CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix MUSA compiler warning (#12704)

Browse Source 
* fix MUSA compiler warning

* replace (void) with GGML_UNUSED
This commit is contained in:
a3sh
2025-04-03 15:32:55 +08:00
committed by GitHub
parent 65cfe136a0
commit 193c3e03a6
2 changed files with 44 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
ggml/src/ggml-cuda/ssm-conv.cu
View File

@@ -4,13 +4,14 @@ template <size_t split_d_inner, size_t d_conv>
static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1, static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1, const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2, float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
const int nc, const int ncs, const int nr, const int n_t, const int n_s) { const int64_t n_t) {
GGML_UNUSED(src0_nb0);
const int tid = threadIdx.x; const int tid = threadIdx.x;
const int bidx = blockIdx.x; const int bidx = blockIdx.x;
const int bidy = blockIdx.y; const int bidy = blockIdx.y;
const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1); const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1);
const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1); const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1);
float * y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0); float * y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0);
const int stride_x = src0_nb1 / sizeof(float); const int stride_x = src0_nb1 / sizeof(float);
@@ -21,15 +22,15 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
float w[d_conv] = { 0.0f }; float w[d_conv] = { 0.0f };
#pragma unroll #pragma unroll
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
w[j] = w_block[tid * stride_w + j]; w[j] = w_block[tid * stride_w + j];
} }
for (int i = 0; i < n_t; i++) { for (int64_t i = 0; i < n_t; i++) {
float sumf = 0.0f; float sumf = 0.0f;
if (i == 0) { if (i == 0) {
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
x[j] = x_block[tid * stride_x + j]; x[j] = x_block[tid * stride_x + j];
} }
} else { } else {
@@ -37,27 +38,26 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
} }
#pragma unroll #pragma unroll
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
sumf += x[(i + j) % d_conv] * w[j]; sumf += x[(i + j) % d_conv] * w[j];
} }
y_block[i * stride_y + tid] = sumf; y_block[i * stride_y + tid] = sumf;
} }
} }
template <size_t split_d_inner, size_t d_conv, size_t split_n_t> template <size_t split_d_inner, size_t d_conv, int64_t split_n_t>
static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1, static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src0_nb0, const int src0_nb1, const int src0_nb2,
const int src1_nb1, float * __restrict__ dst, const int dst_nb0, const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
const int dst_nb1, const int dst_nb2, const int nc, const int ncs, const int dst_nb1, const int dst_nb2, const int64_t n_t) {
const int nr, const int n_t, const int n_s) {
const int tid = threadIdx.x; const int tid = threadIdx.x;
const int bidx = blockIdx.x; const int bidx = blockIdx.x;
const int bidy = blockIdx.y; const int bidy = blockIdx.y;
const int bidz = blockIdx.z; const int bidz = blockIdx.z;
const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 + const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 +
bidz * split_n_t * src0_nb0); bidz * split_n_t * src0_nb0);
const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1); const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1);
float * y_block = float * y_block =
(float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0); (float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0);
@@ -69,17 +69,17 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
float w[d_conv] = { 0.0f }; float w[d_conv] = { 0.0f };
#pragma unroll #pragma unroll
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
w[j] = w_block[tid * stride_w + j]; w[j] = w_block[tid * stride_w + j];
} }
#pragma unroll #pragma unroll
for (int i = 0; i < split_n_t; i++) { for (int64_t i = 0; i < split_n_t; i++) {
if (bidz * split_n_t + i < n_t) { if (bidz * split_n_t + i < n_t) {
float sumf = 0.0f; float sumf = 0.0f;
if (i == 0) { if (i == 0) {
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
x[j] = x_block[tid * stride_x + j]; x[j] = x_block[tid * stride_x + j];
} }
} else { } else {
@@ -87,7 +87,7 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
} }
#pragma unroll #pragma unroll
for (int j = 0; j < d_conv; j++) { for (size_t j = 0; j < d_conv; j++) {
sumf += x[(i + j) % d_conv] * w[j]; sumf += x[(i + j) % d_conv] * w[j];
} }
y_block[i * stride_y + tid] = sumf; y_block[i * stride_y + tid] = sumf;
@@ -97,8 +97,8 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1, static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1, const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
const int dst_nb2, const int nc, const int ncs, const int nr, const int n_t, const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t,
const int n_s, cudaStream_t stream) { const int64_t n_s, cudaStream_t stream) {
const int threads = 128; const int threads = 128;
GGML_ASSERT(nr % threads == 0); GGML_ASSERT(nr % threads == 0);
@@ -106,18 +106,16 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
const dim3 blocks(n_s, (nr + threads - 1) / threads, 1); const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
if (nc == 4) { if (nc == 4) {
ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
dst, dst_nb0, dst_nb1, dst_nb2, nc, ncs, nr, n_t, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
n_s);
} else { } else {
GGML_ABORT("Only support kernel size = 4 now."); GGML_ABORT("Only support kernel size = 4 now.");
} }
} else { } else {
if (nc == 4) { if (nc == 4) {
const int split_n_t = 32; const int64_t split_n_t = 32;
dim3 blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t); dim3 blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
ssm_conv_long_token_f32<threads, 4, split_n_t> ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
<<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
dst_nb1, dst_nb2, nc, ncs, nr, n_t, n_s);
} else { } else {
GGML_ABORT("Only support kernel size = 4 right now."); GGML_ABORT("Only support kernel size = 4 right now.");
} }
@@ -128,11 +126,10 @@ void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0]; // conv_x const struct ggml_tensor * src0 = dst->src[0]; // conv_x
const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight
const int nc = src1->ne[0]; // d_conv const int64_t nc = src1->ne[0]; // d_conv
const int ncs = src0->ne[0]; // d_conv - 1 + n_t const int64_t nr = src0->ne[1]; // d_inner
const int nr = src0->ne[1]; // d_inner const int64_t n_t = dst->ne[1]; // tokens per sequence
const int n_t = dst->ne[1]; // tokens per sequence const int64_t n_s = dst->ne[2]; // number of sequences in the batch
const int n_s = dst->ne[2]; // number of sequences in the batch
GGML_ASSERT(dst->ne[0] == nr); GGML_ASSERT(dst->ne[0] == nr);
GGML_ASSERT(src0->nb[0] == sizeof(float)); GGML_ASSERT(src0->nb[0] == sizeof(float));
@@ -147,5 +144,5 @@ void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32);
ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1], ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1],
dst->nb[2], nc, ncs, nr, n_t, n_s, stream); dst->nb[2], nc, nr, n_t, n_s, stream);
} }

