2. Introduction#
rocWMMA is AMD’s C++ library for accelerating mixed precision matrix multiply-accumulate operations leveraging specialized GPU matrix cores on AMD’s latest discrete GPUs.
A C++ API is provided to facilitate decomposition of matrix multiply-accumulate problems into discretized block fragments and to parallelize block-wise operations across multiple GPU wavefronts.
The API is implemented in GPU device code: it empowers user device kernel code with direct use of GPU matrix cores. Moreover, this code can benefit from inline compiler optimization passes and does not incur additional overhead of external runtime calls or extra kernel launches.
Acronym |
Expansion |
|---|---|
GEMM |
GEneral Matrix Multiply |
WMMA |
Wavefront Mixed precision Multiply Accumulate |
ROCm |
Radeon Open Compute platform |
HIP |
Heterogeneous-Compute Interface for Portability |
rocWMMA is written in C++14 and may be applied directly in device kernel code. Library code is templated for modularity and uses available meta-data to provide opportunities for compile-time inferences and optimizations.
The rocWMMA API exposes block-wise data load / store and matrix multiply-accumulate functions appropriately sized for thread-block execution on data fragments. Matrix multiply-accumulate functionality supports mixed precision inputs and outputs with native fixed-precision accumulation. The rocWMMA Coop API provides wave/warp collaborations within the thread-blocks for block-wise data load and stores. Supporting code is required for GPU device management and for kernel invocation. Kernel code samples and tests provided are built and launched via the HIP ecosystem within ROCm.
Below is a simple example code for calling rocWMMA functions load_matrix_sync, store_matrix_sync, fill_fragment, mma_sync.
#include <hip/hip_ext.h>
#include <hip/hip_fp16.h>
#include <hip/hip_runtime.h>
#include <iostream>
#include <vector>
#include <rocwmma/rocwmma.hpp>
using rocwmma::float16_t;
using rocwmma::float32_t;
// Matrix data initialization
template <typename DataT>
__host__ static inline void fill(DataT* mat, uint32_t m, uint32_t n)
{
auto ld = n;
for(int i = 0; i < m; ++i)
{
for(int j = 0; j < n; ++j)
{
// Generated data
// Alternate sign every 3 elements
auto value = (i * n + j) % 13;
mat[i * ld + j] = (value % 3) ? -static_cast<DataT>(value) : static_cast<DataT>(value);
}
}
}
// Supports BlockM/N square sizes of
// : 16 x 16
// : 32 x 32
const int ROCWMMA_M = 16;
const int ROCWMMA_N = 16;
// Supports ROCWMMA_K sizes as
// : multiples of 16.
const int ROCWMMA_K = 16;
// AMDGCN default wave size
const int WAVE_SIZE = rocwmma::AMDGCN_WAVE_SIZE;
// Thread block
// : T_BLOCK_X must be multiple of WAVE_SIZE.
// Note: Each wave will compute one BLOCK_M x BLOCK_N output block
// Note: Workgroup will compute
// T_BLOCK_X / WAVE_SIZE x T_BLOCK_Y output blocks
// This thread block will compute (4 x 4 output blocks)
const int T_BLOCK_X = 4 * WAVE_SIZE;
const int T_BLOCK_Y = 4;
// The following device kernel is a naive implementation
// of blocked GEMM. Each wave will compute one BLOCK_M x BLOCK_N
// output block of the M x N x K GEMM, generalized as:
// D = alpha * (A x B) + beta * C
//
// In this simplified example, we assume:
// : A is in row-major format (m x k)
// : B is in col-major format (k x n)
// : C, D are in row-major format (m x n)
// : Multiplication is NOT in-place, output is written to D matrix
// : No LDS required
//
// Disclaimer: This is a simplified implementation to demonstrate API usage in
// context of wave-level GEMM computation, and is not optimized.
