Merge branch 'main' of http://192.168.0.100:3000/squall/torch_ext

csrc/core.h

@@ -1,6 +1,15 @@
 #ifndef CORE_H
 #define CORE_H
 #include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)           \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
+    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+    AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
 #define TYPING_DISPATCH(scalar_t, ...) \
     switch (scalar_t)                  \
@@ -49,5 +58,6 @@ void test_cute_tensor();
 void md_mm(const torch::Tensor &src);
 void block_sum(const torch::Tensor &src, torch::Tensor &dest);
 void md_block_sum(const torch::Tensor &src, torch::Tensor &dest);
+void softmax(const torch::Tensor &src, torch::Tensor &dest);
 void rms_norm(torch::Tensor &states, float eps, float gamma);
 #endif

csrc/core_bind.cpp

@@ -18,5 +18,6 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
   m.def("md_mm", &md_mm, "just a test of multi dimension mm");
   m.def("block_sum", &block_sum, "test block sum");
   m.def("md_block_sum", &md_block_sum, "multi dimension block sum");
+  m.def("softmax", &softmax, "test softmax example");
   m.def("rms_norm", &rms_norm, "rms noram");
 }

csrc/md.cu

@@ -117,15 +117,133 @@ __global__ void md_row_sum_kernel(const float *src, float *dest, int stride_a, i
 void md_block_sum(const torch::Tensor &src, torch::Tensor &dest)
 {
     int block_size = 1024;
+
     dim3 grid(src.size(0), src.size(1), (src.size(2) + block_size - 1) / block_size);
     dim3 block(block_size);
-    md_row_sum_kernel<<<grid, block>>>(src.data_ptr<float>(),
-                                       dest.data_ptr<float>(),
-                                       src.stride(0),
-                                       src.stride(1),
-                                       src.size(0),
-                                       src.size(1),
-                                       src.size(2));
+
+    printf("this is the device num:%d\n", src.get_device());
+    int dev = src.get_device();
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream(dev);
+    md_row_sum_kernel<<<grid, block, 0, stream>>>(src.data_ptr<float>(),
+                                                  dest.data_ptr<float>(),
+                                                  src.stride(0),
+                                                  src.stride(1),
+                                                  src.size(0),
+                                                  src.size(1),
+                                                  src.size(2));
+}
+
+void interaction(const torch::Tensor &src)
+{
+    int block_size = 1024;
+
+    dim3 grid(src.size(0), src.size(1), (src.size(2) + block_size - 1) / block_size);
+    dim3 block(block_size);
+
+    printf("this is the device num:%d\n", src.get_device());
+    int dev = src.get_device();
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream(dev);
+    // seems can do some other things.
+}
+
+template <typename s_scalar>
+__device__ s_scalar exp(s_scalar a)
+{
+    return expf(a);
+}
+
+template <>
+__device__ __nv_bfloat16 exp(__nv_bfloat16 a)
+{
+    float tmp = __bfloat162float(a);
+    float tmp_score = expf(tmp);
+    return __float2bfloat16(tmp_score);
+}
+
+template <>
+__device__ __half exp(__half a)
+{
+    float tmp = __half2float(a);
+    float tmp_score = expf(tmp);
+    return __float2half(tmp_score);
+}
+
+template <>
+__device__ float exp(float a)
+{
+    return expf(a);
+}
+
+template <typename scalar_t>
+__device__ float fi_cast(scalar_t a)
+{
+    return a;
+}
+
+template <>
+__device__ float fi_cast(__nv_bfloat16 a)
+{
+    return __bfloat162float(a);
+}
+
+template <>
+__device__ float fi_cast(__half a)
+{
+    return __half2float(a);
+}
+
+template <int BLOCK_SIZE, typename scalar_t>
+__global__ void softmax_kernel(const scalar_t *src, scalar_t *dest, int hidden_dim)
+{
+    int tid = threadIdx.x;
+    int offset = blockIdx.x * hidden_dim;
+    __shared__ scalar_t smem[BLOCK_SIZE];
+    float local_sum = 0.0f;
+    for (int i = tid; i < hidden_dim; i += blockDim.x)
+    {
+        // sum the res;
+        int tmp_index = offset + i;
+        scalar_t tmp_score = exp(src[tmp_index]);
+        dest[tmp_index] = tmp_score;
+        local_sum += tmp_score;
+    }
+    if (tid < BLOCK_SIZE)
+        smem[tid] = local_sum;
+    else
+        smem[tid] = 0.0f;
+    __syncthreads();
+    typedef cub::BlockReduce<scalar_t, BLOCK_SIZE> BlockReduce;
+    __shared__ typename BlockReduce::TempStorage temp_storage;
+    scalar_t sum = BlockReduce(temp_storage).Sum(smem[tid]);
+    // remember the block reduce sum means only the first thread has the real sum.
+    if (tid == 0)
+        smem[0] = sum;
+    __syncthreads();
+    for (int i = tid; i < hidden_dim; i += blockDim.x)
+    {
+        int tmp_index = offset + i;
+        scalar_t tmp_score = dest[tmp_index] / smem[0];
+        dest[tmp_index] = tmp_score;
+    }
+}
+
+void softmax(const torch::Tensor &src, torch::Tensor &dest)
+{
+    int batch_num = src.size(0);
+    int hidden_dim = src.size(1);
+    int block_size = 1024;
+    dim3 grid(batch_num);
+    dim3 block(block_size);
+    VLLM_DISPATCH_FLOATING_TYPES(
+        src.scalar_type(), "softmax",
+        [&]
+        { 
+            int dev = src.get_device();
+            cudaStream_t stream = at::cuda::getCurrentCUDAStream(dev);
+            softmax_kernel<1024, scalar_t><<<grid, block, 0, stream>>>(
+                                       src.data_ptr<scalar_t>(),
+                                       dest.data_ptr<scalar_t>(),
+                                       hidden_dim); });
 }
 
