b/src/compiler/nir/nir_lower_atomics_to_ssbo.c
>> @@ -58,6 +58,9 @@ lower_instr(nir_intrinsic_instr *instr, unsigned ssbo_offset, nir_builder *b)
>> case nir_intrinsic_ssbo_atomic_exchange:
>> case nir_intrinsic_ssbo_atomic_comp_swap:
>> case nir_intrinsic_ssbo_atomic_fadd:
>> + case nir_intrinsic_ssbo_atomic_fmin:
>> + case nir_intrinsic_ssbo_atomic_fmax:
>> + case nir_intrinsic_ssbo_atomic_fcomp_swap:
>> case nir_intrinsic_store_ssbo:
>> case nir_intrinsic_load_ssbo:
>> case nir_intrinsic_get_buffer_size:
>> @@ -138,7 +141,8 @@ lower_instr(nir_intrinsic_instr *instr, unsigned ssbo_offset, nir_builder *b)
>> new_instr->src[0] = nir_src_for_ssa(buffer);
>> nir_src_copy(&new_instr->src[1], &instr->src[0], new_instr);
>> nir_src_copy(&new_instr->src[2], &instr->src[1], new_instr);
>> - if (op == nir_intrinsic_ssbo_atomic_comp_swap)
>> + if (op == nir_intrinsic_ssbo_atomic_comp_swap ||
>> + op == nir_intrinsic_ssbo_atomic_fcomp_swap)
>> nir_src_copy(&new_instr->src[3], &instr->src[2], new_instr);
>> break;
>> }
>> diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c
>> index 9a30564d988..01367b37dc9 100644
>> --- a/src/compiler/nir/nir_lower_io.c
>> +++ b/src/compiler/nir/nir_lower_io.c
>> @@ -287,6 +287,9 @@ lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
>> OP(atomic_or)
>> OP(atomic_xor)
>> OP(atomic_fadd)
>> + OP(atomic_fmin)
>> + OP(atomic_fmax)
>> + OP(atomic_fcomp_swap)
>> #undef OP
>> default:
>> unreachable("Invalid atomic");
>> @@ -387,6 +390,9 @@ nir_lower_io_block(nir_block *block,
>> case nir_intrinsic_var_atomic_exchange:
>> case nir_intrinsic_var_atomic_comp_swap:
>> case nir_intrinsic_var_atomic_fadd:
>> + case nir_intrinsic_var_atomic_fmin:
>> + case nir_intrinsic_var_atomic_fmax:
>> + case nir_intrinsic_var_atomic_fcomp_swap:
>> /* We can lower the io for this nir instrinsic */
>> break;
>> case nir_intrinsic_interp_var_at_centroid:
>> @@ -448,6 +454,9 @@ nir_lower_io_block(nir_block *block,
>> case nir_intrinsic_var_atomic_exchange:
>> case nir_intrinsic_var_atomic_comp_swap:
>> case nir_intrinsic_var_atomic_fadd:
>> + case nir_intrinsic_var_atomic_fmin:
>> + case nir_intrinsic_var_atomic_fmax:
>> + case nir_intrinsic_var_atomic_fcomp_swap:
>> assert(vertex_index == NULL);
>> replacement = lower_atomic(intrin, state, offset);
>> break;
>> --
>> 2.14.4
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
More information about the mesa-dev
mailing list