Skip to content

Fixed __dp4a compute capability: 6.0 -> 6.1 #2189

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jul 12, 2023
Merged

Fixed __dp4a compute capability: 6.0 -> 6.1 #2189

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 11 additions & 11 deletions ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1258,7 +1258,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, float * __
}

static __device__ __forceinline__ float vec_dot_q4_0_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int iqs) {
#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
#if __CUDA_ARCH__ >= 610 // lowest compute capability for integer intrinsics
const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;

int vi;
Expand All @@ -1279,11 +1279,11 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1(const void * __restric
return sumi*d;
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= 600
#endif // __CUDA_ARCH__ >= 610
}

static __device__ __forceinline__ float vec_dot_q4_1_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int iqs) {
#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
#if __CUDA_ARCH__ >= 610 // lowest compute capability for integer intrinsics
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;

const int vi = *((int *) &bq4_1->qs[sizeof(int) * (iqs + 0)]);
Expand All @@ -1304,11 +1304,11 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1(const void * __restric
return sumi*d + m*s / QI4_1; // scale sum by QI4_1 because there are QI4_1 threads working on this block
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= 600
#endif // __CUDA_ARCH__ >= 610
}

static __device__ __forceinline__ float vec_dot_q5_0_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int iqs) {
#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
#if __CUDA_ARCH__ >= 610 // lowest compute capability for integer intrinsics
const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;

int qs;
Expand Down Expand Up @@ -1339,11 +1339,11 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1(const void * __restric
return sumi*d;
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= 600
#endif // __CUDA_ARCH__ >= 610
}

static __device__ __forceinline__ float vec_dot_q5_1_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int iqs) {
#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
#if __CUDA_ARCH__ >= 610 // lowest compute capability for integer intrinsics
const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;

const int qs = *((int *) &bq5_1->qs[sizeof(int) * (iqs + 0)]);
Expand Down Expand Up @@ -1373,11 +1373,11 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1(const void * __restric
return sumi*d + m*s / QI5_1; // scale sum by QI5_1 because there are QI5_1 threads working on this block
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= 600
#endif // __CUDA_ARCH__ >= 610
}

static __device__ __forceinline__ float vec_dot_q8_0_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int iqs) {
#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
#if __CUDA_ARCH__ >= 610 // lowest compute capability for integer intrinsics
const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;

int vi;
Expand All @@ -1392,7 +1392,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1(const void * __restric
return sumi*d;
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= 600
#endif // __CUDA_ARCH__ >= 610
}

template <int qk, int qi, typename block_q_t, vec_dot_q_cuda_t vec_dot_q_cuda>
Expand Down Expand Up @@ -2406,7 +2406,7 @@ inline void ggml_cuda_op_mul_mat_vec(
src0->type == GGML_TYPE_Q5_1 ||
src0->type == GGML_TYPE_Q8_0;

const bool use_mul_mat_vec_q = g_compute_capabilities[id] >= 600 && mul_mat_vec_q_implemented;
const bool use_mul_mat_vec_q = g_compute_capabilities[id] >= 610 && mul_mat_vec_q_implemented;
#endif

if (use_mul_mat_vec_q) {
Expand Down