 template <int head_num = 8>

csrc/quantize.cu

@@ -0,0 +1,36 @@
+#include <cuda_fp16.h>
+#include <cuda_fp8.h>
+
+__global__ void quantize(const half *src, __nv_fp8_storage_t *dest, int x_len, int y_len)
+{
+    int x_start = threadIdx.x * blockDim.x;
+    int y_start = threadIdx.y * blockDim.y;
+    __shared__ half max_value;
+
+    max_value = __float2half(-10000.0f);
+    for (int i = 0; i < blockDim.x; i++)
+    {
+        for (int j = 0; j < blockDim.x; j++)
+        {
+            if (x_start + i < x_len && y_start + j < y_len)
+            {
+
+                int real_offset = (y_start + j) * x_len + x_start + i;
+                max_value = __hmax(src[real_offset], max_value);
+            }
+        }
+    }
+    for (int i = 0; i < blockDim.x; i++)
+    {
+        for (int j = 0; j < blockDim.y; j++)
+        {
+            if (x_start + i < x_len && y_start + j < y_len)
+            {
+
+                int real_offset = (y_start + j) * x_len + x_start + i;
+                half tmp = __hdiv(src[real_offset], max_value);
+                dest[real_offset] = __nv_cvt_halfraw_to_fp8(__nv_half_raw(tmp), __NV_SATFINITE, __NV_E5M2);
+            }
+        }
+    }
+}

setup.py

@@ -11,6 +11,7 @@ files = [
     "csrc/core_bind.cpp",
     "csrc/max.cu",
     "csrc/md.cu",
+    "csrc/quantize.cu",
     "csrc/layernorm.cu",
 ]
 extension = CUDAExtension(

