Skip to content
KernelIndex
Search⌘K

submission 16190

KernelBot user 7 · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 1 lines, June 9 Researcher Reciprocity License v1.0.

amd-fp8-mm-hip-v12.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-fp8-mm-16190?include=source"
interfacepython
Compatibility
measured onAMD Instinct MI300X
declared hardwareAMD Instinct MI300X
architecturesgfx942
dtypesfp32, fp8_e4m3fnuz

Benchmark evidence

18 measurements across 1 GPU, fastest first.

Operation / workload
Hardware
Latency
Rank
Observed
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 256]
AMD Instinct MI300X
49.9µs
#59 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 512]
AMD Instinct MI300X
53.0µs
#61 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 1536]
AMD Instinct MI300X
82.6µs
#59 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 256]
AMD Instinct MI300X
112.5µs
#41 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 512]
AMD Instinct MI300X
121.6µs
#46 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=576 · seed=12346
AMD Instinct MI300X
135.2µs
#58 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=512 · seed=6563
AMD Instinct MI300X
136.2µs
#56 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 2048]
AMD Instinct MI300X
173.4µs
#71 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=1536 · seed=8135
AMD Instinct MI300X
178.1µs
#58 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 2304]
AMD Instinct MI300X
196.2µs
#71 of 98
2025-05-05
Show all 18 measurements ›
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 1536]
AMD Instinct MI300X
226.1µs
#68 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=4608 · seed=7531
AMD Instinct MI300X
408.0µs
#74 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=512 · seed=12341
AMD Instinct MI300X
510.5µs
#79 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=576 · seed=9863
AMD Instinct MI300X
510.6µs
#77 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 2048]
AMD Instinct MI300X
528.9µs
#54 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 2304]
AMD Instinct MI300X
634.1µs
#61 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=1536 · seed=6543
AMD Instinct MI300X
731.8µs
#74 of 98
2025-05-05
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=4608 · seed=65436
AMD Instinct MI300X
1.39ms
#65 of 98
2025-05-05

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:40bfe394aa66a192cdc1b122742ce8faf2f1d71983d131a25c59c71f80c8b5c3
license declaredunknown
license concludedunknown
imported2026-08-15

Kernel source

amd-fp8-mm-hip-v12.py1 lines
# This script provides a template for using load_inline to run a HIP kernel for
from torch.utils.cpp_extension import load_inline
from task import input_t, output_t
CPP_WRAPPER = """
void fp8_mm(torch::Tensor a, torch::Tensor b, torch::Tensor as, torch::Tensor bs, torch::Tensor c);
"""

CUDA_SRC = """
#include <hip/hip_runtime.h>
#include <hip/amd_detail/amd_hip_fp8.h>
#include <hip/amd_detail/amd_hip_bf16.h>

constexpr const int BLOCK = 128;

__global__ void custom_kernel(const __hip_fp8_e4m3_fnuz* a, const __hip_fp8_e4m3_fnuz* b, const float* as, const float* bs,
                   __hip_bfloat16* c, int m, int n, int k) {

    // Your implementation here
    // int cx = threadIdx.x + blockDim.x * blockIdx.x;
    // int cy = threadIdx.y + blockDim.y * blockIdx.y;
    // if(cx >= m || cy >= n) return;

    using float4 = float __attribute__((ext_vector_type(4)));

    __shared__ long lds_a[64][17];
    __shared__ long lds_b[64][17];

    int thread_block_m = blockIdx.x * 64;
    int thread_block_n = blockIdx.y * 64;

    int wave_m = (threadIdx.y / 2) * 16;
    int wave_n = (threadIdx.y % 2) * 32;

    int thread_c_m = (threadIdx.x / 16) * 4;
    int thread_c_n =  threadIdx.x % 16; // by float

    int cp_shared_mn = threadIdx.y * 8 + threadIdx.x / 8;
    int cp_shared_k  = (threadIdx.x % 8) * 2; // by long

    int cp_global_m = thread_block_m + cp_shared_mn;
    int cp_global_n = thread_block_n + cp_shared_mn;
    int cp_global_k = cp_shared_k;

    int is_lgd_a = cp_global_m < m;
    int is_lgd_b = cp_global_n < n;

    int lds_m = wave_m + threadIdx.x % 16;
    int lds_n = wave_n + threadIdx.x % 16;
    int lds_k = threadIdx.x / 16; // by long

    int c_m = thread_block_m + wave_m + thread_c_m;
    int c_n = thread_block_n + wave_n + thread_c_n;

    float4 *a_ptr = &(((float4 *)a)[cp_global_m * k/16 + cp_global_k/2]);
    float4 *b_ptr = &(((float4 *)b)[cp_global_n * k/16 + cp_global_k/2]);

    float4 tile_d[2] = {0};

    int as_m =  thread_block_m + wave_m + thread_c_m;
    int bs_n = (thread_block_n + wave_n + thread_c_n) / BLOCK;
    float *as_ptr = (float*)&as[as_m];
    float *bs_ptr = (float*)&bs[bs_n];

