Prototype (#2486)

Summary: Pull Request resolved: #2486

Differential Revision: D61055780

torchbenchmark/operators/fused_ffn/__init__.py

@@ -0,0 +1 @@
+from .operator import Operator

torchbenchmark/operators/fused_ffn/kernel.py

@@ -0,0 +1,306 @@
+
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.autotune(
+    configs=[
+        triton.Config(
+            # B_T, H_D (8192), D (2048)
+            {"BLOCK_M": BLOCK_M, "BLOCK_N": BLOCK_N, "BLOCK_K": BLOCK_K},
+            num_stages=num_stages,
+            num_warps=num_warps,
+        )
+        for BLOCK_M in [64]
+        for BLOCK_N in [128]
+        for BLOCK_K in [128, 256]
+        for num_stages in [2]
+        for num_warps in [8]
+    ],
+    key=["B_T", "D", "H_D"],
+)
+@triton.jit
+def fused_ffn_fwd(
+    x_ptr,
+    w13_ptr,
+    w2_ptr,
+    output_ptr,
+    p_ptr,
+    B_T,
+    stride_xa,
+    stride_xb,
+    stride_w13a,
+    stride_w13b,
+    stride_w2a,
+    stride_w2b,
+    stride_oa,
+    stride_ob,
+    stride_pa,
+    stride_pb,
+    HAS_P: tl.constexpr,
+    D: tl.constexpr,
+    H_D: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+):
+    pid_m = tl.program_id(axis=0)
+    dtype = x_ptr.dtype.element_ty
+
+    X_block_ptr = tl.make_block_ptr(
+        base=x_ptr,
+        shape=(B_T, D),
+        strides=(stride_xa, stride_xb),
+        offsets=(pid_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_K),
+        order=(1, 0),
+    )
+    O_block_ptr = tl.make_block_ptr(
+        base=output_ptr,
+        shape=(B_T, D),
+        strides=(stride_oa, stride_ob),
+        offsets=(pid_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_K),
+        order=(1, 0),
+    )
+
+    for start_n in range(0, H_D, BLOCK_N):
+        if HAS_P:
+            P_block_ptr = tl.make_block_ptr(
+                base=p_ptr,
+                shape=(B_T, H_D),
+                strides=(stride_pa, stride_pb),
+                offsets=(pid_m * BLOCK_M, start_n),
+                block_shape=(BLOCK_M, BLOCK_N),
+                order=(1, 0),
+            )
+        else:
+            P_block_ptr = None
+
+        w1t_bptr = tl.make_block_ptr(
+            base=w13_ptr,
+            shape=(D, H_D),
+            strides=(stride_w13b, stride_w13a),
+            offsets=(0, start_n),
+            block_shape=(BLOCK_K, BLOCK_N),
+            order=(0, 1),
+        )
+        w3t_bptr = tl.make_block_ptr(
+            base=w13_ptr,
+            shape=(D, H_D),
+            strides=(stride_w13b, stride_w13a),
+            offsets=(0, H_D + start_n),
+            block_shape=(BLOCK_K, BLOCK_N),
+            order=(0, 1),
+        )
+        w2_bptr = tl.make_block_ptr(
+            base=w2_ptr,
+            shape=(H_D, D),
+            strides=(stride_w2a, stride_w2b),
+            offsets=(0, 0),
+            block_shape=(BLOCK_N, BLOCK_K),
+            order=(1, 0),
+        )
+
+        x_bptr = X_block_ptr
+        o_bptr = O_block_ptr
+        acc_1 = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+        acc_3 = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+        # first GEMM
+        w1t_bptr_inner = w1t_bptr
+        w3t_bptr_inner = w3t_bptr
+        w2_bptr_inner = w2_bptr
+        for _ in range(0, D, BLOCK_K):
+            x = tl.load(x_bptr)
+            w1t = tl.load(w1t_bptr_inner)
+            w3t = tl.load(w3t_bptr_inner)
+            acc_1 = tl.dot(x, w1t, acc_1)
+            acc_3 = tl.dot(x, w3t, acc_3)
+            x_bptr = tl.advance(x_bptr, (0, BLOCK_K))
+            w1t_bptr_inner = tl.advance(w1t_bptr_inner, (BLOCK_K, 0))
+            w3t_bptr_inner = tl.advance(w3t_bptr_inner, (BLOCK_K, 0))
+        # acc_1 = acc_1.to(dtype).to(tl.float32)
+        # acc_3 = acc_3.to(dtype).to(tl.float32)
+        p = acc_1 * tl.sigmoid(acc_1) * acc_3
+        p = p.to(dtype)
+        if HAS_P:
+            tl.store(P_block_ptr, p)
+        # second GEMM
+        for _ in range(0, BLOCK_K, BLOCK_K):
+            w2 = tl.load(w2_bptr)
+            o = tl.load(o_bptr)
+            tl.store(o_bptr, (tl.dot(p, w2) + o).to(dtype))
+            w2_bptr_inner = tl.advance(w2_bptr_inner, (0, BLOCK_K))
+            o_bptr = tl.advance(o_bptr, (0, BLOCK_K))
+
+
+def fused_ffn(
+    x: torch.Tensor, w13: torch.Tensor, w2: torch.Tensor, has_p: bool = False
+):
+    # x: [B_T, D]
+    # w13: [H_D*2, D]
+    # w2: [H_D, D]
+    # output: [B_T, D]
+    B_T, D = x.shape
+    H_D_2, D = w13.shape
+    H_D = w2.shape[0]
+    assert H_D_2 == 2 * H_D, f"H_D_2 must be 2 times of H_D but got {H_D_2=} and {H_D=}"
+
+    def grid(META):
+        return (triton.cdiv(B_T, META["BLOCK_M"]),)
+
+    output = torch.empty_like(x)
+    if has_p:
