mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 21:21:22 +00:00
[rocm-libraries] ROCm/rocm-libraries#5939 (commit 6fb1791)
[CK_TILE] Flatten nested static_for loops into static_ford (#5939) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Summary Mechanical conversion of 129 nested `static_for`/`static_ford` patterns to flat `static_ford` across 29 ck_tile header files. Each conversion eliminates intermediate lambda closure instantiations by replacing nested compile-time loops with a single flat iteration using index decomposition. ### What `static_ford` eliminates When `static_for` loops are nested, each level creates unique closure types: ```cpp // BEFORE: M + M×N = 20 IR functions (for M=4, N=4) static_for<0, 4, 1>{}([&](auto m) { // 4 closure instantiations static_for<0, 4, 1>{}([&](auto n) { // 4×4 = 16 closure instantiations body(m, n); }); }); // AFTER: M×N = 16 IR functions (with ford_applier, no intermediates) static_ford<sequence<4, 4>>{}([&](auto mn) { constexpr auto m = number<mn[number<0>{}]>{}; constexpr auto n = number<mn[number<1>{}]>{}; body(m, n); }); ``` ### Pattern categories converted | Category | Count | Description | |----------|-------|-------------| | C (2-level `static_for` chains) | 112 | Nested `static_for` → `static_ford` | | C3 (3-level `static_for` chains) | 9 | Three consecutive nests → `static_ford` | | Partial rescue | 3 | Outer 2 levels of blocked 4-level nests | | B (nested `static_ford` merge) | 5 | Two nested `static_ford` → single higher-dim `static_ford` | | **Total** | **129** | Across 29 files | 6 false positives were detected and reverted (in `tensor_adaptor.hpp`, `tile_distribution.hpp`, `tile_distribution_encoding.hpp`) where the inner loop bound depended on the outer variable. ### Files changed by family | Family | Files | Sites | |--------|-------|-------| | Block GEMM | 12 | ~20 | | FlatMM pipelines | 4 | ~69 (including 5 ford-ford merges) | | GEMM quant | 7 | ~22 | | FlatMM kernel | 1 | 2 | | FMHA | 1 | 2 | | Reduce/norm | 2 | 2 | | Epilogue | 1 | 1 | ### Blocked locations from review comments - **block_gemm_areg_breg_creg_v1.hpp:356** — BLOCKED: runtime scale loads (`scale_a_slice`, `scale_b_slice`, A warp tensor load) between every nesting level - **block_universal_gemm_ar_aquant_flatbr_bquant_cr.hpp:228** — BLOCKED: `zero_accumulators()` before inner loop; `sched_barrier` + conditional `block_sync_lds()` after inner loop - **block_universal_gemm_as_aquant_bs_bquant_cr.hpp:298** — BLOCKED: runtime `CWarpTensor` construction before inner loop; quantization scale application code after inner loop - **block_universal_gemm_as_aquant_bs_cr.hpp:277** — BLOCKED: same pattern as above - **block_universal_gemm_as_bs_bquant_cr.hpp:367** — BLOCKED: same pattern as above ## Depends on - #5938 ([CK_TILE] Optimize static_ford and sequence compile-time infrastructure) — provides the `ford_applier` that makes these conversions beneficial. Without it, `static_ford` uses a recursive implementation that provides no IR function savings. ## Results (combined with #5938) ### Build Time (Wilcoxon signed-rank, 7 paired trials, gfx942) | Target | Base (s) | Treat (s) | Delta | % | Significant? | |--------|----------|-----------|-------|---|-------------| | **flatmm** | 161.1 | 149.0 | **-12.1s** | **-7.5%** | **YES** (p<0.01, 7/7 wins) | | **universal_gemm** | 225.4 | 220.3 | **-5.1s** | **-2.3%** | **YES** (p<0.01, 7/7 wins) | ### IR Function Counts (device trace, gfx942) | Target | InstFunc | CodeGen | |--------|----------|---------| | universal_gemm | **-8.5%** | **-9.2%** | | flatmm | **-7.6%** | **-10.5%** | ### ASM Equivalence 5/5 PASS — 650,151 lines verified identical (gfx942). TUs: universal_gemm, flatmm_basic, fmha_bwd, reduce, bscale. ## Test plan - [x] ASM equivalence verified (650K lines, gfx942) - [x] Wilcoxon timing verified (7 trials, p<0.01) - [x] IR function counts verified (-7.6% to -10.5% CodeGen reduction) - [ ] CI 🤖 Generated with [Claude Code](https://claude.com/claude-code)