    for (int ib = 0; ib < k/BLOCK; ib++) {
        float4 ldg_a = is_lgd_a ? a_ptr[0] : 0;
        float4 ldg_b = is_lgd_b ? b_ptr[0] : 0;
        a_ptr += 8;
        b_ptr += 8;
        float4 this_as = *(float4 *)(&as_ptr[0]);
        float  this_bs = bs_ptr[0];
        as_ptr += m;
        bs_ptr += (n+BLOCK-1)/BLOCK;
        *(float4 *)(&lds_a[cp_shared_mn][cp_shared_k]) = ldg_a;
        *(float4 *)(&lds_b[cp_shared_mn][cp_shared_k]) = ldg_b;
        __syncthreads();
        float4 block_tile_d[2] = {0};
        block_tile_d[0] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  0], lds_b[lds_n][lds_k +  0], block_tile_d[0], 0, 0, 0);
        block_tile_d[0] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  4], lds_b[lds_n][lds_k +  4], block_tile_d[0], 0, 0, 0);
        block_tile_d[0] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  8], lds_b[lds_n][lds_k +  8], block_tile_d[0], 0, 0, 0);
        block_tile_d[0] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k + 12], lds_b[lds_n][lds_k + 12], block_tile_d[0], 0, 0, 0);
        tile_d[0][0] += block_tile_d[0][0] * this_as[0] * this_bs;
        tile_d[0][1] += block_tile_d[0][1] * this_as[1] * this_bs;
        tile_d[0][2] += block_tile_d[0][2] * this_as[2] * this_bs;
        tile_d[0][3] += block_tile_d[0][3] * this_as[3] * this_bs;
        block_tile_d[1] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  0], lds_b[lds_n+16][lds_k +  0], block_tile_d[1], 0, 0, 0);
        block_tile_d[1] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  4], lds_b[lds_n+16][lds_k +  4], block_tile_d[1], 0, 0, 0);
        block_tile_d[1] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k +  8], lds_b[lds_n+16][lds_k +  8], block_tile_d[1], 0, 0, 0);
        block_tile_d[1] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(lds_a[lds_m][lds_k + 12], lds_b[lds_n+16][lds_k + 12], block_tile_d[1], 0, 0, 0);
        tile_d[1][0] += block_tile_d[1][0] * this_as[0] * this_bs;
        tile_d[1][1] += block_tile_d[1][1] * this_as[1] * this_bs;
        tile_d[1][2] += block_tile_d[1][2] * this_as[2] * this_bs;
        tile_d[1][3] += block_tile_d[1][3] * this_as[3] * this_bs;
        __syncthreads();
    }

    // finally, write the result as bf16
    c[(c_m + 0) * n + c_n] = (__hip_bfloat16)(tile_d[0][0]);
    c[(c_m + 1) * n + c_n] = (__hip_bfloat16)(tile_d[0][1]);
    c[(c_m + 2) * n + c_n] = (__hip_bfloat16)(tile_d[0][2]);
    c[(c_m + 3) * n + c_n] = (__hip_bfloat16)(tile_d[0][3]);
    c[(c_m + 0) * n + c_n + 16] = (__hip_bfloat16)(tile_d[1][0]);
    c[(c_m + 1) * n + c_n + 16] = (__hip_bfloat16)(tile_d[1][1]);
    c[(c_m + 2) * n + c_n + 16] = (__hip_bfloat16)(tile_d[1][2]);
    c[(c_m + 3) * n + c_n + 16] = (__hip_bfloat16)(tile_d[1][3]);
}

void fp8_mm(torch::Tensor a, torch::Tensor b, torch::Tensor as, torch::Tensor bs, torch::Tensor c) {
    int m = a.size(0);
    int n = b.size(0);
    int k = a.size(1);
    custom_kernel<<<dim3((m+63)/64, (n+63)/64), dim3(64, 8), 0, 0>>> ((__hip_fp8_e4m3_fnuz*)a.data_ptr(), (__hip_fp8_e4m3_fnuz*)b.data_ptr(),
    as.data_ptr<float>(), bs.data_ptr<float>(), (__hip_bfloat16*)c.data_ptr(), m, n, k);
    //C10_CUDA_CHECK(cudaGetLastError());
}
"""

import os
os.environ["CXX"] = "clang++"

module = load_inline(
    name='fp8_mm',
    cpp_sources=[CPP_WRAPPER],
    cuda_sources=[CUDA_SRC],
    functions=['fp8_mm'],
    verbose=True,
    extra_cuda_cflags=["--offload-arch=gfx942", "-std=c++20", "-O3"],
)


def custom_kernel(data: input_t) -> output_t:
    a, b, a_scale, b_scale, c = data
    a = a.contiguous()
    b = b.contiguous()
    module.fp8_mm(a, b, a_scale, b_scale, c)
    return c



Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0

Changes from previous submission

Against this author's previous submission submission 16188.

Best evidence level for this revision: reported

JSON