luboslenco
79 lines
2.0 KiB
Plaintext
79 lines
2.0 KiB
Plaintext
#version 450
|
|
|
|
/*
|
|
* Iris Vulkan GEMM (bf16 weights)
|
|
*
|
|
* Computes C[M,N] = alpha * op(A)[M,K] @ op(B)[K,N] + beta * C[M,N]
|
|
* where A and C are f32 and B is bfloat16. Used for weight-bound linear
|
|
* layers: A is the f32 activation, B is the bf16 weight matrix.
|
|
*
|
|
* bf16 values are packed two-per-uint (little endian) so this needs no
|
|
* 16-bit storage extension. bf16 -> f32 is simply (bits << 16).
|
|
*/
|
|
|
|
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 { uint Bw[]; };
|
|
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];
|
|
|
|
float load_bf16(uint e) {
|
|
uint w = Bw[e >> 1u];
|
|
uint h = ((e & 1u) == 0u) ? (w & 0xffffu) : (w >> 16u);
|
|
return uintBitsToFloat(h << 16u);
|
|
}
|
|
|
|
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;
|
|
uint kB = t * 16u + ty;
|
|
|
|
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) {
|
|
uint e = (pc.tb == 0u) ? (kB * pc.ldb + col)
|
|
: (col * pc.ldb + kB);
|
|
bv = load_bf16(e);
|
|
}
|
|
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;
|
|
}
|
|
}
|