[GCN] Extract stats from the 'AMDGPU.csdata' section. More tests.

Author: Keluaa

src/stats/gcn.jl

@@ -8,6 +8,10 @@ struct GCNStats
     vgpr_spill_count :: Int
     uses_dyn_stack   :: Bool
 
+    # Memory stats
+    code_size :: Int
+    LDS_size  :: Int  # Static allocations of shared memory only
+
     # Arguments
     arguments_count :: Int
     arguments_size  :: Int  # in bytes
@@ -51,8 +55,13 @@ function extract_stats(::Val{:gcn}, gcn_source, stats_opts)
 
     code = extract_gcn_code_stats(gcn_source; arch, arch_version)
 
+    csdata = extract_gcn_csdata(gcn_source)
+    code_bytes = get(csdata, "codeLenInByte", 0)
+    LDS_bytes  = get(csdata, "LDSByteSize", 0)
+
     return GCNStats(
         sgpr_count, vgpr_count, agpr_count, sgpr_spills, vgpr_spills, uses_dyn_stack,
+        code_bytes, LDS_bytes,
         arguments_count, arguments_size,
         target_triple, processor,
         string(arch), arch_version, wavefront_size,
@@ -62,14 +71,6 @@ function extract_stats(::Val{:gcn}, gcn_source, stats_opts)
 end
 
 
-parse_match_group(::Nothing, group, default) = default
-parse_match_group(m::RegexMatch, group, default::String) = @something m[group] default
-function parse_match_group(m::RegexMatch, group, default::T) where {T}
-    isnothing(m[group]) && return default
-    return @something tryparse(T, m[group]) default
-end
-
-
 function parse_amdgpu_target(target_triple)
     # Docs: https://llvm.org/docs/AMDGPUUsage.html#target-triples
     # It is made of a "<target triple>-<processor>", with 4 '-' in total.
@@ -191,6 +192,37 @@ function extract_gcn_code_stats(gcn_source; arch=:unknown, arch_version=v"0")
 end
 
 
+function extract_gcn_csdata(gcn_source)
+    # Between the GCN kernel source and the YAML metadata, there is a "AMDGPU.csdata" section.
+    # I cannot find where it is documented (if it is), therefore it shouldn't be relied on too much.
+    # It likely contains important information for the GPU driver, therefore it may be reliable, but
+    # it may also depend on the target.
+    # Some of the information is also found in the YAML metadata, which should be preferred over
+    # this one, as its format and content does not depend on the target.
+    csdata_start = findfirst(r"\.section\s+\.AMDGPU\.csdata", gcn_source)
+    isnothing(csdata_start) && return Dict{String, Int}()
+
+    csdata_end = findnext(".text", gcn_source, last(csdata_start))
+    isnothing(csdata_end) && return Dict{String, Int}()
+
+    csdata_section = @view gcn_source[last(csdata_end):first(csdata_end)]
+
+    # Captures the key and value (decimal) of all entries within the csdata section.
+    # Examples:
+    #  "; NumSgprs: 8" => key="NumSgprs", value="8"
+    #  "; COMPUTE_PGM_RSRC2:SCRATCH_EN: 0" => key="COMPUTE_PGM_RSRC2:SCRATCH_EN", value="0"
+    csdata_regex = r"^;\s(?<key>[\w:]+)\s?[=:]\s(?<value>\d+)\b"m
+
+    csdata = Dict{String, Int}()
+    for m in eachmatch(csdata_regex, csdata_section)
+        value = @something tryparse(Int, m[:value]) 0
+        csdata[m[:key]] = value
+    end
+
+    return csdata
+end
+
+
 function extract_gcn_yaml_metadata(gcn_source)
     # As per the LLVM docs, this section is in YAML.
     # https://llvm.org/docs/AMDGPUUsage.html#id112
@@ -231,6 +263,10 @@ function Base.show(io::IO, stats::GCNStats)
     println(io, " - Accum. registers (AGPR)  ", stats.accu_registers)
     println(io, " - Uses dynamic stack       ", stats.uses_dyn_stack)
     println(io)
+    println(io, "Kernel memory stats:")
+    println(io, " - Instructions             ", Base.format_bytes(stats.code_size))
+    println(io, " - Shared mem. (LDS)        ", Base.format_bytes(stats.LDS_size), " (static allocations)")
+    println(io)
     println(io, "Kernel target:")
     println(io, " - Target                   ", stats.target_triple)
     println(io, " - ISA                      ", stats.ISA)

test/stats.jl

@@ -62,8 +62,6 @@ end
         @test sass_stats.warp_sync      == 0
         @test sass_stats.calls          == 12
 
-        disp_sass_stats = display_str(sass_stats)
-
         cleaned_sass = CDC.cleanup_code(Val(:sass), sass_sample)
         cleaned_sass_stats = CDS.extract_stats(Val(:sass), cleaned_sass)
         @test sass_stats == cleaned_sass_stats
@@ -102,6 +100,9 @@ end
         @test gcn_stats.arguments_count == 12
         @test gcn_stats.arguments_size  == 304
 
+        @test gcn_stats.code_size == 325 * 8  # each instruction is 8 bytes long
+        @test gcn_stats.LDS_size  == 0
+
         @test gcn_stats.target_triple  == "amdgcn-amd-amdhsa"
         @test gcn_stats.ISA            == "gfx90a"
         @test gcn_stats.architecture   == "CDNA"
@@ -122,6 +123,33 @@ end
         @test cleaned_gcn_stats.scalar_registers == 0
         @test cleaned_gcn_stats.inst_count == gcn_stats.inst_count
         @test gcn_stats != cleaned_gcn_stats
+
+        disp_gcn_stats = display_str(gcn_stats)
+        @test occursin("CDNA 2", disp_gcn_stats)
+    end
+
+    @testset "Architecture" begin
+        @test CDS.amdgpu_processor_to_architecture("unknown") == (:unknown, v"0")
+        @test CDS.amdgpu_processor_to_architecture("gfx0")    == (:unknown, v"0")
+
+        @test CDS.amdgpu_processor_to_architecture("gfx600")  == (:GCN,  v"1")
+        @test CDS.amdgpu_processor_to_architecture("gfx700")  == (:GCN,  v"2")
+        @test CDS.amdgpu_processor_to_architecture("gfx800")  == (:GCN,  v"3")
+        @test CDS.amdgpu_processor_to_architecture("gfx900")  == (:GCN,  v"5")
+
+        @test CDS.amdgpu_processor_to_architecture("gfx908")  == (:CDNA, v"1")
+        @test CDS.amdgpu_processor_to_architecture("gfx90a")  == (:CDNA, v"2")
+        @test CDS.amdgpu_processor_to_architecture("gfx942")  == (:CDNA, v"3")
+        @test CDS.amdgpu_processor_to_architecture("gfx950")  == (:CDNA, v"3")
+
+        @test CDS.amdgpu_processor_to_architecture("gfx1010") == (:RDNA, v"1")
+        @test CDS.amdgpu_processor_to_architecture("gfx1100") == (:RDNA, v"3")
+        @test CDS.amdgpu_processor_to_architecture("gfx1200") == (:RDNA, v"4")
+
+        @test CDS.amdgpu_processor_to_architecture("gfx9-generic")    == (:GCN,  v"5")
+        @test CDS.amdgpu_processor_to_architecture("gfx10-3-generic") == (:RDNA, v"3")
+        @test CDS.amdgpu_processor_to_architecture("gfx11-generic")   == (:RDNA, v"3")
+        @test CDS.amdgpu_processor_to_architecture("gfx12-generic")   == (:RDNA, v"4")
     end
 end