test_reducemax.py

@@ -2,7 +2,7 @@ import torch
 import torch_cuda_ext.core as core
 
 n = 1000000
-for i in range(1000):
+for i in range(100):
     src = torch.randn(size=(n,)).float().cuda()
     dest_n = int((n + 1024 - 1) / 1024)
     dest = torch.zeros(size=(dest_n,)).float().cuda()
@@ -31,3 +31,48 @@ core.md_block_sum(src, dest)
 real_sum = src.sum(dim=-1)
 diff = real_sum - dest
 print(diff)
+
+for k in range(128, 4096, 128):
+    for j in range(1024, 4096, 1024):
+        a = torch.randn(size=(k, j)).half().cuda()
+        b = torch.empty_like(a)
+        num_runs = 100
+        times = []
+        for _ in range(num_runs):
+            start = torch.cuda.Event(enable_timing=True)
+            end = torch.cuda.Event(enable_timing=True)
+            start.record()
+            core.softmax(a, b)
+            end.record()
+            torch.cuda.synchronize()  # 等待 CUDA 操作完成
+            elapsed_time = start.elapsed_time(end) / 1000  # 转换为秒
+            times.append(elapsed_time)
+
+        own_avg_time = sum(times) / num_runs
+        own_std_time = (sum((t - own_avg_time) ** 2 for t in times) / num_runs) ** 0.5
+        print(f"own softmax cost time: {own_avg_time}, {own_std_time}")
+
+        times = []
+        for _ in range(num_runs):
+            start = torch.cuda.Event(enable_timing=True)
+            end = torch.cuda.Event(enable_timing=True)
+            start.record()
+
+            res = torch.softmax(a, dim=1)
+            end.record()
+            torch.cuda.synchronize()  # 等待 CUDA 操作完成
+            elapsed_time = start.elapsed_time(end) / 1000  # 转换为秒
+            times.append(elapsed_time)
+
+        avg_time = sum(times) / num_runs
+        std_time = (sum((t - avg_time) ** 2 for t in times) / num_runs) ** 0.5
+        print(f"torch softmax cost time: {avg_time}, {std_time}")
+
+        # print("this is b", b)
+        diff = (res - b).abs().max()
+        if diff < 1e-4:
+            print("softmax is good")
+            time_diff_rate = (own_avg_time - avg_time) / avg_time
+            print(f"{k}, {j} matrix result {time_diff_rate}")
+        else:
+            print("softmax is not equal")

