Skip to content

Commit 4b4ba0f

Browse files
yeahdongcnyeahdongcn
committed
musa: enable muBlas and MMA
Signed-off-by: Xiaodong Ye <[email protected]>
1 parent 1f73301 commit 4b4ba0f

File tree

3 files changed

+85
-36
lines changed

3 files changed

+85
-36
lines changed

‎ggml/src/ggml-cuda/common.cuh‎

Lines changed: 20 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -76,17 +76,17 @@
7676
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
7777

7878
// Moore Threads
79-
#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
80-
81-
#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
82-
#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
83-
#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
79+
#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
80+
#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
81+
#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
8482

8583
#define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
8684
#define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
8785
#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
8886
#define GGML_CUDA_CC_IS_NG(cc) (cc >= GGML_CUDA_CC_NG)
8987

88+
#define GGML_CUDA_CC_IS_QY1_OR_EARLIER (__MUSA_ARCH__ < 220)
89+
9090
#ifdef __CUDA_ARCH_LIST__
9191
constexpr bool ggml_cuda_has_arch_impl(int) {
9292
return false;
@@ -203,9 +203,9 @@ typedef float2 dfloat2;
203203
#define FP16_AVAILABLE
204204
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
205205

206-
#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
206+
#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != GGML_CUDA_CC_DP4A
207207
#define FAST_FP16_AVAILABLE
208-
#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
208+
#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != GGML_CUDA_CC_DP4A
209209

210210
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
211211
#define FP16_MMA_AVAILABLE
@@ -215,6 +215,10 @@ typedef float2 dfloat2;
215215
#define FP16_MMA_AVAILABLE
216216
#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
217217

218+
#if defined(GGML_USE_MUSA)
219+
#define FP16_MMA_AVAILABLE
220+
#endif // defined(GGML_USE_MUSA)
221+
218222
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
219223
#define NEW_MMA_AVAILABLE
220224
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -223,21 +227,22 @@ typedef float2 dfloat2;
223227
#define CP_ASYNC_AVAILABLE
224228
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
225229

226-
#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
230+
#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_CC_IS_QY1_OR_EARLIER)
227231
#define FLASH_ATTN_AVAILABLE
228-
#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
232+
#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_CC_IS_QY1_OR_EARLIER)
229233

230234
static bool fp16_available(const int cc) {
231235
return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
232236
}
233237

234238
static bool fast_fp16_available(const int cc) {
235-
return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
239+
return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != GGML_CUDA_CC_DP4A) || GGML_CUDA_CC_IS_AMD(cc);
236240
}
237241

238242
// To be used for feature selection of external libraries, e.g. cuBLAS.
239243
static bool fast_fp16_hardware_available(const int cc) {
240-
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
244+
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) ||
245+
GGML_CUDA_CC_IS_AMD(cc) || GGML_CUDA_CC_IS_MTHREADS(cc);
241246
}
242247

243248
// Any FP16 tensor core instructions are available for ggml code.
@@ -246,14 +251,16 @@ static bool fp16_mma_available(const int cc) {
246251
return false;
247252
#else
248253
return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
249-
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
254+
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
255+
GGML_CUDA_CC_IS_MTHREADS(cc);
250256
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
251257
}
252258

253259
// To be used for feature selection of external libraries, e.g. cuBLAS.
254260
static bool fp16_mma_hardware_available(const int cc) {
255261
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
256-
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
262+
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
263+
GGML_CUDA_CC_IS_MTHREADS(cc);
257264
}
258265

259266
// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.

‎ggml/src/ggml-cuda/fattn-wmma-f16.cu‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9,7 +9,11 @@
99
#ifdef FP16_MMA_AVAILABLE
1010
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
1111
#include <mma.h>
12+
#ifdef GGML_USE_MUSA
13+
namespace wmma = mtmusa::wmma;
14+
#else // GGML_USE_MUSA
1215
namespace wmma = nvcuda::wmma;
16+
#endif // GGML_USE_MUSA
1317
#elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
1418
#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
1519
#include <rocwmma/rocwmma.hpp>

‎ggml/src/ggml-cuda/ggml-cuda.cu‎

Lines changed: 61 additions & 23 deletions
Original file line numberDiff line numberDiff line change
@@ -1200,7 +1200,9 @@ static void ggml_cuda_op_mul_mat_cublas(
12001200

12011201
const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
12021202

1203-
if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
1203+
if ((GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) ||
1204+
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2)) &&
1205+
src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
12041206
ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
12051207
if (src1->type != GGML_TYPE_BF16) {
12061208
const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type);
@@ -1228,7 +1230,9 @@ static void ggml_cuda_op_mul_mat_cublas(
12281230

12291231
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
12301232
to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream);
1231-
} else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
1233+
} else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
1234+
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2) ||
1235+
GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
12321236
// convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
12331237
ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
12341238
if (src0->type != GGML_TYPE_F16) {
@@ -1872,13 +1876,24 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18721876
// use cublasGemmBatchedEx
18731877
const int64_t ne23 = ne12*ne13;
18741878

