ZLUDA v3.8.7 (#66)

* Add dummy cuFFTW library. * Bump version. * Implement fft functions required to run torch fftn, ifftn, and rfftn.
lshqqytiger · Jan 15, 2025 · c4994b3 · c4994b3
1 parent d60bddb
commit c4994b3
Show file tree

Hide file tree

Showing 14 changed files with 1,049 additions and 193 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/Cargo.toml b/Cargo.toml
@@ -1,65 +1,66 @@
-[workspace]
-
-resolver = "2"
-
-# Remember to also update the project's Cargo.toml
-# if it's a top-level project
-members = [
-    "atiadlxx-sys",
-    "comgr",
-    "cuda_base",
-    "cuda_types",
-    "detours-sys",
-    "ext/llvm-sys.rs",
-    "hip_common",
-    "hip_runtime-sys",
-    "hipblaslt-sys",
-    "hipfft-sys",
-    "hiprt-sys",
-    "miopen-sys",
-    "offline_compiler",
-    "optix_base",
-    "optix_dump",
-    "process_address_table",
-    "ptx",
-    "rocblas-sys",
-    "rocm_smi-sys",
-    "rocsparse-sys",
-    "xtask",
-    "zluda",
-    "zluda_api",
-    "zluda_blas",
-    "zluda_blaslt",
-    "zluda_ccl",
-    "zluda_dark_api",
-    "zluda_dnn",
-    "zluda_dump",
-    "zluda_fft",
-    "zluda_inject",
-    "zluda_lib",
-    "zluda_llvm",
-    "zluda_ml",
-    "zluda_redirect",
-    "zluda_rt",
-    "zluda_rtc",
-    "zluda_runtime",
-    "zluda_sparse",
-]
-
-# Cargo does not support OS-specific or profile-specific
-# targets. We keep list here to bare minimum and rely on xtask
-default-members = [
-    "zluda_lib",
-    "zluda_ml",
-    "zluda_inject",
-    "zluda_redirect"
-]
-
-[profile.dev.package.blake3]
-opt-level = 3
-
-[profile.dev.package.lz4-sys]
-opt-level = 3
-
-[profile.dev.package.xtask]
-opt-level = 2
+[workspace]
+
+resolver = "2"
+
+# Remember to also update the project's Cargo.toml
+# if it's a top-level project
+members = [
+    "atiadlxx-sys",
+    "comgr",
+    "cuda_base",
+    "cuda_types",
+    "detours-sys",
+    "ext/llvm-sys.rs",
+    "hip_common",
+    "hip_runtime-sys",
+    "hipblaslt-sys",
+    "hipfft-sys",
+    "hiprt-sys",
+    "miopen-sys",
+    "offline_compiler",
+    "optix_base",
+    "optix_dump",
+    "process_address_table",
+    "ptx",
+    "rocblas-sys",
+    "rocm_smi-sys",
+    "rocsparse-sys",
+    "xtask",
+    "zluda",
+    "zluda_api",
+    "zluda_blas",
+    "zluda_blaslt",
+    "zluda_ccl",
+    "zluda_dark_api",
+    "zluda_dnn",
+    "zluda_dump",
+    "zluda_fft",
+    "zluda_fftw",
+    "zluda_inject",
+    "zluda_lib",
+    "zluda_llvm",
+    "zluda_ml",
+    "zluda_redirect",
+    "zluda_rt",
+    "zluda_rtc",
+    "zluda_runtime",
+    "zluda_sparse",
+]
+
+# Cargo does not support OS-specific or profile-specific
+# targets. We keep list here to bare minimum and rely on xtask
+default-members = [
+    "zluda_lib",
+    "zluda_ml",
+    "zluda_inject",
+    "zluda_redirect"
+]
+
+[profile.dev.package.blake3]
+opt-level = 3
+
+[profile.dev.package.lz4-sys]
+opt-level = 3
+
+[profile.dev.package.xtask]
+opt-level = 2
diff --git a/hipblaslt-sys/src/lib.rs b/hipblaslt-sys/src/lib.rs
@@ -11,6 +11,7 @@ impl hipblasOperation_t {
 impl hipblasOperation_t {
     pub const HIPBLAS_OP_C: hipblasOperation_t = hipblasOperation_t(113);
 }
+#[allow(non_camel_case_types)]
 #[repr(transparent)]
 #[derive(Copy, Clone, Hash, PartialEq, Eq)]
 pub struct hipblasOperation_t(pub ::std::os::raw::c_int);
diff --git a/zluda_blas/src/lib.rs b/zluda_blas/src/lib.rs
@@ -1,6 +1,8 @@
-#![allow(warnings)]
+#[allow(warnings)]
 mod common;
+#[allow(warnings)]
 mod cublas;
+#[allow(warnings)]
 mod cublasxt;
 