tests/test_mma.py

@@ -0,0 +1,468 @@
+import torch
+
+import triton
+import triton.language as tl
+
+DEVICE = "cuda"
+
+
+def is_cuda():
+    return triton.runtime.driver.active.get_current_target().backend == "cuda"
+
+
+def is_hip_cdna2():
+    target = triton.runtime.driver.active.get_current_target()
+    return target.backend == "hip" and target.arch == "gfx90a"
+
+
+def get_cuda_autotune_config():
+    return [
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 64,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=3,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 32,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 32,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=5,
+            num_warps=2,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 32,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=5,
+            num_warps=2,
+        ),
+        # Good config for fp8 inputs.
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=3,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 256,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=3,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 256,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 64,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 64,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 32,
+                "BLOCK_SIZE_K": 64,
+                "GROUP_SIZE_M": 8,
+            },
+            num_stages=4,
+            num_warps=4,
+        ),
+    ]
+
+
+def get_hip_autotune_config():
+    return [
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 16,
+                "GROUP_SIZE_M": 1,
+                "waves_per_eu": 2,
+            },
+            num_warps=4,
+            num_stages=2,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 256,
+                "BLOCK_SIZE_N": 256,
+                "BLOCK_SIZE_K": 16,
+                "GROUP_SIZE_M": 4,
+                "waves_per_eu": 2,
+            },
+            num_warps=8,
+            num_stages=2,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 1,
+                "waves_per_eu": 2,
+            },
+            num_warps=8,
+            num_stages=2,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 128,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 8,
+                "waves_per_eu": 3,
+            },
+            num_warps=4,
+            num_stages=2,
+        ),
+        triton.Config(
+            {
+                "BLOCK_SIZE_M": 64,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 32,
+                "GROUP_SIZE_M": 1,
+                "waves_per_eu": 8,
+            },
+            num_warps=4,
+            num_stages=2,
+        ),
+    ]
+
+
+def get_autotune_config():
+    if is_cuda():
+        return get_cuda_autotune_config()
+    else:
+        return get_hip_autotune_config()
+
+
+# `triton.jit`'ed functions can be auto-tuned by using the `triton.autotune` decorator, which consumes:
+#   - A list of `triton.Config` objects that define different configurations of
+#       meta-parameters (e.g., `BLOCK_SIZE_M`) and compilation options (e.g., `num_warps`) to try
+#   - An auto-tuning *key* whose change in values will trigger evaluation of all the
+#       provided configs
+@triton.autotune(
+    configs=get_autotune_config(),
+    key=["M", "N", "K"],
+)
+@triton.jit
+def matmul_kernel(
+    # Pointers to matrices
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    # Matrix dimensions
+    M,
+    N,
+    K,
+    # The stride variables represent how much to increase the ptr by when moving by 1
+    # element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
+    # by to get the element one row down (A has M rows).
+    stride_am,
+    stride_ak,  #
+    stride_bk,
+    stride_bn,  #
+    stride_cm,
+    stride_cn,
+    # Meta-parameters
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,  #
+    GROUP_SIZE_M: tl.constexpr,  #
+    ACTIVATION: tl.constexpr,  #
+):
+    """Kernel for computing the matmul C = A x B.
+    A has shape (M, K), B has shape (K, N) and C has shape (M, N)
+    """
+    # -----------------------------------------------------------
+    # Map program ids `pid` to the block of C it should compute.
+    # This is done in a grouped ordering to promote L2 data reuse.
+    # See above `L2 Cache Optimizations` section for details.
+    pid = tl.program_id(axis=0)
+    num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+    num_pid_in_group = GROUP_SIZE_M * num_pid_n
+    group_id = pid // num_pid_in_group
+    first_pid_m = group_id * GROUP_SIZE_M
+    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
+    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    # ----------------------------------------------------------
+    # Create pointers for the first blocks of A and B.
+    # We will advance this pointer as we move in the K direction
+    # and accumulate
+    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
+    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
+    # See above `Pointer Arithmetic` section for details
+    offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
+    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
+    b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
+
+    # -----------------------------------------------------------
+    # Iterate to compute a block of the C matrix.
+    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
+    # of fp32 values for higher accuracy.