1879+
#ifdef GGML_USE_MUSA
1880+
const void ** ptrs_src;
1881+
void ** ptrs_dst;
1882+
CUDA_CHECK(cudaMalloc((void **)&ptrs_src, sizeof(void *)*2*ne23));
1883+
CUDA_CHECK(cudaMalloc((void **)&ptrs_dst, sizeof(void *)*1*ne23));
1884+
#else // GGML_USE_MUSA
18751885
ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
18761886
ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23);
1887+
#endif // GGML_USE_MUSA
18771888

18781889
dim3 block_dims(ne13, ne12);
18791890
k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
18801891
src0_f16, src1_f16, dst_t,
1892+
#ifdef GGML_USE_MUSA
1893+
ptrs_src, ptrs_dst,
1894+
#else // GGML_USE_MUSA
18811895
ptrs_src.get(), ptrs_dst.get(),
1896+
#endif // GGML_USE_MUSA
18821897
ne12, ne13,
18831898
ne23,
18841899
nb02, nb03,
@@ -1888,15 +1903,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18881903
r2, r3);
18891904
CUDA_CHECK(cudaGetLastError());
18901905

1891-
CUBLAS_CHECK(
1906+
#ifdef GGML_USE_MUSA
1907+
CUDA_CHECK(cudaDeviceSynchronize());
1908+
const void **Aarray = (const void **) (ptrs_src + 0*ne23);
1909+
const void **Barray = (const void **) (ptrs_src + 1*ne23);
1910+
void **Carray = ( void **) (ptrs_dst + 0*ne23);
1911+
#else // GGML_USE_MUSA
1912+
const void **Aarray = (const void **) (ptrs_src.get() + 0*ne23);
1913+
const void **Barray = (const void **) (ptrs_src.get() + 1*ne23);
1914+
void **Carray = ( void **) (ptrs_dst.get() + 0*ne23);
1915+
#endif // GGML_USE_MUSA
1916+
1917+
CUBLAS_CHECK(
18921918
cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
18931919
ne01, ne11, ne10,
1894-
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
1895-
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, s11,
1896-
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
1920+
alpha, Aarray, CUDA_R_16F, nb01/nb00,
1921+
Barray, CUDA_R_16F, s11,
1922+
beta, Carray, cu_data_type, ne0,
18971923
ne23,
18981924
cu_compute_type,
18991925
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
1926+
1927+
#ifdef GGML_USE_MUSA
1928+
CUDA_CHECK(cudaFree(ptrs_src));
1929+
CUDA_CHECK(cudaFree(ptrs_dst));
1930+
#endif // GGML_USE_MUSA
19001931
}
19011932
#endif
19021933

@@ -1920,6 +1951,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19201951

19211952
bool any_gpus_with_slow_fp16 = false;
19221953
bool any_gpus_without_fp16_mma = false;
1954+
bool any_gpus_without_cublas_gemm = false;
19231955

19241956
if (split) {
19251957
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1930,16 +1962,18 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19301962
continue;
19311963
}
19321964

1933-
const int cc = ggml_cuda_info().devices[id].cc;
1934-
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
1935-
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
1936-
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
1965+
const int cc = ggml_cuda_info().devices[id].cc;
1966+
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
1967+
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
1968+
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
1969+
any_gpus_without_cublas_gemm = any_gpus_without_cublas_gemm || !(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
19371970
}
19381971
} else {
1939-
const int cc = ggml_cuda_info().devices[ctx.device].cc;
1940-
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
1941-
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
1942-
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
1972+
const int cc = ggml_cuda_info().devices[ctx.device].cc;
1973+
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
1974+
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
1975+
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
1976+
any_gpus_without_cublas_gemm = any_gpus_without_cublas_gemm || !(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
19431977
}
19441978

19451979
// debug helpers
@@ -1958,8 +1992,9 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19581992
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
19591993
} else if (!split && use_mul_mat_q) {
19601994
ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
1961-
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
1962-
!ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
1995+
} else if (!split && !any_gpus_without_cublas_gemm && src0->type == GGML_TYPE_F16 &&
1996+
(src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
1997+
!ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
19631998
// general KQ + KQV multi-batch without FlashAttention
19641999
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
19652000
} else if (use_mul_mat_vec) {
@@ -2999,9 +3034,17 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
29993034
return false;
30003035
}
30013036
#ifdef GGML_USE_MUSA
3002-
if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
3037+
const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
3038+
if (GGML_CUDA_CC_IS_MTHREADS(cc) && b->ne[2]*b->ne[3] > 1 &&
30033039
!ggml_is_transposed(a) && !ggml_is_transposed(b)) {
3004-
return false;
3040+
if (GGML_CUDA_CC_IS_QY1(cc) && op->op == GGML_OP_MUL_MAT &&
3041+
a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) {
3042+
return false;
3043+
}
3044+
if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID &&
3045+
a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) {
3046+
return false;
3047+
}
30053048
}
30063049
#endif // GGML_USE_MUSA
30073050
switch (a->type) {
@@ -3028,11 +3071,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
30283071
case GGML_TYPE_IQ4_NL:
30293072
case GGML_TYPE_IQ4_XS:
30303073
case GGML_TYPE_BF16:
3031-
#ifdef GGML_USE_MUSA
3032-
if (a->type == GGML_TYPE_Q3_K) {
3033-
return false;
3034-
}
3035-
#endif // GGML_USE_MUSA
30363074
return true;
30373075
default:
30383076
return false;

0 commit comments

Comments
 (0)