//
// Launchable device kernel function:
//
__global__ void gemm_wmma_d(uint32_t m, // matrix free dim m
uint32_t n, // matrix free dim n
uint32_t k, // matrix fixed dim k
float16_t const* a, // device data ptr for matrix A
float16_t const* b, // device data ptr for matrix B
float32_t const* c, // device data ptr for matrix C
float32_t* d, // device data ptr for matrix D
uint32_t lda, // leading dimension for matrix A
uint32_t ldb, // leading dimension for matrix B
uint32_t ldc, // leading dimension for matrix C
uint32_t ldd, // leading dimension for matrix D
float32_t alpha, // uniform scalar
float32_t beta) // uniform scalar
{
// Create frags with meta-data context for block-wise GEMM decomposition
// @tp0: fragment context = matrix_a, matrix_b or accumulator
// @tp1: block size M
// @tp2: block size N
// @tp3: block size K
// @tp4: fragment data type
// @tp5: data layout = row_major, col_major or void (default)
auto fragA = rocwmma::fragment<rocwmma::matrix_a, ROCWMMA_M, ROCWMMA_N, ROCWMMA_K, float16_t, rocwmma::row_major>();
auto fragB = rocwmma::fragment<rocwmma::matrix_b, ROCWMMA_M, ROCWMMA_N, ROCWMMA_K, float16_t, rocwmma::col_major>();
auto fragC = rocwmma::fragment<rocwmma::accumulator, ROCWMMA_M, ROCWMMA_N, ROCWMMA_K, float32_t>();
auto fragAcc = rocwmma::fragment<rocwmma::accumulator, ROCWMMA_M, ROCWMMA_N, ROCWMMA_K, float32_t>();
// Initialize accumulator fragment
rocwmma::fill_fragment(fragAcc, 0.0f);
// Tile using a 2D grid
auto majorWarp = (blockIdx.x * blockDim.x + threadIdx.x) / WAVE_SIZE;
auto minorWarp = (blockIdx.y * blockDim.y + threadIdx.y);
// Target C block
auto cRow = majorWarp * ROCWMMA_M;
auto cCol = minorWarp * ROCWMMA_N;
// Bounds check
if(cRow < m && cCol < n)
{
// fragAcc = A x B
for(int i = 0; i < k; i += ROCWMMA_K)
{
// Load the inputs
rocwmma::load_matrix_sync(fragA, a + (cRow * lda + i), lda);
rocwmma::load_matrix_sync(fragB, b + (i + cCol * ldb), ldb);
// Matrix multiply - accumulate using MFMA units
rocwmma::mma_sync(fragAcc, fragA, fragB, fragAcc);
}
// Fetch C matrix
rocwmma::load_matrix_sync(fragC, c + (cRow * ldc + cCol), ldc, rocwmma::mem_row_major);
// D = alpha * A x B + beta * C
for(int i = 0; i < fragC.num_elements; ++i)
{
fragC.x[i] = alpha * fragAcc.x[i] + beta * fragC.x[i];
}
// Store to D
rocwmma::store_matrix_sync(d + (cRow * ldd + cCol), fragC, ldd, rocwmma::mem_row_major);
}
}
// Host side supporting device mgmt and launch code
__host__ void gemm_test(uint32_t m, uint32_t n, uint32_t k, float32_t alpha, float32_t beta)
{
// Problem size check
if((m < (ROCWMMA_M * T_BLOCK_X / WAVE_SIZE) || n < (ROCWMMA_N * T_BLOCK_Y) || k < ROCWMMA_K)
|| (m % ROCWMMA_M || n % ROCWMMA_N || k % ROCWMMA_K))
{
std::cout << "Unsupported size!\n";
return;
}
int lda = k;
int ldb = k;
int ldc = n;
int ldd = ldc;