+    # `accumulator` will be converted back to fp16 after the loop.
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        # Load the next block of A and B, generate a mask by checking the K dimension.
+        # If it is out of bounds, set it to 0.
+        a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
+        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
+        # We accumulate along the K dimension.
+        accumulator = tl.dot(a, b, accumulator)
+        # Advance the ptrs to the next K block.
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+    # You can fuse arbitrary activation functions here
+    # while the accumulator is still in FP32!
+    if ACTIVATION == "leaky_relu":
+        accumulator = leaky_relu(accumulator)
+    c = accumulator.to(tl.float16)
+
+    # -----------------------------------------------------------
+    # Write back the block of the output matrix C with masks.
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+# We can fuse `leaky_relu` by providing it as an `ACTIVATION` meta-parameter in `matmul_kernel`.
+@triton.jit
+def leaky_relu(x):
+    return tl.where(x >= 0, x, 0.01 * x)
+
+
+def matmul(a, b, activation=""):
+    # Check constraints.
+    assert a.shape[1] == b.shape[0], "Incompatible dimensions"
+    assert a.is_contiguous(), "Matrix A must be contiguous"
+    M, K = a.shape
+    K, N = b.shape
+    # Allocates output.
+    c = torch.empty((M, N), device=a.device, dtype=torch.float16)
+    # 1D launch kernel where each block gets its own program.
+    grid = lambda META: (
+        triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+    )
+    matmul_kernel[grid](
+        a,
+        b,
+        c,  #
+        M,
+        N,
+        K,  #
+        a.stride(0),
+        a.stride(1),  #
+        b.stride(0),
+        b.stride(1),  #
+        c.stride(0),
+        c.stride(1),  #
+        ACTIVATION=activation,  #
+    )
+    return c
+
+
+torch.manual_seed(0)
+a = torch.randn((512, 512), device=DEVICE, dtype=torch.float16)
+b = torch.randn((512, 512), device=DEVICE, dtype=torch.float16)
+triton_output = matmul(a, b)
+torch_output = torch.matmul(a, b)
+print(f"triton_output_with_fp16_inputs={triton_output}")
+print(f"torch_output_with_fp16_inputs={torch_output}")
+# Bigger tolerance for AMD CDNA2 devices.
+# CDNA2 devices use reduced precision fp16 and bf16 and flush input and
+# output denormal values to zero. Detailed info is at: https://pytorch.org/docs/stable/notes/numerical_accuracy.html#reduced-precision-fp16-and-bf16-gemms-and-convolutions-on-amd-instinct-mi200-devices
+rtol = 1e-2 if is_hip_cdna2() else 0
+if torch.allclose(triton_output, torch_output, atol=1e-2, rtol=rtol):
+    print("✅ Triton and Torch match")
+else:
+    print("❌ Triton and Torch differ")
+
+TORCH_HAS_FP8 = hasattr(torch, "float8_e5m2")
+if TORCH_HAS_FP8 and is_cuda():
+    torch.manual_seed(0)
+    a = torch.randn((512, 512), device=DEVICE, dtype=torch.float16)
+    b = torch.randn((512, 512), device=DEVICE, dtype=torch.float16)
+    a = a.to(torch.float8_e5m2)
+    # pre-transpose b for efficiency.
+    b = b.T
+    b = b.to(torch.float8_e5m2)
+    triton_output = matmul(a, b)
+    torch_output = torch.matmul(a.to(torch.float16), b.to(torch.float16))
+    print(f"triton_output_with_fp8_inputs={triton_output}")
+    print(f"torch_output_with_fp8_inputs={torch_output}")
+    if torch.allclose(triton_output, torch_output, atol=0.125, rtol=0):
+        print("✅ Triton and Torch match")
+    else:
+        print("❌ Triton and Torch differ")
+
+ref_lib = "cuBLAS" if is_cuda() else "rocBLAS"
+
+configs = []
+for fp8_inputs in [False, True]:
+    if fp8_inputs and (not TORCH_HAS_FP8 or not is_cuda()):
+        continue
+    configs.append(
+        triton.testing.Benchmark(
+            x_names=["M", "N", "K"],  # Argument names to use as an x-axis for the plot
+            x_vals=[
+                128 * i for i in range(2, 33)
+            ],  # 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=(
+                ["triton"] if fp8_inputs else [ref_lib.lower(), "triton"]
+            ),  # Label name for the lines
+            line_names=["Triton"] if fp8_inputs else [ref_lib, "Triton"],  # Line styles
+            styles=[("green", "-"), ("blue", "-")],
+            ylabel="TFLOPS",  # Label name for the y-axis
+            plot_name="matmul-performance-"
+            + (
+                "fp16" if not fp8_inputs else "fp8"
+            ),  # Name for the plot, used also as a file name for saving the plot.
+            args={"fp8_inputs": fp8_inputs},
+        )
+    )
+
+
+@triton.testing.perf_report(configs)
+def benchmark(M, N, K, provider, fp8_inputs):
+    a = torch.randn((M, K), device=DEVICE, dtype=torch.float16)
+    b = torch.randn((K, N), device=DEVICE, dtype=torch.float16)
+    if TORCH_HAS_FP8 and fp8_inputs:
+        a = a.to(torch.float8_e5m2)
+        b = b.T
+        b = b.to(torch.float8_e5m2)
+    quantiles = [0.5, 0.2, 0.8]
+    if provider == ref_lib.lower():
+        ms, min_ms, max_ms = triton.testing.do_bench(
+            lambda: torch.matmul(a, b), quantiles=quantiles
+        )
+    if provider == "triton":
+        ms, min_ms, max_ms = triton.testing.do_bench(
+            lambda: matmul(a, b), quantiles=quantiles
+        )
+    perf = lambda ms: 2 * M * N * K * 1e-12 / (ms * 1e-3)
+    return perf(ms), perf(max_ms), perf(min_ms)
+
+
+benchmark.run(show_plots=True, print_data=True)