This commit is contained in:
Christopher Millette
committed by
assistant-librarian[bot]
assistant-librarian[bot]
parent
c2ac7aa7b0
commit
a170e2bd9d
@@ -213,38 +213,38 @@ struct BlockGemmARegBRegCRegV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
using c_iter_idx = std::
|
||||
conditional_t<TransposeC, sequence<nIter, mIter>, sequence<mIter, nIter>>;
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
// read C warp tensor from C block tensor
|
||||
using c_iter_idx =
|
||||
std::conditional_t<TransposeC, sequence<nIter, mIter>, sequence<mIter, nIter>>;
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
@@ -323,73 +323,69 @@ struct BlockGemmARegBRegCRegV1
|
||||
|
||||
// hot loop with MX scaling and pre-packed int32_t scales:
|
||||
// Outer loops iterate over pack groups (scale tile indices)
|
||||
static_for<0, KPackIterPerWarp, 1>{}([&](auto ikpack) {
|
||||
static_for<0, MPackIterPerWarp, 1>{}([&](auto impack) {
|
||||
// Get pre-packed int32_t A scale (already contains MXdlPack*KXdlPack e8m0_t)
|
||||
auto scale_a_slice = scale_a_tensor.get_y_sliced_thread_data(
|
||||
sequence<ikpack, impack, 0>{}, sequence<1, 1, 1>{});
|
||||
const int32_t a_scale_packed = bit_cast<int32_t>(scale_a_slice[number<0>{}]);
|
||||
static_ford<sequence<KPackIterPerWarp, MPackIterPerWarp>>{}([&](auto ii) {
|
||||
constexpr auto ikpack = number<ii[number<0>{}]>{};
|
||||
constexpr auto impack = number<ii[number<1>{}]>{};
|
||||
// Get pre-packed int32_t A scale (already contains MXdlPack*KXdlPack e8m0_t)
|
||||
auto scale_a_slice = scale_a_tensor.get_y_sliced_thread_data(
|
||||
sequence<ikpack, impack, 0>{}, sequence<1, 1, 1>{});
|
||||
const int32_t a_scale_packed = bit_cast<int32_t>(scale_a_slice[number<0>{}]);
|
||||
|
||||
static_for<0, NPackIterPerWarp, 1>{}([&](auto inpack) {
|
||||
// Get pre-packed int32_t B scale
|
||||
auto scale_b_slice = scale_b_tensor.get_y_sliced_thread_data(
|
||||
sequence<ikpack, inpack, 0>{}, sequence<1, 1, 1>{});
|
||||
const int32_t b_scale_packed = bit_cast<int32_t>(scale_b_slice[number<0>{}]);
|
||||
static_for<0, NPackIterPerWarp, 1>{}([&](auto inpack) {
|
||||
// Get pre-packed int32_t B scale
|
||||
auto scale_b_slice = scale_b_tensor.get_y_sliced_thread_data(
|
||||
sequence<ikpack, inpack, 0>{}, sequence<1, 1, 1>{});
|
||||
const int32_t b_scale_packed = bit_cast<int32_t>(scale_b_slice[number<0>{}]);
|
||||
|
||||
// Inner loops: issue MFMAs within the pack group using OpSel
|
||||
static_for<0, KXdlPack, 1>{}([&](auto ikxdl) {
|
||||
static_for<0, MXdlPack, 1>{}([&](auto imxdl) {
|
||||
constexpr auto kIter = ikpack * KXdlPack + ikxdl;
|
||||
constexpr auto mIter = impack * MXdlPack + imxdl;
|
||||
// Inner loops: issue MFMAs within the pack group using OpSel
|
||||
static_ford<sequence<KXdlPack, MXdlPack>>{}([&](auto jj) {
|
||||
constexpr auto ikxdl = number<jj[number<0>{}]>{};
|
||||
constexpr auto imxdl = number<jj[number<1>{}]>{};
|
||||
constexpr auto kIter = ikpack * KXdlPack + ikxdl;
|
||||
constexpr auto mIter = impack * MXdlPack + imxdl;
|
||||
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() =
|
||||
a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
// OpSel for A: selects byte within packed int32_t
|
||||
constexpr index_t kOpSelA = ikxdl * MXdlPack + imxdl;
|
||||
// OpSel for A: selects byte within packed int32_t
|
||||
constexpr index_t kOpSelA = ikxdl * MXdlPack + imxdl;
|
||||
|
||||
static_for<0, NXdlPack, 1>{}([&](auto inxdl) {
|
||||
constexpr auto nIter = inpack * NXdlPack + inxdl;
|
||||
static_for<0, NXdlPack, 1>{}([&](auto inxdl) {
|
||||
constexpr auto nIter = inpack * NXdlPack + inxdl;
|
||||
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() =
|
||||
b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{},
|
||||
b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
// OpSel for B: selects byte within packed int32_t
|
||||
constexpr index_t kOpSelB = ikxdl * NXdlPack + inxdl;
|
||||
// OpSel for B: selects byte within packed int32_t
|
||||