std::cout << "Initializing host data..." << std::endl;
// Initialize input matrices
std::vector<float16_t> matrixA(m * k);
std::vector<float16_t> matrixB(k * n);
std::vector<float32_t> matrixC(m * n);
// Fill outputs with NaN to catch contamination
std::vector<float32_t> matrixD(m * n, std::numeric_limits<float32_t>::signaling_NaN());
fill(matrixA.data(), m, k);
fill(matrixB.data(), k, n);
fill(matrixC.data(), m, n);
std::cout << "Initializing device data..." << std::endl;
// Allocate and copy device memory
float16_t* d_a;
float16_t* d_b;
float32_t* d_c;
float32_t* d_d;
const size_t bytesA = matrixA.size() * sizeof(float16_t);
const size_t bytesB = matrixB.size() * sizeof(float16_t);
const size_t bytesC = matrixC.size() * sizeof(float32_t);
const size_t bytesD = matrixD.size() * sizeof(float32_t);
CHECK_HIP_ERROR(hipMalloc(&d_a, bytesA));
CHECK_HIP_ERROR(hipMalloc(&d_b, bytesB));
CHECK_HIP_ERROR(hipMalloc(&d_c, bytesC));
CHECK_HIP_ERROR(hipMalloc(&d_d, bytesD));
CHECK_HIP_ERROR(hipMemcpy(d_a, matrixA.data(), bytesA, hipMemcpyHostToDevice));
CHECK_HIP_ERROR(hipMemcpy(d_b, matrixB.data(), bytesB, hipMemcpyHostToDevice));
CHECK_HIP_ERROR(hipMemcpy(d_c, matrixC.data(), bytesC, hipMemcpyHostToDevice));
CHECK_HIP_ERROR(hipMemcpy(d_d, matrixD.data(), bytesD, hipMemcpyHostToDevice));
auto blockDim = dim3(T_BLOCK_X, T_BLOCK_Y);
auto gridDim = dim3(rocwmma::ceilDiv(m, ROCWMMA_M * T_BLOCK_X / WAVE_SIZE),
rocwmma::ceilDiv(n, ROCWMMA_N * T_BLOCK_Y));
std::cout << "Launching GEMM kernel..." << std::endl;
hipEvent_t startEvent, stopEvent;
CHECK_HIP_ERROR(hipEventCreate(&startEvent));
CHECK_HIP_ERROR(hipEventCreate(&stopEvent));
hipExtLaunchKernelGGL(gemm_wmma_d,
gridDim,
blockDim,
0, // sharedMemBytes
0, // stream
startEvent, // Event start
stopEvent, // event stop
0, // flags
m,
n,
k,
d_a,
d_b,
d_c,
d_d,
lda,
ldb,
ldc,
ldd,
alpha,
beta);
auto elapsedTimeMs = 0.0f;
CHECK_HIP_ERROR(hipEventSynchronize(stopEvent));
CHECK_HIP_ERROR(hipEventElapsedTime(&elapsedTimeMs, startEvent, stopEvent));
CHECK_HIP_ERROR(hipEventDestroy(startEvent));
CHECK_HIP_ERROR(hipEventDestroy(stopEvent));
// Release device memory
CHECK_HIP_ERROR(hipFree(d_a));
CHECK_HIP_ERROR(hipFree(d_b));
CHECK_HIP_ERROR(hipFree(d_c));
CHECK_HIP_ERROR(hipFree(d_d));
std::cout << "Finished!" << std::endl;
}
int main()
{
gemm_test(256, 256, 256, 2.1f, 2.1f);
return 0;
}
2.1. Synchronous API#
In general, rocWMMA API functions ( load_matrix_sync, store_matrix_sync, mma_sync ) are assumed to be synchronous when used in context of global memory.
When using these functions in the context of shared memory (e.g. LDS memory), additional explicit workgroup synchronization may be required due to the nature this memory usage.
2.2. Supported Data Types#
rocWMMA mixed precision multiply-accumulate operations support the following data type combinations.