tests/test_profille.py

@@ -0,0 +1,27 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.profiler import profile, record_function, ProfilerActivity
+
+# 定义模型和优化器
+model = nn.Linear(100, 10).cuda()
+optimizer = optim.SGD(model.parameters(), lr=0.01)
+
+# 启动 Profiler
+with profile(
+    activities=[ProfilerActivity.CUDA, ProfilerActivity.CPU],  # 监控 GPU 和 CPU
+    record_shapes=True,  # 记录张量形状
+    profile_memory=True,  # 分析内存使用
+    with_stack=True,  # 记录调用栈
+) as prof:
+    for _ in range(10):
+        x = torch.randn(64, 100).cuda()
+        y = model(x)
+        loss = y.sum()
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+
+# 输出分析结果
+print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
+prof.export_chrome_trace("./trace.json")

tests/test_triton.py

@@ -0,0 +1,84 @@
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def softmax_kernel(
+    output_ptr,
+    input_ptr,
+    input_row_stride,
+    output_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    # 获取当前程序的行索引
+    row_idx = tl.program_id(0)
+
+    # 计算输入和输出行的起始指针
+    row_start_ptr = input_ptr + row_idx * input_row_stride
+    output_row_start_ptr = output_ptr + row_idx * output_row_stride
+
+    # 将输入数据加载到本地内存
+    row_offsets = tl.arange(0, BLOCK_SIZE)
+    input_ptrs = row_start_ptr + row_offsets
+    row = tl.load(input_ptrs, mask=row_offsets < n_cols, other=-float("inf"))
+
+    # 计算 Softmax
+    row_minus_max = row - tl.max(row, axis=0)  # 数值稳定性：减去最大值
+    numerator = tl.exp(row_minus_max)
+    denominator = tl.sum(numerator, axis=0)
+    softmax_output = numerator / denominator
+
+    # 将结果写回输出
+    output_ptrs = output_row_start_ptr + row_offsets
+    tl.store(output_ptrs, softmax_output, mask=row_offsets < n_cols)
+
+
+def softmax(x):
+    n_rows, n_cols = x.shape
+
+    # 分配输出张量
+    output = torch.empty_like(x)
+
+    # 定义 GPU 内核的网格和块大小
+    BLOCK_SIZE = triton.next_power_of_2(n_cols)
+    num_warps = 4
+    if BLOCK_SIZE >= 2048:
+        num_warps = 8
+    if BLOCK_SIZE >= 4096:
+        num_warps = 16
+
+    # 启动 Triton 内核
+    softmax_kernel[(n_rows,)](
+        output,
+        x,
+        x.stride(0),
+        output.stride(0),
+        n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+
+    return output
+
+
+# 测试 Softmax
+if __name__ == "__main__":
+    # 创建一个随机矩阵
+    x = torch.randn(4, 16, device="cuda")
+
+    # 使用 Triton 计算 Softmax
+    output_triton = softmax(x)
+
+    # 使用 PyTorch 计算 Softmax 作为参考
+    output_torch = torch.softmax(x, dim=1)
+
+    # 检查结果是否一致
+    print("Input:")
+    print(x)
+    print("Triton Softmax:")
+    print(output_triton)
+    print("PyTorch Softmax:")
+    print(output_torch)
+    print(f"Are close: {torch.allclose(output_triton, output_torch, atol=1e-5)}")