constexpr index_t kOpSelB = ikxdl * NXdlPack + inxdl;
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
using c_iter_idx = std::conditional_t<TransposeC,
|
||||
sequence<nIter, mIter>,
|
||||
sequence<mIter, nIter>>;
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() =
|
||||
c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
// read C warp tensor from C block tensor
|
||||
using c_iter_idx = std::conditional_t<TransposeC,
|
||||
sequence<nIter, mIter>,
|
||||
sequence<mIter, nIter>>;
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM with MX scaling using pre-packed scale and OpSel
|
||||
WarpGemm{}.template operator()<kOpSelA, kOpSelB>(c_warp_tensor,
|
||||
a_warp_tensor,
|
||||
b_warp_tensor,
|
||||
a_scale_packed,
|
||||
b_scale_packed);
|
||||
// warp GEMM with MX scaling using pre-packed scale and OpSel
|
||||
WarpGemm{}.template operator()<kOpSelA, kOpSelB>(c_warp_tensor,
|
||||
a_warp_tensor,
|
||||
b_warp_tensor,
|
||||
a_scale_packed,
|
||||
b_scale_packed);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
@@ -250,74 +250,74 @@ struct BlockGemmARegBRegCRegV2
|
||||
// hot loop:
|
||||
if constexpr(BlockGemmLoopOrder == GemmLoopOrder::KMN)
|
||||
{
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<kIter, mIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<kIter, mIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<kIter, nIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<kIter, nIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
CWarpTensor c_warp_tensor;
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
else if constexpr(BlockGemmLoopOrder == GemmLoopOrder::MNK)
|
||||
{
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_ford<sequence<MIterPerWarp, NIterPerWarp, KIterPerWarp>>{}([&](auto mnk) {
|
||||
constexpr auto mIter = number<mnk[number<0>{}]>{};
|
||||
constexpr auto nIter = number<mnk[number<1>{}]>{};
|
||||
constexpr auto kIter = number<mnk[number<2>{}]>{};
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
// read A warp tensor from A Block window
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
@@ -109,13 +109,13 @@ struct BlockGemmARegBSmemCRegOneWarpV1
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
static_ford<sequence<NIterPerWarp, KIterPerWarp>>{}([&](auto nk) {
|
||||
constexpr auto nIter = number<nk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<nk[number<1>{}]>{};
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -141,35 +141,35 @@ struct BlockGemmARegBSmemCRegOneWarpV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -116,13 +116,13 @@ struct BlockGemmARegBSmemCRegV1
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
static_ford<sequence<NIterPerWarp, KIterPerWarp>>{}([&](auto nk) {
|
||||
constexpr auto nIter = number<nk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<nk[number<1>{}]>{};
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -148,35 +148,35 @@ struct BlockGemmARegBSmemCRegV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -103,13 +103,13 @@ struct BlockGemmARegBSmemCRegV2
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
static_ford<sequence<NIterPerWarp, KIterPerWarp>>{}([&](auto nk) {
|
||||
constexpr auto nIter = number<nk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<nk[number<1>{}]>{};
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -135,36 +135,36 @@ struct BlockGemmARegBSmemCRegV2
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
static_ford<sequence<KIterPerWarp, NIterPerWarp>>{}([&](auto kn) {
|
||||
constexpr auto kIter = number<kn[number<0>{}]>{};
|
||||
constexpr auto nIter = number<kn[number<1>{}]>{};
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -90,13 +90,13 @@ struct BlockGemmARegBSmemCRegV2R1
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
static_ford<sequence<NIterPerWarp, KIterPerWarp>>{}([&](auto nk) {
|
||||
constexpr auto nIter = number<nk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<nk[number<1>{}]>{};
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
|
||||
