FP8 roofline support (#592)

Adding FP8 datatype to roofline feature in rocprof-compute on MI300-based systems.
FP8 now shows in terminal output and roofline csv, and outputs a standalone PDF.

---------

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

docs/how-to/profile/mode.rst

@@ -441,9 +441,9 @@ successfully.
 
 .. note::
 
-   ROCm Compute Profiler generates two roofline outputs to organize results and reduce
-   clutter. One chart plots FP32/FP64 performance while the other plots I8/FP16
-   performance.
+   ROCm Compute Profiler generates three roofline outputs to organize results and reduce
+   clutter. One chart plots FP32/FP64 performance, one plots I8/FP16
+   performance, and the other plots FP8 performance.
 
 The following image is a sample ``empirRoof_gpu-0_int8_fp16.pdf`` roofline
 plot.

src/rocprof_compute_soc/profile_configs/gfx940/roofline/pmc_roof_perf.txt

@@ -4,6 +4,9 @@ pmc: SQ_INSTS_VALU_ADD_F64   SQ_INSTS_VALU_MUL_F64   SQ_INSTS_VALU_FMA_F64   SQ_
 
 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
+pmc: SQ_INSTS_VALU_MFMA_MOPS_F8
+
 range:
 gpu:
 kernels:

src/roofline.py

@@ -130,6 +130,9 @@ class Roofline:
             dtype="I8",
             fig=fp16_fig,
         )
+        if self.__mspec.gpu_series != "MI200":
+            fig_fp8 = self.generate_plot(dtype="FP8")
+
         # Create a legend and distinct kernel markers. This can be saved, optionally
         self.__figure = go.Figure(
             go.Scatter(
@@ -163,6 +166,11 @@ class Roofline:
                 self.__run_parameters["workload_dir"]
                 + "/empirRoof_gpu-{}_int8_fp16.pdf".format(dev_id)
             )
+            if self.__mspec.gpu_series != "MI200":
+                fig_fp8.write_image(
+                    self.__run_parameters["workload_dir"]
+                    + "/empirRoof_gpu-{}_fp8.pdf".format(dev_id)
+                )
             # only save a legend if kernel_names option is toggled
             if self.__run_parameters["include_kernel_names"]:
                 self.__figure.write_image(
@@ -178,6 +186,11 @@ class Roofline:
                 self.__run_parameters["workload_dir"]
                 + "/empirRoof_gpu-{}_int8_fp16.pdf".format(dev_id)
             )
+            if self.__mspec.gpu_series != "MI200":
+                fig_fp8.write_image(
+                    self.__run_parameters["workload_dir"]
+                    + "/empirRoof_gpu-{}_fp8.pdf".format(dev_id)
+                )
             if self.__run_parameters["include_kernel_names"]:
                 self.__figure.write_image(
                     self.__run_parameters["workload_dir"] + "/kernelName_legend.pdf"

src/utils/roofline_calc.py

@@ -56,6 +56,7 @@ class AI_Data:
 
     total_flops: float
     valu_flops: float
+    mfma_flops_f8: float
     mfma_flops_f16: float
     mfma_flops_bf16: float
     mfma_flops_f32: float
@@ -193,11 +194,11 @@ 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_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_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 = ""
 
@@ -250,6 +251,8 @@ def calc_ai(mspec, sort_type, ret_df):
                 + (df["SQ_INSTS_VALU_MFMA_MOPS_F32"][idx] * 512)
                 + (df["SQ_INSTS_VALU_MFMA_MOPS_F64"][idx] * 512)
             )
+            if mspec.gpu_series != "MI200":
+                total_flops += df["SQ_INSTS_VALU_MFMA_MOPS_F8"][idx] * 512
         except KeyError:
             console_debug(
                 "roofline",
@@ -288,6 +291,8 @@ def calc_ai(mspec, sort_type, ret_df):
             pass
 
         try:
+            if mspec.gpu_series != "MI200":
+                mfma_flops_f8 += df["SQ_INSTS_VALU_MFMA_MOPS_F8"][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
@@ -388,6 +393,7 @@ def calc_ai(mspec, sort_type, ret_df):
                     calls,
                     total_flops / calls,
                     valu_flops / calls,
+                    mfma_flops_f8 / calls,
                     mfma_flops_f16 / calls,
                     mfma_flops_bf16 / calls,
                     mfma_flops_f32 / calls,
@@ -406,11 +412,11 @@ def calc_ai(mspec, sort_type, ret_df):
                     kernelName, idx, calls
                 )
             )
-            total_flops = valu_flops = 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_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(
@@ -419,6 +425,7 @@ def calc_ai(mspec, sort_type, ret_df):
                     calls,
                     total_flops,
                     valu_flops,
+                    mfma_flops_f8,
                     mfma_flops_f16,
                     mfma_flops_bf16,
                     mfma_flops_f32,
@@ -432,11 +439,11 @@ def calc_ai(mspec, sort_type, ret_df):
                     avgDuration,
                 )
             )
-            total_flops = valu_flops = 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_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)
 

tests/test_profile_general.py

@@ -121,6 +121,7 @@ ROOF_ONLY_FILES = sorted(
     [
         "empirRoof_gpu-0_fp32_fp64.pdf",
         "empirRoof_gpu-0_int8_fp16.pdf",
+        "empirRoof_gpu-0_fp8.pdf",
         "pmc_perf.csv",
         "pmc_perf_0.csv",
         "pmc_perf_1.csv",