tests/test_triton_mma.py

@@ -0,0 +1,176 @@
+# coding=utf-8
+import torch
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def add_kernel(a_ptr, b_ptr, c_ptr, numel, BLOCK_SIZE: tl.constexpr):
+    xidx = tl.program_id(0)
+    index = xidx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    mask = index < numel
+    a = tl.load(a_ptr + index, mask=mask)
+    b = tl.load(b_ptr + index, mask=mask)
+    c = a + b
+    tl.store(c_ptr + index, c, mask=mask)
+
+
+@triton.jit
+def add_mat_kernel(
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    stride_m,
+    stride_n,
+    m,
+    n,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+):
+    midx = tl.program_id(0)
+    nidx = tl.program_id(1)
+    m_offset = midx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    n_offset = nidx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    mask = (m_offset[:, None] < m) & (n_offset[None, :] < n)
+    index = m_offset[:, None] * stride_m + n_offset[None, :] * stride_n
+    a = tl.load(a_ptr + index, mask=mask)
+    b = tl.load(b_ptr + index, mask=mask)
+    c = a + b
+    tl.store(c_ptr + index, c, mask=mask)
+
+
+@triton.jit
+def threed_mat_kernel(
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    stride_1,
+    stride_m,
+    stride_n,
+    num_token,
+    m,
+    n,
+    TOKEN_BLOCK: tl.constexpr,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+):
+    token_idx = tl.program_id(0)
+    midx = tl.program_id(1)
+    nidx = tl.program_id(2)
+    # tl.device_print("token idx:", token_idx)
+    m_offset = midx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    n_offset = nidx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    mask = (m_offset[:, None] < m) & (n_offset[None, :] < n)
+    index = (
+        token_idx * stride_1
+        + m_offset[:, None] * stride_m
+        + n_offset[None, :] * stride_n
+    )
+    a = tl.load(a_ptr + index, mask=mask)
+    b = tl.load(b_ptr + index, mask=mask)
+    c = a + b
+    tl.store(c_ptr + index, c, mask=mask)
+
+
+@triton.jit
+def mma_kernel(
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    m,
+    n,
+    k,
+    stride_am,
+    stride_an,
+    stride_bm,
+    stride_bk,
+    stride_cm,
+    stride_ck,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+):
+    midx = tl.program_id(0)
+    nidx = tl.program_id(1)
+    a_m_offset = midx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    a_n_offset = nidx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    a_mask = (a_m_offset[:, None] < m) & (a_n_offset[None, :] < n)
+    a_index = a_m_offset[:, None] * stride_am + a_n_offset[None, :] * stride_an
+    a = tl.load(a_ptr + a_index, mask=a_mask)
+    acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+
+
+def test_add_kernel():
+    a = torch.randn(size=(1024,), device="cuda")
+    b = torch.randn(size=(1024,), device="cuda")
+    c = torch.empty_like(a)
+    BLOCK_SIZE = 32
+    grid = lambda meta: (triton.cdiv(a.numel(), meta["BLOCK_SIZE"]),)
+    add_kernel[grid](a, b, c, a.numel(), BLOCK_SIZE)
+    real_c = a + b
+    assert torch.allclose(real_c, c), "not equal"
+    print("all right")
+
+
+def test_add_mat_kernel():
+    a = torch.randn(size=(127, 255), device="cuda")
+    b = torch.randn(size=(127, 255), device="cuda")
+    c = torch.empty_like(a)
+    BLOCK_SIZE_M = 32
+    BLOCK_SIZE_N = 16
+    grid = lambda meta: (
+        triton.cdiv(a.size(0), meta["BLOCK_SIZE_M"]),
+        triton.cdiv(a.size(1), meta["BLOCK_SIZE_N"]),
+    )
+    add_mat_kernel[grid](
+        a,
+        b,
+        c,
+        a.stride(0),
+        a.stride(1),
+        a.size(0),
+        a.size(1),
+        BLOCK_SIZE_M,
+        BLOCK_SIZE_N,
+    )
+    real_c = a + b
+    assert torch.allclose(c, real_c), "not equal"
+    print("all right")
+
+
+def test_three_dimension():
+    num_token = 128
+    a = torch.randn(size=(num_token, 127, 255), device="cuda")
+    b = torch.randn(size=(num_token, 127, 255), device="cuda")
+    c = torch.empty_like(a)
+    BLOCK_SIZE_M = 32
+    BLOCK_SIZE_N = 16
+    TOKEN_BLOCK = a.size(0)
+    grid = lambda meta: (
+        a.size(0),
+        triton.cdiv(a.size(1), meta["BLOCK_SIZE_M"]),
+        triton.cdiv(a.size(2), meta["BLOCK_SIZE_N"]),
+    )
+    threed_mat_kernel[grid](
+        a,
+        b,
+        c,
+        a.stride(0),
+        a.stride(1),
+        a.stride(2),
+        a.size(0),
+        a.size(1),
+        a.size(2),
+        TOKEN_BLOCK,
+        BLOCK_SIZE_M,
+        BLOCK_SIZE_N,
+    )
+    real_c = a + b
+    assert torch.allclose(c, real_c), "not equal"
+    print("all right")
+
+
+if __name__ == "__main__":
+    test_add_kernel()
+    test_add_mat_kernel()
+    test_three_dimension()