 pub use common::*;
@@ -13,7 +15,7 @@ use rocsolver_sys::{
     rocsolver_cgetrf_batched, rocsolver_cgetri_outofplace_batched, rocsolver_dgetrs_batched,
     rocsolver_sgetrs_batched, rocsolver_zgetrf_batched, rocsolver_zgetri_outofplace_batched,
 };
-use std::{mem, ptr};
+use std::ptr;
 
 #[cfg(debug_assertions)]
 pub(crate) fn unsupported() -> cublasStatus_t {
@@ -223,61 +225,20 @@ unsafe fn set_stream(handle: cublasHandle_t, stream_id: cudaStream_t) -> cublasS
         ) -> CUresult>(b"cuGetExportTable\0")
         .unwrap();
     let mut export_table = ptr::null();
-    (cu_get_export_table)(&mut export_table, &zluda_dark_api::ZludaExt::GUID);
+    assert_eq!(
+        (cu_get_export_table)(&mut export_table, &zluda_dark_api::ZludaExt::GUID),
+        CUresult::CUDA_SUCCESS
+    );
     let zluda_ext = zluda_dark_api::ZludaExt::new(export_table);
     let stream: Result<_, _> = zluda_ext.get_hip_stream(stream_id as _).into();
     to_cuda(rocblas_set_stream(handle as _, stream.unwrap() as _))
 }
 
-fn set_math_mode(handle: cublasHandle_t, mode: cublasMath_t) -> cublasStatus_t {
+fn set_math_mode(_handle: cublasHandle_t, _mode: cublasMath_t) -> cublasStatus_t {
     // llama.cpp uses CUBLAS_TF32_TENSOR_OP_MATH
     cublasStatus_t::CUBLAS_STATUS_SUCCESS
 }
 
-unsafe fn sgemm(
-    transa: std::ffi::c_char,
-    transb: std::ffi::c_char,
-    m: i32,
-    n: i32,
-    k: i32,
-    alpha: f32,
-    a: *const f32,
-    lda: i32,
-    b: *const f32,
-    ldb: i32,
-    beta: f32,
-    c: *mut f32,
-    ldc: i32,
-) -> cublasStatus_t {
-    let mut handle = mem::zeroed();
-    let mut status = to_cuda(rocblas_create_handle(handle));
-    if status != cublasStatus_t::CUBLAS_STATUS_SUCCESS {
-        return status;
-    }
-    let transa = op_from_cuda(cublasOperation_t(transa as _));
-    let transb = op_from_cuda(cublasOperation_t(transb as _));
-    status = to_cuda(rocblas_sgemm(
-        handle.cast(),
-        transa,
-        transb,
-        m,
-        n,
-        k,
-        &alpha,
-        a,
-        lda,
-        b,
-        ldb,
-        &beta,
-        c,
-        ldc,
-    ));
-    if status != cublasStatus_t::CUBLAS_STATUS_SUCCESS {
-        return status;
-    }
-    to_cuda(rocblas_destroy_handle(*handle))
-}
-
 unsafe fn sgemm_v2(
     handle: cublasHandle_t,
     transa: cublasOperation_t,
@@ -495,7 +456,7 @@ unsafe fn gemm_ex(
     ))
 }
 
