Add F4 F6 to roofline for MI350 series (#709)

Add roofline bins with FP4 FP6 datatypes enabled for gfx950 arch

---------

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

[ROCm/rocprofiler-compute commit: cb2d928ecf]

projects/rocprofiler-compute/src/utils/roofline_calc.py

@@ -44,15 +44,61 @@ FONT_WEIGHT = "bold"
 # SUPPORTED_DATATYPES table is based on datatype support in rocm-amdgpu-bench repository
 # Indicates which datatypes per gpu arch can be generated by the roofline binary
 SUPPORTED_DATATYPES = {
-    "gfx90a": ["FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"],  # Unsupported: F8
-    "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:
+    "gfx90a": [
+        "FP16",
+        "BF16",
+        "FP32",
+        "FP64",
+        "I8",
+        "I32",
+        "I64",
+    ],  # Unsupported: F4, F6, F8
+    "gfx940": [
+        "FP8",
+        "FP16",
+        "BF16",
+        "FP32",
+        "FP64",
+        "I8",
+        "I32",
+        "I64",
+    ],  # Unsupported: F4, F6
+    "gfx941": [
+        "FP8",
+        "FP16",
+        "BF16",
+        "FP32",
+        "FP64",
+        "I8",
+        "I32",
+        "I64",
+    ],  # Unsupported: F4, F6
+    "gfx942": [
+        "FP8",
+        "FP16",
+        "BF16",
+        "FP32",
+        "FP64",
+        "I8",
+        "I32",
+        "I64",
+    ],  # Unsupported: F4, F6
+    "gfx950": [
+        "FP4",
+        "FP6",
+        "FP8",
+        "FP16",
+        "BF16",
+        "FP32",
+        "FP64",
+        "I8",
+        "I32",
+        "I64",
+    ],  # Unsupported:
 }
 
 PEAK_OPS_DATATYPES = ["FP8", "FP16", "BF16", "FP32", "FP64", "I8", "I32", "I64"]
-MFMA_DATATYPES = ["FP8", "FP16", "BF16", "FP32", "FP64", "I8"]
+MFMA_DATATYPES = ["FP4", "FP6", "FP8", "FP16", "BF16", "FP32", "FP64", "I8"]
 
 TOP_N = 10
 
@@ -67,6 +113,7 @@ class AI_Data:
 
     total_flops: float
     valu_flops: float
+    mfma_flops_f6f4: float
     mfma_flops_f8: float
     mfma_flops_f16: float
     mfma_flops_bf16: float
@@ -212,11 +259,11 @@ def calc_ai(mspec, sort_type, ret_df):
     df = df.sort_values(by=["Kernel_Name"])
     df = df.reset_index(drop=True)
 
-    total_flops = valu_flops = mfma_flops_f8 = mfma_flops_bf16 = mfma_flops_f16 = (
-        mfma_iops_i8
-    ) = mfma_flops_f32 = mfma_flops_f64 = lds_data = L1cache_data = L2cache_data = (
-        hbm_data
-    ) = calls = totalDuration = avgDuration = 0.0
+    total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = mfma_flops_bf16 = (
+        mfma_flops_f16
+    ) = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = lds_data = L1cache_data = (
+        L2cache_data
+    ) = hbm_data = calls = totalDuration = avgDuration = 0.0
 
     kernelName = ""
 
@@ -273,6 +320,8 @@ def calc_ai(mspec, sort_type, ret_df):
             )
             if "FP8" in supported_dt:
                 total_flops += df["SQ_INSTS_VALU_MFMA_MOPS_F8"][idx] * 512
+            if ("FP4" in supported_dt) or ("FP6" in supported_dt):
+                total_flops += df["SQ_INSTS_VALU_MFMA_MOPS_F6F4"][idx] * 512
         except KeyError:
             console_debug(
                 "roofline",
@@ -313,6 +362,8 @@ def calc_ai(mspec, sort_type, ret_df):
         try:
             if "FP8" in supported_dt:
                 mfma_flops_f8 += df["SQ_INSTS_VALU_MFMA_MOPS_F8"][idx] * 512
+            if ("FP4" in supported_dt) or ("FP6" in supported_dt):
+                mfma_flops_f6f4 += df["SQ_INSTS_VALU_MFMA_MOPS_F6F4"][idx] * 512
             mfma_flops_f16 += df["SQ_INSTS_VALU_MFMA_MOPS_F16"][idx] * 512
             mfma_flops_bf16 += df["SQ_INSTS_VALU_MFMA_MOPS_BF16"][idx] * 512
             mfma_flops_f32 += df["SQ_INSTS_VALU_MFMA_MOPS_F32"][idx] * 512
@@ -413,6 +464,7 @@ def calc_ai(mspec, sort_type, ret_df):
                     calls,
                     total_flops / calls,
                     valu_flops / calls,
+                    mfma_flops_f6f4 / calls,
                     mfma_flops_f8 / calls,
                     mfma_flops_f16 / calls,
                     mfma_flops_bf16 / calls,
@@ -432,11 +484,13 @@ def calc_ai(mspec, sort_type, ret_df):
                     kernelName, idx, calls
                 )
             )
-            total_flops = valu_flops = mfma_flops_f8 = mfma_flops_bf16 = (
-                mfma_flops_f16
-            ) = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = lds_data = (
-                L1cache_data
-            ) = L2cache_data = hbm_data = calls = totalDuration = avgDuration = 0.0
+            total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = (
+                mfma_flops_bf16
+            ) = mfma_flops_f16 = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = (
+                lds_data
+            ) = L1cache_data = L2cache_data = hbm_data = calls = totalDuration = (
+                avgDuration
+            ) = 0.0
 
         if sort_type == "dispatches":
             myList.append(
@@ -445,6 +499,7 @@ def calc_ai(mspec, sort_type, ret_df):
                     calls,
                     total_flops,
                     valu_flops,
+                    mfma_flops_f6f4,
                     mfma_flops_f8,
                     mfma_flops_f16,
                     mfma_flops_bf16,
@@ -459,11 +514,13 @@ def calc_ai(mspec, sort_type, ret_df):
                     avgDuration,
                 )
             )
-            total_flops = valu_flops = mfma_flops_f8 = mfma_flops_bf16 = (
-                mfma_flops_f16
-            ) = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = lds_data = (
-                L1cache_data
-            ) = L2cache_data = hbm_data = calls = totalDuration = avgDuration = 0.0
+            total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = (
+                mfma_flops_bf16
+            ) = mfma_flops_f16 = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = (
+                lds_data
+            ) = L1cache_data = L2cache_data = hbm_data = calls = totalDuration = (
+                avgDuration
+            ) = 0.0
 
     myList.sort(key=lambda x: x.totalDuration, reverse=True)