// check C-block-distribution
|
||||
@@ -126,43 +126,43 @@ struct BlockGemmARegBSmemCRegV2R1
|
||||
NIterPerWarp>
|
||||
b_warp_tensors;
|
||||
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
b_warp_tensors(nIter)(kIter) = load_tile(b_warp_windows(nIter)(kIter));
|
||||
});
|
||||
static_ford<sequence<KIterPerWarp, NIterPerWarp>>{}([&](auto kn) {
|
||||
constexpr auto kIter = number<kn[number<0>{}]>{};
|
||||
constexpr auto nIter = number<kn[number<1>{}]>{};
|
||||
b_warp_tensors(nIter)(kIter) = load_tile(b_warp_windows(nIter)(kIter));
|
||||
});
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = b_warp_tensors(nIter)(kIter);
|
||||
static_ford<sequence<KIterPerWarp, NIterPerWarp>>{}([&](auto kn) {
|
||||
constexpr auto kIter = number<kn[number<0>{}]>{};
|
||||
constexpr auto nIter = number<kn[number<1>{}]>{};
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = b_warp_tensors(nIter)(kIter);
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
|
||||
|
||||
@@ -116,13 +116,13 @@ struct BlockGemmASmemBRegCRegV1
|
||||
MIterPerWarp>
|
||||
a_warp_windows;
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
|
||||
static_ford<sequence<MIterPerWarp, KIterPerWarp>>{}([&](auto mk) {
|
||||
constexpr auto mIter = number<mk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<mk[number<1>{}]>{};
|
||||
a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
|
||||
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -148,34 +148,34 @@ struct BlockGemmASmemBRegCRegV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A Block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A Block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -85,13 +85,13 @@ struct BlockGemmASmemBSmemCRegV1
|
||||
MIterPerWarp>
|
||||
a_warp_windows;
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
|
||||
static_ford<sequence<MIterPerWarp, KIterPerWarp>>{}([&](auto mk) {
|
||||
constexpr auto mIter = number<mk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<mk[number<1>{}]>{};
|
||||
a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
|
||||
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -120,13 +120,13 @@ struct BlockGemmASmemBSmemCRegV1
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
static_ford<sequence<NIterPerWarp, KIterPerWarp>>{}([&](auto nk) {
|
||||
constexpr auto nIter = number<nk[number<0>{}]>{};
|
||||
constexpr auto kIter = number<nk[number<1>{}]>{};
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
#endif
|
||||
|
||||
@@ -138,31 +138,31 @@ struct BlockGemmASmemBSmemCRegV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -165,61 +165,60 @@ struct BlockGemmMxARegBSmemCRegV1
|
||||
uniform_sequence_gen_t<BScaleWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
auto b_warp_window = b_warp_window_tmp;
|
||||
move_tile_window(
|
||||
b_warp_window,
|
||||
{nIter * (NPerBlock / NIterPerWarp), kIter * (KPerBlock / KIterPerWarp)});
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_window);
|
||||
static_ford<sequence<KIterPerWarp, NIterPerWarp>>{}([&](auto kn) {
|
||||
constexpr auto kIter = number<kn[number<0>{}]>{};
|
||||
constexpr auto nIter = number<kn[number<1>{}]>{};
|
||||
auto b_warp_window = b_warp_window_tmp;
|
||||
move_tile_window(
|
||||
b_warp_window,
|
||||
{nIter * (NPerBlock / NIterPerWarp), kIter * (KPerBlock / KIterPerWarp)});
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_window);
|
||||
|
||||
BScaleWarpTensor b_scale_warp_tensor;
|
||||
BScaleWarpTensor b_scale_warp_tensor;
|
||||
|
||||
b_scale_warp_tensor.get_thread_buffer() =
|
||||
b_scale_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter / NIterPack, nIter % NIterPack, kIter>{},
|
||||
b_scale_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, b_scale_warp_y_lengths));
|
||||
b_scale_warp_tensor.get_thread_buffer() = b_scale_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter / NIterPack, nIter % NIterPack, kIter>{},
|
||||
b_scale_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, b_scale_warp_y_lengths));