-fn to_algo(algo: cublasGemmAlgo_t) -> rocblas_gemm_algo_ {
+fn to_algo(_algo: cublasGemmAlgo_t) -> rocblas_gemm_algo_ {
     // only option
     rocblas_gemm_algo::rocblas_gemm_algo_standard
 }
@@ -807,7 +768,7 @@ unsafe fn sgetrs_batched(
     dev_ipiv: *const i32,
     b: *const *mut f32,
     ldb: i32,
-    info: *mut i32,
+    _info: *mut i32,
     batch_size: i32,
 ) -> cublasStatus_t {
     let trans = op_from_cuda_for_solver(trans);
@@ -837,7 +798,7 @@ unsafe fn dgetrs_batched(
     dev_ipiv: *const i32,
     b: *const *mut f64,
     ldb: i32,
-    info: *mut i32,
+    _info: *mut i32,
     batch_size: i32,
 ) -> cublasStatus_t {
     let trans = op_from_cuda_for_solver(trans);
@@ -1048,50 +1009,6 @@ unsafe fn dger(
     ))
 }
 
-unsafe fn dgemm(
-    transa: std::ffi::c_char,
-    transb: std::ffi::c_char,
-    m: i32,
-    n: i32,
-    k: i32,
-    alpha: f64,
-    a: *const f64,
-    lda: i32,
-    b: *const f64,
-    ldb: i32,
-    beta: f64,
-    c: *mut f64,
-    ldc: i32,
-) -> cublasStatus_t {
-    let mut handle = mem::zeroed();
-    let mut status = to_cuda(rocblas_create_handle(handle));
-    if status != cublasStatus_t::CUBLAS_STATUS_SUCCESS {
-        return status;
-    }
-    let transa = op_from_cuda(cublasOperation_t(transa as _));
-    let transb = op_from_cuda(cublasOperation_t(transb as _));
-    status = to_cuda(rocblas_dgemm(
-        handle.cast(),
-        transa,
-        transb,
-        m,
-        n,
-        k,
-        &alpha,
-        a,
-        lda,
-        b,
-        ldb,
-        &beta,
-        c,
-        ldc,
-    ));
-    if status != cublasStatus_t::CUBLAS_STATUS_SUCCESS {
-        return status;
-    }
-    to_cuda(rocblas_destroy_handle(*handle))
-}
-
 unsafe fn dgemm_v2(
     handle: *mut cublasContext,
     transa: cublasOperation_t,

diff --git a/zluda_fft/src/cufft.rs b/zluda_fft/src/cufft.rs
@@ -380,15 +380,15 @@ pub unsafe extern "system" fn cufftSetWorkArea(
     plan: cufftHandle,
     workArea: *mut ::std::os::raw::c_void,
 ) -> cufftResult {
-    crate::unsupported()
+    crate::set_work_area(plan, workArea)
 }
 
 #[no_mangle]
 pub unsafe extern "system" fn cufftSetAutoAllocation(
     plan: cufftHandle,
     autoAllocate: ::std::os::raw::c_int,
 ) -> cufftResult {
-    crate::unsupported()
+    crate::set_auto_allocation(plan, autoAllocate)
 }
 
 #[no_mangle]

diff --git a/zluda_fft/src/cufftxt.rs b/zluda_fft/src/cufftxt.rs
@@ -376,7 +376,22 @@ pub unsafe extern "system" fn cufftXtMakePlanMany(
     workSize: *mut usize,
     executiontype: cudaDataType,
 ) -> cufftResult {
-    crate::unsupported()
+    crate::xt_make_plan_many(
+        plan,
+        rank,
+        n,
+        inembed,
+        istride,
+        idist,
+        inputtype,
+        onembed,
+        ostride,
+        odist,
+        outputtype,
+        batch,
+        workSize,
+        executiontype,
+    )
 }
 
 #[no_mangle]
@@ -406,7 +421,7 @@ pub unsafe extern "system" fn cufftXtExec(
     output: *mut ::std::os::raw::c_void,
     direction: ::std::os::raw::c_int,
 ) -> cufftResult {
-    crate::unsupported()
+    crate::xt_exec(plan, input, output, direction)
 }
 
 #[no_mangle]