p35189024
/
InfiniTensor
forked from jiuyuan/InfiniTensor
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

kv register is fp16

This commit is contained in:
xiaonans 2024-03-18 17:25:57 +08:00
parent 1e797d4ffe
commit db053e32a4
1 changed files with 17 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
src/kernels/cuda/attention_kvcache.cu
View File

@ -27,13 +27,14 @@ __global__ void _attention_kvcache_kernel_128_1(float* input_k_cache,
if(parallel_idx >= compMeta.dimSize[0] * compMeta.dimSize[1])
return;
float ptr_V[SEQ_UNIT*4]; // V
float ptr_K[SEQ_UNIT*4]; // K
half ptr_V[4]; // V
half ptr_K[4]; // K
float ptr_Q[4]; // Q
float ptr_P[SEQ_UNIT] = {0};
float ptr_O[4] = {0};
float ptr_sum[1] = {0};
float temp[4];
// readin Q
(float4 &)ptr_Q[0] = (float4 &)input_q[lane_id_x4 + (parallel_idx * 128)];
@ -44,43 +45,34 @@ __global__ void _attention_kvcache_kernel_128_1(float* input_k_cache,
#pragma unroll
for (int idx_SEQ_UNIT = 0; idx_SEQ_UNIT < SEQ_UNIT && idx_SEQ_UNIT + idx_seq < seq_length; idx_SEQ_UNIT ++) {
int idx_kvcache = common_idx + ((idx_SEQ_UNIT + idx_seq) * compMeta.stride[2]);
int idx_SEQ_UNIT_x4 = idx_SEQ_UNIT * 4;
half temp[4];
if(idx_SEQ_UNIT + idx_seq < seq_length - 1){
*((int2*)(&temp[0])) = *((int2*)(&((half*)input_k_cache)[idx_kvcache]));
for(int i = 0; i < 4; i += 2)
(float2 &)ptr_K[idx_SEQ_UNIT_x4 + i] = __half22float2((half2 &)temp[i]);
*((int2*)(&ptr_K[0])) = *((int2*)(&((half*)input_k_cache)[idx_kvcache]));
}
else{
(float4 &)ptr_K[idx_SEQ_UNIT_x4] = (float4 &) input_k[idx_kv];
for(int i = 0; i < 4; i += 2){
*((half2*)(&temp[i])) = __float22half2_rn(*((float2*)(&ptr_K[idx_SEQ_UNIT_x4 + i])));
}
*((int2*)(&((half*)input_k_cache)[idx_kvcache])) = *((int2*)(&temp[0]));
(float4 &)temp[0] = (float4 &) input_k[idx_kv];
for(int i = 0; i < 4; i += 2)
*((half2*)(&ptr_K[i])) = __float22half2_rn(*((float2*)(&temp[i])));
*((int2*)(&((half*)input_k_cache)[idx_kvcache])) = *((int2*)(&ptr_K[0]));
}
// * V
if(idx_SEQ_UNIT + idx_seq < seq_length - 1){
*((int2*)(&temp[0])) = *((int2*)(&((half*)input_v_cache)[idx_kvcache]));
for(int i = 0; i < 4; i += 2)
(float2 &)ptr_V[idx_SEQ_UNIT_x4 + i] = __half22float2((half2 &)temp[i]);
*((int2*)(&ptr_V[0])) = *((int2*)(&((half*)input_v_cache)[idx_kvcache]));
}
else{
(float4 &)ptr_V[idx_SEQ_UNIT_x4] = (float4 &) input_v[idx_kv];
for(int i = 0; i < 4; i += 2){
*((half2*)(&temp[i])) = __float22half2_rn(*((float2*)(&ptr_V[idx_SEQ_UNIT_x4 + i])));
}
*((int2*)(&((half*)input_v_cache)[idx_kvcache])) = *((int2*)(&temp[0]));
(float4 &)temp[0] = (float4 &) input_v[idx_kv];
for(int i = 0; i < 4; i += 2)
*((half2*)(&ptr_V[i])) = __float22half2_rn(*((float2*)(&temp[i])));
*((int2*)(&((half*)input_v_cache)[idx_kvcache])) = *((int2*)(&ptr_V[0]));
}
#pragma unroll
for (int i = 0; i < 4; i ++){
ptr_K[idx_SEQ_UNIT_x4 + i] = ptr_Q[i] * ptr_K[idx_SEQ_UNIT_x4 + i];
ptr_K[i] = __float2half(ptr_Q[i]) * ptr_K[i];
#pragma unroll
for (int offset = 16; offset > 0; offset /= 2) {
ptr_K[idx_SEQ_UNIT_x4 + i] += __shfl_down_sync(0xffffffff, ptr_K[idx_SEQ_UNIT_x4 + i], offset);
ptr_K[i] += __shfl_down_sync(0xffffffff, ptr_K[i], offset);
}
ptr_P[idx_SEQ_UNIT] += ptr_K[idx_SEQ_UNIT_x4 + i];
ptr_P[idx_SEQ_UNIT] += __half2float(ptr_K[i]);
}
// div sqrt(d)
@ -93,7 +85,7 @@ __global__ void _attention_kvcache_kernel_128_1(float* input_k_cache,
#pragma unroll
for (int i = 0; i < 4; i ++)
ptr_O[i] = fmaf(ptr_P[idx_SEQ_UNIT], ptr_V[(idx_SEQ_UNIT_x4 + i)], ptr_O[i]);
ptr_O[i] = fmaf(ptr_P[idx_SEQ_UNIT], __half2float(ptr_V[i]), ptr_O[i]);
}
#pragma unroll