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
AScaleWarpTensor a_scale_warp_tensor;
|
||||
AScaleWarpTensor a_scale_warp_tensor;
|
||||
|
||||
a_scale_warp_tensor.get_thread_buffer() =
|
||||
a_scale_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_scale_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_scale_warp_y_lengths));
|
||||
a_scale_warp_tensor.get_thread_buffer() =
|
||||
a_scale_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_scale_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_scale_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
|
||||
c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
|
||||
c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}.template operator()<0, 0>(
|
||||
c_warp_tensor,
|
||||
a_warp_tensor,
|
||||
b_warp_tensor,
|
||||
int32_t(a_scale_warp_tensor.get_thread_buffer()[0]),
|
||||
int32_t(b_scale_warp_tensor.get_thread_buffer()[0]));
|
||||
// warp GEMM
|
||||
WarpGemm{}.template operator()<0, 0>(
|
||||
c_warp_tensor,
|
||||
a_warp_tensor,
|
||||
b_warp_tensor,
|
||||
int32_t(a_scale_warp_tensor.get_thread_buffer()[0]),
|
||||
int32_t(b_scale_warp_tensor.get_thread_buffer()[0]));
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
|
||||
c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
|
||||
c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
@@ -239,39 +239,39 @@ struct BlockUniversalGemmAsBsCr
|
||||
"C block tensor data type!");
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
|
||||
b_warp_tensor.get_thread_buffer() = b_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
b_warp_tensor.get_thread_buffer() = b_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
@@ -392,63 +392,59 @@ struct BlockUniversalGemmAsBsCr
|
||||
0); // Prevents instruction reordering across this boundary
|
||||
}
|
||||
|
||||
static_for<0, KInnerLoopIter, 1>{}([&](auto kInnerIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
static_ford<sequence<KInnerLoopIter, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kInnerIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kInnerIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kInnerIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
|
||||
b_warp_tensor.get_thread_buffer() =
|
||||
b_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kInnerIter>{},
|
||||
b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
// read C warp tensor from C block tensor-
|
||||
CWarpTensor c_warp_tensor;
|
||||
b_warp_tensor.get_thread_buffer() = b_warp_tile_.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kInnerIter>{}, b_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
|
||||
// read C warp tensor from C block tensor-
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() =
|
||||
c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// The block_sync_lds() here performs double duty:
|
||||
// A) safeguard against data hazard because barrier from
|
||||
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion
|
||||
// by applying small delays to different wavefronts It is
|
||||
// performed near the end of MAC cluster to minimize lgkmcnt
|
||||
// penalty
|
||||
if constexpr(kIter.value == KRepeat - 1 &&
|
||||
kInnerIter.value == KInnerLoopIter - 1 &&
|
||||
mIter.value == MIterPerWarp - 1 &&
|
||||
nIter.value == NIterPerWarp - 1)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
block_sync_lds();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// The block_sync_lds() here performs double duty:
|
||||
// A) safeguard against data hazard because barrier from
|
||||
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion
|
||||
// by applying small delays to different wavefronts It is
|
||||
// performed near the end of MAC cluster to minimize lgkmcnt
|
||||
// penalty
|
||||
if constexpr(kIter.value == KRepeat - 1 &&
|
||||
kInnerIter.value == KInnerLoopIter - 1 &&
|
||||
mIter.value == MIterPerWarp - 1 &&
|
||||
nIter.value == NIterPerWarp - 1)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