+        p = torch.empty((B_T, H_D), dtype=x.dtype, device=x.device)
+    else:
+        p = None
+
+    fused_ffn_fwd[grid](
+        x,
+        w13,
+        w2,
+        output,
+        p,
+        B_T,
+        x.stride(0),
+        x.stride(1),
+        w13.stride(0),
+        w13.stride(1),
+        w2.stride(0),
+        w2.stride(1),
+        output.stride(0),
+        output.stride(1),
+        p.stride(0) if has_p else 0,
+        p.stride(1) if has_p else 0,
+        has_p,
+        D,
+        H_D,
+    )
+
+    return output, p
+
+
+@triton.jit
+# pyre-fixme[3]: Return type must be annotated.
+def _silu_mul_kernel(
+    # pyre-fixme[2]: Parameter must be annotated.
+    x1_ptr,
+    x1_stride: tl.constexpr,
+    # pyre-fixme[2]: Parameter must be annotated.
+    x2_ptr,
+    x2_stride: tl.constexpr,
+    # pyre-fixme[2]: Parameter must be annotated.
+    y_ptr,
+    D: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    b = tl.program_id(0).to(tl.int64)
+
+    x1_start = x1_ptr + b * x1_stride
+    x2_start = x2_ptr + b * x2_stride
+    y_start = y_ptr + b * D
+
+    for offset in range(0, D, BLOCK_SIZE):
+        cols = offset + tl.arange(0, BLOCK_SIZE)
+        mask = cols < D
+        x1v = tl.load(x1_start + cols, mask=mask, other=0).to(tl.float32)
+        x2v = tl.load(x2_start + cols, mask=mask, other=0).to(tl.float32)
+        yv = (x1v * tl.sigmoid(x1v) * x2v).to(tl.bfloat16)
+        tl.store(y_start + cols, yv, mask=mask)
+
+
+sigmoid = torch.nn.Sigmoid()
+
+
+def silu_mul(x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
+    assert x1.shape == x2.shape
+    (B_T, D) = x1.shape
+    out = torch.empty_like(x1)
+    assert x1.stride(1) == x2.stride(1) == 1
+    assert out.is_contiguous()
+    grid = (B_T,)
+    _silu_mul_kernel[grid](x1, x1.stride(0), x2, x2.stride(0), out, D, BLOCK_SIZE=1024)
+    return out
+
+
+def eager_ffn(x, w13, w2):
+    p = x @ w13.T
+    H_D_2, D = w13.shape
+    H_D = H_D_2 // 2
+    p1 = p[:, :H_D]  # B_T, H_D
+    p2 = p[:, H_D:]  # B_T, H_D
+    p_out = silu_mul(p1, p2)  # B_T, H_D
+    out = p_out @ w2
+    return out, p_out
+
+
+def nunerics_check(shape):
+    B_T, H_D, D = shape
+    x = torch.randn((B_T, D), dtype=torch.bfloat16, device="cuda")
+    w13 = torch.randn((H_D * 2, D), dtype=torch.bfloat16, device="cuda")
+    w2 = torch.randn((H_D, D), dtype=torch.bfloat16, device="cuda")
+    triton_out, triton_p = fused_ffn(x, w13, w2, has_p=True)
+    eager_out, eager_p = eager_ffn(x, w13, w2)
+
+    print("P numeric check: ", torch.allclose(triton_p, eager_p, atol=1e-2, rtol=1e-2))
+    # print("P numeric check: ", torch.allclose(eager_p, ref_p, atol=1e-2, rtol=0))
+    # print(triton_p[-1])
+    # print(eager_p[-1])
+    # print(ref_p[-1])
+
+
+def do_benchmark():
+
+    D = 2048
+    H_D = 8192
+
+    configs = []
+    configs.append(
+        triton.testing.Benchmark(
+            x_names=[
+                "B_T",
+                "H_D",
+                "D",
+            ],  # Argument names to use as an x-axis for the plot
+            x_vals=[
+                (i, H_D, D)
+                for H_D, D in [(128, 256), (1024, 512), (8192, 2048)]
+                for i in [1024, 2048, 4096, 8192, 16384]
+            ],  # Different possible values for `x_name`
+            line_arg="provider",  # Argument name whose value corresponds to a different line in the plot
+            # Possible values for `line_arg`
+            # Don't compare to cublas for fp8 cases as torch.matmul doesn't support fp8 at the moment.
+            line_vals=["eager", "fused"],
+            line_names=["Eager", "Fused"],
+            styles=[("green", "-"), ("blue", "-")],
+            ylabel="Latency(ms)",  # Label name for the y-axis
+            plot_name="fused_ffn-benchmark",
+            args={},
+        )
+    )
+
+    @triton.testing.perf_report(configs)
+    def benchmark(B_T, H_D, D, provider):
+        # breakpoint()
+        x = torch.randn((B_T, D), dtype=torch.bfloat16, device="cuda")
+        w13 = torch.randn((H_D * 2, D), dtype=torch.bfloat16, device="cuda")
+        w2 = torch.randn((H_D, D), dtype=torch.bfloat16, device="cuda")
+        quantiles = [0.5, 0.2, 0.8]
+        if provider == "eager":
+            return triton.testing.do_bench(
+                lambda: eager_ffn(x, w13, w2), quantiles=quantiles
+            )
+        if provider == "fused":
+            return triton.testing.do_bench(
+                lambda: fused_ffn(x, w13, w2), quantiles=quantiles
+            )
+
+    benchmark.run(show_plots=True, print_data=True)
+
+
+if __name__ == "__main__":
+    # B_T, H_D, D
+    nunerics_check((64, 128, 128))
+    # nunerics_check((256, 8192, 2048))
+    # do_benchmark()