add nvfp4 cast benchmarks (#3188)

Update

[ghstack-poisoned]

Author: vkuzo

benchmarks/mx_formats/cast_bench.py

@@ -17,6 +17,7 @@
     triton_to_mxfp8_dim1,
 )
 from torchao.prototype.mx_formats.mx_tensor import to_mx
+from torchao.prototype.mx_formats.nvfp4_tensor import NVFP4Tensor
 
 torch.manual_seed(0)
 
@@ -76,6 +77,18 @@ def to_mx_dim1_reference(
     return data_d1.t(), scale_d1
 
 
+def to_nvfp4_reference(x_hp):
+    nvfp4_tensor = NVFP4Tensor.to_nvfp4(x_hp, use_triton_kernel=False)
+    return nvfp4_tensor.qdata, nvfp4_tensor.scale
+
+
+def to_nvfp4_reference_triton_swizzle(x_hp):
+    nvfp4_tensor = NVFP4Tensor.to_nvfp4(
+        x_hp, use_triton_kernel=True, is_swizzled_scales=True
+    )
+    return nvfp4_tensor.qdata, nvfp4_tensor.scale
+
+
 def benchmark_cuda_function_in_microseconds(f, *args):
     return do_bench(lambda: f(*args), return_mode="median") * 1e3
 
@@ -99,6 +112,8 @@ def run(
         "dim0_mxfp4_floor",
         "dim0_mxfp8_rceil",
         "dim0_mxfp8_triton_floor",
+        "dim0_nvfp4",
+        "dim0_nvfp4_triton_swizzle",
         "dim1_mxfp8_floor",
         "dim1_mxfp8_rceil",
         "dim1_mxfp8_triton_floor",
@@ -240,6 +255,37 @@ def run(
         bytes_w = (y_d0.numel() + s_d0.numel()) * bytes_per_el_fp8
         bps = (bytes_r + bytes_w) / (time_us / 1e6)
 
+    elif mode == "dim0_nvfp4":
+        to_nvfp4_reference_c = torch.compile(to_nvfp4_reference)
+        y_d0, s_d0 = to_nvfp4_reference_c(x, use_triton_kernel=False)
+
+        for _ in range(2):
+            __ = to_nvfp4_reference_c(x, use_triton_kernel=False)
+        time_us = benchmark_cuda_function_in_microseconds(
+            lambda x: to_nvfp4_reference_c(x, use_triton_kernel=False),
+            x,
+        )
+        assert y_d0.dtype == torch.uint8
+        assert s_d0.dtype == torch.float8_e4m3fn
+        bytes_r = x.numel() * bytes_per_el_bf16
+        bytes_w = (y_d0.numel() + s_d0.numel()) * bytes_per_el_fp8
+        bps = (bytes_r + bytes_w) / (time_us / 1e6)
+
+    elif mode == "dim0_nvfp4_triton_swizzle":
+        y_d0, s_d0 = to_nvfp4_reference_triton_swizzle(x)
+
+        for _ in range(2):
+            __ = to_nvfp4_reference_triton_swizzle(x)
+        time_us = benchmark_cuda_function_in_microseconds(
+            lambda x: to_nvfp4_reference_triton_swizzle(x),
+            x,
+        )
+        assert y_d0.dtype == torch.uint8
+        assert s_d0.dtype == torch.float8_e4m3fn
+        bytes_r = x.numel() * bytes_per_el_bf16
+        bytes_w = (y_d0.numel() + s_d0.numel()) * bytes_per_el_fp8
+        bps = (bytes_r + bytes_w) / (time_us / 1e6)
+
     elif mode == "dim1_mxfp8_floor":
         to_mx_dim1_reference_c = torch.compile(to_mx_dim1_reference)
         y_d1, s_d1 = to_mx_dim1_reference_c(x, BLOCK_SIZE)