leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 26 Packages Projects Releases Wiki Activity

Release v4.0.0 (#2294)

Browse Source
This commit is contained in:
Kihiro Bando
2025-05-13 15:55:29 -04:00
committed by GitHub
parent ad7b2f5e84
commit f115c3f854
299 changed files with 51495 additions and 4413 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cute/arch/config.hpp
+1
View File
@@ -93,6 +93,7 @@
# define CUTE_ARCH_STSM_SM90_ENABLED
#endif
#if (defined(CUTLASS_ARCH_MMA_SM100A_ENABLED) || defined(CUTLASS_ARCH_MMA_SM101A_ENABLED))
# define CUTE_ARCH_TCGEN05_S8_MMA_ENABLED
#endif
include/cute/arch/mma_sm100_umma.hpp
+32
View File
@@ -1394,7 +1394,11 @@ struct SM100_MMA_MXF4_SS
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.cta_group::1.kind::mxf4nvf4.block_scale.block16 [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.cta_group::1.kind::mxf4nvf4.block_scale.scale_vec::4X [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1411,7 +1415,11 @@ struct SM100_MMA_MXF4_SS
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.cta_group::1.kind::mxf4.block_scale.block32 [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.cta_group::1.kind::mxf4.block_scale.scale_vec::2X [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1457,7 +1465,11 @@ struct SM100_MMA_MXF4NVF4_SS_SPARSE
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.sp.cta_group::1.kind::mxf4nvf4.block_scale.block16 [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.sp.cta_group::1.kind::mxf4nvf4.block_scale.scale_vec::4X [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1475,7 +1487,11 @@ struct SM100_MMA_MXF4NVF4_SS_SPARSE
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.sp.cta_group::1.kind::mxf4.block_scale.block32 [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.sp.cta_group::1.kind::mxf4.block_scale.scale_vec::2X [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1520,7 +1536,11 @@ struct SM100_MMA_MXF4_2x1SM_SS
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.cta_group::2.kind::mxf4nvf4.block_scale.block16 [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.cta_group::2.kind::mxf4nvf4.block_scale.scale_vec::4X [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1537,7 +1557,11 @@ struct SM100_MMA_MXF4_2x1SM_SS
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.cta_group::2.kind::mxf4.block_scale.block32 [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.cta_group::2.kind::mxf4.block_scale.scale_vec::2X [%0], %1, %2, %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1582,7 +1606,11 @@ struct SM100_MMA_MXF4NVF4_2x1SM_SS_SPARSE
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.sp.cta_group::2.kind::mxf4nvf4.block_scale.block16 [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.sp.cta_group::2.kind::mxf4nvf4.block_scale.scale_vec::4X [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
@@ -1600,7 +1628,11 @@ struct SM100_MMA_MXF4NVF4_2x1SM_SS_SPARSE
"{\n\t"
".reg .pred p;\n\t"
"setp.ne.b32 p, %4, 0;\n\t"
#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
"tcgen05.mma.sp.cta_group::2.kind::mxf4.block_scale.block32 [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#else
"tcgen05.mma.sp.cta_group::2.kind::mxf4.block_scale.scale_vec::2X [%0], %1, %2, [%7], %3, [%5], [%6], p; \n\t"
#endif
"}\n"
:
: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(uint32_t(idescE>>32)), "r"(scaleC),
include/cute/atom/copy_traits_sm100_tma.hpp
+2 -2
View File
@@ -104,7 +104,7 @@ struct Copy_Traits<SM100_TMA_2SM_LOAD, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM100_TMA_2SM_LOAD before calling .with()
@@ -192,7 +192,7 @@ struct Copy_Traits<SM100_TMA_2SM_LOAD_MULTICAST, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM100_TMA_2SM_LOAD_MULTICAST_OP before calling .with()
include/cute/atom/copy_traits_sm90_tma.hpp
+4 -4
View File
@@ -146,7 +146,7 @@ struct Copy_Traits<SM90_TMA_LOAD, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM90_TMA_LOAD before calling .with()
@@ -276,7 +276,7 @@ struct Copy_Traits<SM90_TMA_LOAD_MULTICAST, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM90_TMA_LOAD_MULTICAST before calling .with()
@@ -350,7 +350,7 @@ struct Copy_Traits<SM90_TMA_STORE, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Construct new TMA_STORE with (unsafe) swapped out TMA descriptor ptr (for grouped gemm/ptr array gemm)
@@ -463,7 +463,7 @@ struct Copy_Traits<SM90_TMA_REDUCE_ADD, NumBitsPerTMA, AuxParams_>
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
return make_coord_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
template <class Coord, int... Is>
include/cute/tensor_impl.hpp
+3 -3
View File
@@ -474,14 +474,14 @@ make_fragment_like(Tensor<Engine,Layout> const& tensor)
}
//
// make_counting_tensor
// make_coord_tensor
// Make a tensor from a layout by binding it to a counting iter with 0-offset of the same profile as the codomain.
//
template <class Layout, __CUTE_REQUIRES(is_layout<Layout>::value)>
CUTE_HOST_DEVICE constexpr
auto
make_counting_tensor(Layout const& layout)
make_coord_tensor(Layout const& layout)
{
return make_tensor(make_inttuple_iter(coprofile(layout)), layout);
}
@@ -496,7 +496,7 @@ CUTE_HOST_DEVICE constexpr
auto
make_identity_tensor(Shape const& shape)
{
return make_counting_tensor(make_identity_layout(shape));
return make_coord_tensor(make_identity_layout(shape));
}
//