Data Types <Ti / To / Tc> = <Input type / Output Type / Compute Type>
where
Input Type = Matrix A/B
Output Type = Matrix C/D
Compute Type = math / accumulation type
Ti / To / Tc |
BlockM |
BlockN |
BlockK |
|---|---|---|---|
i8 / i32 / i32 |
16 |
16 |
Min: 16, pow2 |
i8 / i32 / i32 |
32 |
32 |
Min: 8, pow2 |
i8 / i8 / i32 |
16 |
16 |
Min: 16, pow2 |
i8 / i32 / i32 |
32 |
32 |
Min: 8, pow2 |
f16 / f32 / f32 |
16 |
16 |
Min: 16, pow2 |
f16 / f32 / f32 |
32 |
32 |
Min: 8, pow2 |
f16 / f16 / f32 |
16 |
16 |
Min: 16, pow2 |
f16 / f16 / f32 |
32 |
32 |
Min: 8, pow2 |
f16 / f16 / f16* |
16 |
16 |
Min: 16, pow2 |
f16 / f16 / f16* |
32 |
32 |
Min: 8, pow2 |
__half / f32 / f32 |
16 |
16 |
Min: 16, pow2 |
__half / f32 / f32 |
32 |
32 |
Min: 8, pow2 |
__half / __half / f32 |
16 |
16 |
Min: 16, pow2 |
__half / __half / f32 |
32 |
32 |
Min: 8, pow2 |
__half / __half / __half* |
16 |
16 |
Min: 16, pow2 |
__half / __half / __half* |
32 |
32 |
Min: 8, pow2 |
bf16 / f32 / f32 |
16 |
16 |
Min: 8, pow2 |
bf16 / f32 / f32 |
32 |
32 |
Min: 4, pow2 |
bf16 / bf16 / f32 |
16 |
16 |
Min: 8, pow2 |
bf16 / bf16 / f32 |
32 |
32 |
Min: 4, pow2 |
bf16 / bf16 / bf16* |
16 |
16 |
Min: 8, pow2 |
bf16 / bf16 / bf16* |
32 |
32 |
Min: 4, pow2 |
f32 / f32 / f32 |
16 |
16 |
Min: 4, pow2 |
f32 / f32 / f32 |
32 |
32 |
Min: 2, pow2 |
f64** / f64** / f64** |
16 |
16 |
Min: 4, pow2 |
*= matrix unit accumulation is natively 32 bit precision, and is converted to desired type.
**= f64 datatype is only supported on MI-200 class AMDGPU and successors.
2.3. Supported Matrix Layouts#
(N = col major, T = row major)
LayoutA |
LayoutB |
Layout C |
LayoutD |
|---|---|---|---|
N |
N |
N |
N |
N |
N |
T |
T |
N |
T |
N |
N |
N |
T |
T |
T |
T |
N |
N |
N |
T |
N |
T |
T |
T |
T |
N |
N |
T |
T |
T |
T |
2.3.1. Using rocWMMA API#
This section describes how to use the rocWMMA library API.
2.4. rocWMMA Datatypes#
-
struct matrix_a#
Input Matrix A.
-
struct matrix_b#
Input Matrix B.
-
struct accumulator#
Input/Output Matrix Accumulator.
-
struct row_major#
Data/In-memory Layout as Row Major.
-
struct col_major#
Data/In-memory Layout as Column Major.
-
class VecT#
HIP vector class.
- Template Parameters:
DataT – vector data type
Rank – vector size
Warning
doxygenstruct: Cannot find class “rocwmma::detail::VectorStorage” in doxygen xml output for project “rocWMMA Documentation” from directory: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-rocwmma/checkouts/docs-5.5.0/docs/.doxygen/docBin/xml
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayoutT>
struct IOConfig# Definition of ROCWMMA fragment input / output configurations in specific matrix context.
- Template Parameters:
- Param BlockDim:
leading block dimension (row / col size)
- Param KDim:
minor block dimension (row / col count)
- Param MaxVectorWidth:
maximum allowable vector width
- Param VectorWidth:
currently used vector width
- Param CoopIndex:
shared wave index (0 = row, 1 = col)
- Param IOTraits:
Input/output traits specific to AMDGCN architecture
- Param Packer:
Packs raw fragment data into register
- Param Unpacker:
Unpacks registers to raw fragment data
- Param Broadcaster:
Sets all fragment data to a desired value
- Param MatrixLayout:
Maps GPU threads to matrix shape or geometry
- Param Loader:
Issues load instructions for raw fragment data
- Param Storer:
Issues store instructions for raw fragment data
- Param CoopLoader:
Issues cooperative load instructions for raw fragment data
- Param CoopStorer:
Issues cooperative store instructions for raw fragment data
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayoutT>
struct IOShape# Definition of rocWMMA data and matrix mapping utilities in specific matrix context.