34
ggml/src/ggml-cuda/ssm-scan.cu
View File

@@ -1,10 +1,5 @@
#include "ssm-scan.cuh" #include "ssm-scan.cuh"
// #include <cuda_runtime.h>
// static __device__ void global_to_shared(const float *src, float *dst) {
// asm volatile("cp.async.");
// }
template <size_t splitD, size_t N> template <size_t splitD, size_t N>
__global__ void __launch_bounds__(splitD, 2) __global__ void __launch_bounds__(splitD, 2)
ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2, ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
@@ -12,7 +7,9 @@ __global__ void __launch_bounds__(splitD, 2)
const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1, const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2, const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * __restrict__ dst, const int D, const int L, const int B) { float * __restrict__ dst, const int64_t L) {
GGML_UNUSED(src1_nb0);
GGML_UNUSED(src2_nb0);
const int bidx = blockIdx.x; // split along B const int bidx = blockIdx.x; // split along B
const int bidy = blockIdx.y; // split along D const int bidy = blockIdx.y; // split along D
const int tid = threadIdx.x; const int tid = threadIdx.x;
@@ -25,12 +22,12 @@ __global__ void __launch_bounds__(splitD, 2)
float * smem_A = smem; float * smem_A = smem;
float * smem_s0 = smem_A + splitD * stride_sA; float * smem_s0 = smem_A + splitD * stride_sA;
const float * s0_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1); const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
const float * x_block = (const float *) ((char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float)); const float * x_block = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
const float * dt_block = (const float *) ((char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float)); const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
const float * A_block = (const float *) ((char *) src3 + bidy * splitD * src3_nb1); const float * A_block = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
const float * B_block = (const float *) ((char *) src4 + (bidx * src4_nb2)); const float * B_block = (const float *) ((const char *) src4 + (bidx * src4_nb2));
const float * C_block = (const float *) ((char *) src5 + (bidx * src5_nb2)); const float * C_block = (const float *) ((const char *) src5 + (bidx * src5_nb2));
float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float)); float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1); float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
@@ -46,7 +43,7 @@ __global__ void __launch_bounds__(splitD, 2)
// can N not be 16? for example 32? // can N not be 16? for example 32?
if (N == 16) { if (N == 16) {
#pragma unroll #pragma unroll
for (int i = 0; i < splitD / 4; i += 2) { for (size_t i = 0; i < splitD / 4; i += 2) {
float value = A_block[(wid * warpSize + i) * stride_A + wtid]; float value = A_block[(wid * warpSize + i) * stride_A + wtid];
// todo: bank conflict // todo: bank conflict
// I am always confused with how to use the swizzling method to solve // I am always confused with how to use the swizzling method to solve
@@ -54,7 +51,7 @@ __global__ void __launch_bounds__(splitD, 2)
smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value; smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
} }
#pragma unroll #pragma unroll
for (int i = 0; i < splitD / 4; i += 2) { for (size_t i = 0; i < splitD / 4; i += 2) {
float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid]; float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value; smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
} }
@@ -62,7 +59,7 @@ __global__ void __launch_bounds__(splitD, 2)
__syncthreads(); __syncthreads();
for (int i = 0; i < L; i++) { for (int64_t i = 0; i < L; i++) {
float dt_soft_plus = dt_block[i * stride_dt + tid]; float dt_soft_plus = dt_block[i * stride_dt + tid];
if (dt_soft_plus <= 20.0f) { if (dt_soft_plus <= 20.0f) {
dt_soft_plus = log1pf(exp(dt_soft_plus)); dt_soft_plus = log1pf(exp(dt_soft_plus));
@@ -70,7 +67,7 @@ __global__ void __launch_bounds__(splitD, 2)
float x_dt = x_block[i * stride_x + tid] * dt_soft_plus; float x_dt = x_block[i * stride_x + tid] * dt_soft_plus;
float sumf = 0.0f; float sumf = 0.0f;
#pragma unroll #pragma unroll
for (int j = 0; j < N; j++) { for (size_t j = 0; j < N; j++) {
float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) + float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) +
(B_block[i * stride_B + j] * x_dt); (B_block[i * stride_B + j] * x_dt);
sumf += state * C_block[i * stride_C + j]; sumf += state * C_block[i * stride_C + j];
@@ -90,7 +87,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3, const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2, const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * dst, const int N, const int D, const int L, const int B, cudaStream_t stream) { float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
cudaStream_t stream) {
const int threads = 128; const int threads = 128;
// todo: consider D cannot be divided,does this situation exist? // todo: consider D cannot be divided,does this situation exist?
GGML_ASSERT(D % threads == 0); GGML_ASSERT(D % threads == 0);
@@ -99,7 +97,7 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
if (N == 16) { if (N == 16) {
ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>( ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0, src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, D, L, B); src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
} else { } else {
GGML_ABORT("doesn't support N!=16."); GGML_ABORT("doesn't support N!=16.");
} }