block_sync_lds();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
// warp GEMM
|
||||
WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
|
||||
if constexpr(kInnerIter.value == 0 && mIter.value == 0 &&
|
||||
nIter.value == 0)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(1);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
});
|
||||
if constexpr(kInnerIter.value == 0 && mIter.value == 0 && nIter.value == 0)
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
__builtin_amdgcn_s_setprio(1);
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
});
|
||||
});
|
||||
|
||||
|
||||
@@ -156,55 +156,54 @@ struct BlockWeightPreshuffleASmemBRegCReg
|
||||
uniform_sequence_gen_t<BFlatDistribution::NDimY, 0>{};
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
constexpr auto AwarpIter = (kIter * MIterPerWarp + mIter) % m_preload;
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read C warp tensor from C block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
CWarpTensor c_warp_tensor;
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
constexpr auto AwarpIter = (kIter * MIterPerWarp + mIter) % m_preload;
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read C warp tensor from C block tensor
|
||||
BWarpTensor b_warp_tensor;
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<nIter, kIter>{},
|
||||
typename sequence_split<decltype(b_block_y_index_zeros),
|
||||
2>::right_type{}),
|
||||
merge_sequences(
|
||||
sequence<1, 1>{},
|
||||
typename sequence_split<decltype(b_block_y_lengths), 2>::right_type{}));
|
||||
b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(
|
||||
sequence<nIter, kIter>{},
|
||||
typename sequence_split<decltype(b_block_y_index_zeros), 2>::right_type{}),
|
||||
merge_sequences(
|
||||
sequence<1, 1>{},
|
||||
typename sequence_split<decltype(b_block_y_lengths), 2>::right_type{}));
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WarpGemm{}(
|
||||
c_warp_tensor, preloaded_a_warp_tensor(number<AwarpIter>{}), b_warp_tensor);
|
||||
// warp GEMM
|
||||
WarpGemm{}(
|
||||
c_warp_tensor, preloaded_a_warp_tensor(number<AwarpIter>{}), b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
|
||||
__builtin_amdgcn_sched_barrier(0x7F6);
|
||||
});
|
||||
// preload next A from lds
|
||||
if constexpr((kIter * MIterPerWarp + mIter) <
|
||||
(KIterPerWarp * MIterPerWarp - m_preload))
|
||||
{
|
||||
constexpr auto AmIter = (mIter + m_preload) % MIterPerWarp;
|
||||
constexpr auto AkIter = (kIter + (mIter + m_preload) / MIterPerWarp);
|
||||
|
||||
load_tile(preloaded_a_warp_tensor(number<AwarpIter>{}),
|
||||
a_load_windows[number<AkIter>{}][number<AmIter>{}]);
|
||||
}
|
||||
|
||||
// barrier
|
||||
if constexpr((kIter == KIterPerWarp - 1) && (mIter == MIter_2nd_last))
|
||||
{
|
||||
block_sync_lds();
|
||||
}
|
||||
__builtin_amdgcn_sched_barrier(0x7F6);
|
||||
});
|
||||
// preload next A from lds
|
||||
if constexpr((kIter * MIterPerWarp + mIter) < (KIterPerWarp * MIterPerWarp - m_preload))
|
||||
{
|
||||
constexpr auto AmIter = (mIter + m_preload) % MIterPerWarp;
|
||||
constexpr auto AkIter = (kIter + (mIter + m_preload) / MIterPerWarp);
|
||||
|
||||
load_tile(preloaded_a_warp_tensor(number<AwarpIter>{}),
|
||||
a_load_windows[number<AkIter>{}][number<AmIter>{}]);
|
||||
}
|
||||
|
||||
// barrier
|
||||
if constexpr((kIter == KIterPerWarp - 1) && (mIter == MIter_2nd_last))
|
||||
{
|
||||
block_sync_lds();
|
||||
}
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
@@ -88,28 +88,28 @@ struct BlockWeightPreshuffleASmemBSmemCRegV1
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
static_ford<sequence<KIterPerWarp, MIterPerWarp>>{}([&](auto km) {
|
||||
constexpr auto kIter = number<km[number<0>{}]>{};
|
||||
constexpr auto mIter = number<km[number<1>{}]>{};
|
||||
// read A warp tensor from A block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor(nIter)(kIter));
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor(nIter)(kIter));
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user