2.5. rocWMMA Enumeration#
Enumeration constants have numbering that is consistent with standard C++ libraries.
2.6. rocWMMA API functions#
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::fill_fragment(fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> &frag, DataT value)# Fills the entire fragment with the desired value.
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::load_matrix_sync(fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> &frag, const DataT *data, uint32_t ldm)# Loads the entire fragment from the data pointer according to its matrix and data layouts. Data pointer may point to either local or global memory.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT>
void rocwmma::load_matrix_sync(fragment<MatrixT, BlockM, BlockN, BlockK, DataT> &frag, const DataT *data, uint32_t ldm, layout_t layout)# Loads the entire fragment from the data pointer according to its matrix layout.Data pointer may point to either local or global memory. This overload provides a run-time ability to choose the data layout of the target fragment.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
layout – Matrix layout
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::store_matrix_sync(DataT *data, fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> const &frag, uint32_t ldm)# Stores the entire fragment to the data pointer according to its matrix and data layouts. Data pointer may point to either local or global memory.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT>
void rocwmma::store_matrix_sync(DataT *data, fragment<MatrixT, BlockM, BlockN, BlockK, DataT> const &frag, uint32_t ldm, layout_t layout)# Stores the entire fragment to the data pointer according to its matrix layout. Data pointer may point to either local or global memory. This overload provides a run-time ability to choose the data layout of the target fragment.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
layout – Data layout
- Template Parameters:
-
template<uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename InputT, typename ComputeT, typename LayoutA, typename LayoutB, typename LayoutC, typename LayoutD>
void rocwmma::mma_sync(fragment<accumulator, BlockM, BlockN, BlockK, ComputeT, LayoutD> &d, fragment<matrix_a, BlockM, BlockN, BlockK, InputT, LayoutA> const &a, fragment<matrix_b, BlockM, BlockN, BlockK, InputT, LayoutB> const &b, fragment<accumulator, BlockM, BlockN, BlockK, ComputeT, LayoutC> const &c)# Performs the Multiply-Accumulate operation on the fragments A, B, C and D(D = A * B + C)
Note
Frag c = d is valid
- Parameters:
d – Accumulator output D
a – Input fragment A
b – Input fragment B
c – Input accumulator fragment C
- Template Parameters:
-
void rocwmma::synchronize_workgroup()#
Synchronization point for all wavefronts in a workgroup.
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::load_matrix_coop_sync(fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> &frag, const DataT *data, uint32_t ldm, uint32_t waveIndex, uint32_t waveCount, uint32_t splitCount)# Cooperative Load Matrix - Loads the entire fragment with data from memory address cooperatively across waves. Each cooperative wave is responsible in loading a portion of the final fragment. Note that the full fragment data is not cohesive for individual waves as they only load a piece of the data. This function may be paired with store_matrix_sync to move a single fragment collaboratively between memory locations.
The full load is split into work items (splitCount). Work items are assigned in round robin fashion to waves in the range of [0, waveCount). The current wave index determines the order of the current wave in the collaboration pool. Work items are consumed in order by waves [0, waveCount) until there are no more work items and the operation is completed.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
waveIndex – Index assignment of current wave in collaboration
waveCount – Number of waves assigned for collaboration
splitCount – Number of work items to split the operation
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
inline void rocwmma::load_matrix_coop_sync(fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> &frag, const DataT *data, uint32_t ldm, uint32_t waveIndex, uint32_t waveCount)# Cooperative Load Matrix - Loads the entire fragment with data from memory address cooperatively across waves. Each cooperative wave is responsible in loading a portion of the final fragment. Note that the full fragment data is not cohesive for individual waves as they only load a piece of the data. This function may be paired with store_matrix_sync to move a single fragment collaboratively between memory locations.
