luboslenco
71 lines
1.9 KiB
Plaintext
71 lines
1.9 KiB
Plaintext
#version 450
|
|
|
|
/*
|
|
* Iris Vulkan GEMM (f32)
|
|
*
|
|
* Computes C[M,N] = alpha * op(A)[M,K] @ op(B)[K,N] + beta * C[M,N]
|
|
* where op() optionally transposes, matching the BLAS sgemm contract used
|
|
* by the Iris kernels layer. Tiled with 16x16 shared-memory blocks.
|
|
*
|
|
* A, B, C are row-major with leading dims lda, ldb, ldc.
|
|
*/
|
|
|
|
layout(local_size_x = 16, local_size_y = 16) in;
|
|
|
|
layout(std430, binding = 0) readonly buffer ABuf { float A[]; };
|
|
layout(std430, binding = 1) readonly buffer BBuf { float B[]; };
|
|
layout(std430, binding = 2) buffer CBuf { float C[]; };
|
|
|
|
layout(push_constant) uniform PC {
|
|
uint M, N, K;
|
|
uint lda, ldb, ldc;
|
|
uint ta, tb;
|
|
float alpha, beta;
|
|
} pc;
|
|
|
|
shared float As[16][16];
|
|
shared float Bs[16][16];
|
|
|
|
void main() {
|
|
uint row = gl_GlobalInvocationID.y;
|
|
uint col = gl_GlobalInvocationID.x;
|
|
uint tx = gl_LocalInvocationID.x;
|
|
uint ty = gl_LocalInvocationID.y;
|
|
|
|
float acc = 0.0;
|
|
uint numTiles = (pc.K + 15u) / 16u;
|
|
|
|
for (uint t = 0u; t < numTiles; t++) {
|
|
uint kA = t * 16u + tx; /* column of A loaded by this lane */
|
|
uint kB = t * 16u + ty; /* row of B loaded by this lane */
|
|
|
|
float av = 0.0;
|
|
if (row < pc.M && kA < pc.K) {
|
|
av = (pc.ta == 0u) ? A[row * pc.lda + kA]
|
|
: A[kA * pc.lda + row];
|
|
}
|
|
As[ty][tx] = av;
|
|
|
|
float bv = 0.0;
|
|
if (col < pc.N && kB < pc.K) {
|
|
bv = (pc.tb == 0u) ? B[kB * pc.ldb + col]
|
|
: B[col * pc.ldb + kB];
|
|
}
|
|
Bs[ty][tx] = bv;
|
|
|
|
barrier();
|
|
|
|
for (uint k = 0u; k < 16u; k++) {
|
|
acc += As[ty][k] * Bs[k][tx];
|
|
}
|
|
|
|
barrier();
|
|
}
|
|
|
|
if (row < pc.M && col < pc.N) {
|
|
uint ci = row * pc.ldc + col;
|
|
float prev = (pc.beta != 0.0) ? C[ci] : 0.0;
|
|
C[ci] = pc.alpha * acc + pc.beta * prev;
|
|
}
|
|
}
|