Enable roofline for MI350 series (#677)

Rework of roofline binaries generated from rocm-amdgpu-bench
- removed arch identifier in bin name
- removed rocm5 bins altogether

Updated required distros for roofline
- updated distro checks and bin naming
- moved up ubuntu20.04->22.04 and sles15.3->15.6 per rocm support

Enabled ctests for mi350 for test_roof_*
- removed mi350 series check to skip these specific tests

---------

Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>

src/rocprof_compute_soc/profile_configs/gfx950/roofline/pmc_roof_perf.txt

@@ -0,0 +1,12 @@
+pmc: SQ_INSTS_VALU_ADD_F16   SQ_INSTS_VALU_MUL_F16   SQ_INSTS_VALU_FMA_F16   SQ_INSTS_VALU_TRANS_F16  SQ_INSTS_VALU_ADD_F32       SQ_INSTS_VALU_MUL_F32         SQ_INSTS_VALU_FMA_F32       SQ_INSTS_VALU_TRANS_F32
+
+pmc: SQ_INSTS_VALU_ADD_F64   SQ_INSTS_VALU_MUL_F64   SQ_INSTS_VALU_FMA_F64   SQ_INSTS_VALU_TRANS_F64  SQ_INSTS_VALU_MFMA_MOPS_F16 SQ_INSTS_VALU_MFMA_MOPS_BF16  SQ_INSTS_VALU_MFMA_MOPS_F32 SQ_INSTS_VALU_MFMA_MOPS_F64 TCP_TCC_READ_REQ_sum TCC_EA0_RDREQ_32B_sum    TCC_EA0_RDREQ_sum    TCC_EA0_WRREQ_64B_sum    TCC_EA0_WRREQ_sum   TCC_BUBBLE_sum
+
+pmc: SQ_LDS_IDX_ACTIVE    SQ_LDS_BANK_CONFLICT    TCP_TOTAL_CACHE_ACCESSES_sum    TCP_TCC_WRITE_REQ_sum    TCP_TCC_ATOMIC_WITH_RET_REQ_sum    TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum
+
+#ROOF:MI300 series and above
+pmc: SQ_INSTS_VALU_MFMA_MOPS_F8
+
+range:
+gpu:
+kernels:

src/rocprof_compute_soc/soc_gfx950.py

@@ -46,17 +46,6 @@ class gfx950_soc(OmniSoC_Base):
                     )
                 )
             )
-        else:
-            # NB: We're using generalized Mi300 perfmon configs
-            self.set_perfmon_dir(
-                str(
-                    Path(str(config.rocprof_compute_home)).joinpath(
-                        "rocprof_compute_soc",
-                        "profile_configs",
-                        "gfx950",
-                    )
-                )
-            )
         self.set_compatible_profilers(["rocprofv3"])
         # Per IP block max number of simultaneous counters. GFX IP Blocks
         self.set_perfmon_config(

src/utils/roofline_calc.py

@@ -48,6 +48,7 @@ SUPPORTED_DATATYPES = {
     "gfx940": ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"],  # Unsupported:
     "gfx941": ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"],  # Unsupported:
     "gfx942": ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"],  # Unsupported:
+    "gfx950": ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"],  # Unsupported:
 }
 
 PEAK_OPS_DATATYPES = ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"]

src/utils/utils.py

@@ -939,14 +939,14 @@ def detect_roofline(mspec):
     elif (
         (type(sles_distro) == str and len(sles_distro) >= 3)
         and sles_distro[:2] == "15"  # confirm string and len
-        and int(sles_distro[3]) >= 3  # SLES15 and SP >= 3
+        and int(sles_distro[3]) >= 6  # SLES15 and SP >= 6
     ):
         # Must be a valid SLES machine
-        # Use SP3 binary for all forward compatible service pack versions
-        distro = "15.3"
-    elif ubuntu_distro == "20.04" or ubuntu_distro == "22.04" or ubuntu_distro == "24.04":
+        # Use SP6 binary for all forward compatible service pack versions
+        distro = "15.6"
+    elif ubuntu_distro == "22.04" or ubuntu_distro == "24.04":
         # Must be a valid Ubuntu machine
-        distro = ubuntu_distro
+        distro = "22.04"
     else:
         console_error("roofline", "Cannot find a valid binary for your operating system")
 
@@ -984,10 +984,8 @@ def mibench(args, mspec):
 
     distro_map = {
         "platform:el8": "rhel8",
-        "15.3": "sles15sp5",
-        "20.04": "ubuntu20_04",
-        "22.04": "ubuntu20_04",
-        "24.04": "ubuntu20_04",
+        "15.6": "sles15sp6",
+        "22.04": "ubuntu22_04",
     }
 
     binary_paths = []
@@ -1008,8 +1006,6 @@ def mibench(args, mspec):
                 dir
                 + "-"
                 + distro_map[target_binary["distro"]]
-                + "-"
-                + mspec.gpu_series.lower()
                 + "-rocm"
                 + target_binary["rocm_ver"]
             )