The full load is split into work items (default = waveCount). Work items are assigned in round robin fashion to waves in the range of [0, waveCount). The current wave index determines the order of the current wave in the collaboration pool. Work items are consumed in order by waves [0, waveCount) until there are no more work items and the operation is completed.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
waveIndex – Index assignment of current wave in collaboration
waveCount – Number of waves assigned for collaboration
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::load_matrix_coop_sync(fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> &frag, const DataT *data, uint32_t ldm)# Cooperative Load Matrix - Loads the entire fragment with data from memory address cooperatively across waves. Each cooperative wave is responsible in loading a portion of the final fragment. Note that the full fragment data is not cohesive for individual waves as they only load a piece of the data. This function may be paired with store_matrix_sync to move a single fragment collaboratively between memory locations.
The full load is split into work items. This overload is conducive to GEMM workload where matrix_a fragments collaboratively load common data with other waves in the same row. Likewise, matrix_b fragments collaboratively load common data with other waves in the same column. Workload is split evenly across all waves in collaborative dimension. Split count = wave count = workgroupDim<1> (matrix_a) | workgroupDim<0> (matrix_b) Wave index = waveCoord<1> (matrix_a) | waveCoord<0> (matrix_b)
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::store_matrix_coop_sync(DataT *data, fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> const &frag, uint32_t ldm, uint32_t waveIndex, uint32_t waveCount, uint32_t splitCount)# Cooperative Store Matrix - Stores the entire fragment to data address cooperatively across waves. Each cooperative wave is responsible in storing a portion of the final fragment. Note that the full fragment data is not required to be cohesive for individual waves as they only store a piece of the data. This function may be paired with load_matrix_sync to move a single fragment collaboratively between memory locations.
The full store is split into work items (splitCount). Work items are assigned in round robin fashion to waves in the range of [0, waveCount). The current wave index determines the order of the current wave in the collaboration pool. Work items are consumed in order by waves [0, waveCount) until there are no more work items and the operation is completed.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
waveIndex – Index assignment of current wave in collaboration
waveCount – Number of waves assigned for collaboration
splitCount – Number of work items to split the operation
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::store_matrix_coop_sync(DataT *data, fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> const &frag, uint32_t ldm, uint32_t waveIndex, uint32_t waveCount)# Cooperative Store Matrix - Stores the entire fragment to data address cooperatively across waves. Each cooperative wave is responsible in storing a portion of the final fragment. Note that the full fragment data is not required to be cohesive for individual waves as they only store a piece of the data. This function may be paired with load_matrix_sync to move a single fragment collaboratively between memory locations.
The full store is split into work items (default = waveCount). Work items are assigned in round robin fashion to waves in the range of [0, waveCount). The current wave index determines the order of the current wave in the collaboration pool. Work items are consumed in order by waves [0, waveCount) until there are no more work items and the operation is completed.
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
waveIndex – Index assignment of current wave in collaboration
waveCount – Number of waves assigned for collaboration
- Template Parameters:
-
template<typename MatrixT, uint32_t BlockM, uint32_t BlockN, uint32_t BlockK, typename DataT, typename DataLayout>
void rocwmma::store_matrix_coop_sync(DataT *data, fragment<MatrixT, BlockM, BlockN, BlockK, DataT, DataLayout> const &frag, uint32_t ldm)# Cooperative Store Matrix - Stores the entire fragment to data address cooperatively across waves. Each cooperative wave is responsible in storing a portion of the final fragment. Note that the full fragment data is not required to be cohesive for individual waves as they only store a piece of the data. This function may be paired with load_matrix_sync to move a single fragment collaboratively between memory locations.
The full store is split into work items. This overload is conducive to GEMM workload where matrix_a fragments collaboratively use common data with other waves in the same row. Likewise, matrix_b fragments collaboratively use common data with other waves in the same column. Workload is split evenly across all waves in collaborative dimension.
Split count = wave count = workgroupDim<1> (matrix_a) | workgroupDim<0> (matrix_b) Wave index = waveCoord<1> (matrix_a) | waveCoord<0> (matrix_b)
- Parameters:
frag – Fragment of type MatrixT with its associated block sizes, data type and layout
data – Data pointer to global/local memory
ldm – Leading dimension size
- Template Parameters: