biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Enforce style for C++ and CUDA code with clang-format (#4722)

Browse Source
This commit is contained in:
Michael Goin
2024-05-22 03:18:41 -04:00
committed by GitHub
parent 9b9a10d6cb
commit 5f6d10c14c
64 changed files with 6398 additions and 6790 deletions
Download Patch File Download Diff File Expand all files Collapse all files

460
csrc/quantization/marlin/dense/marlin_cuda_kernel.cu
View File

@@ -25,7 +25,10 @@
#include <iostream>
template <typename T> inline std::string str(T x) { return std::to_string(x); }
template <typename T>
inline std::string str(T x) {
return std::to_string(x);
}
namespace marlin {
@@ -38,9 +41,10 @@ constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
// corresponding index accesses must be compile-time constants, which is why we
// extensively use `#pragma unroll` throughout the kernel code to guarantee
// this.
template <typename T, int n> struct Vec {
template <typename T, int n>
struct Vec {
T elems[n];
__device__ T &operator[](int i) { return elems[i]; }
__device__ T& operator[](int i) { return elems[i]; }
};
using I4 = Vec<int, 4>;
@@ -51,29 +55,32 @@ using I4 = Vec<int, 4>;
using FragA = Vec<half2, 4>;
using FragB = Vec<half2, 2>;
using FragC = Vec<float, 4>;
using FragS = Vec<half2, 1>; // quantization scales
using FragS = Vec<half2, 1>; // quantization scales
// Predicated asynchronous global->shared copy; used for inputs A where we apply
// predication to handle batchsizes that are not multiples of 16.
__device__ inline void cp_async4_pred(void *smem_ptr, const void *glob_ptr,
__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
bool pred = true) {
const int BYTES = 16;
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
asm volatile("{\n"
" .reg .pred p;\n"
" setp.ne.b32 p, %0, 0;\n"
" @p cp.async.cg.shared.global [%1], [%2], %3;\n"
"}\n" ::"r"((int)pred),
"r"(smem), "l"(glob_ptr), "n"(BYTES));
asm volatile(
"{\n"
" .reg .pred p;\n"
" setp.ne.b32 p, %0, 0;\n"
" @p cp.async.cg.shared.global [%1], [%2], %3;\n"
"}\n" ::"r"((int)pred),
"r"(smem), "l"(glob_ptr), "n"(BYTES));
}
// Asynchronous global->shared copy
__device__ inline void cp_async4(void *smem_ptr, const void *glob_ptr) {
__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
const int BYTES = 16;
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
asm volatile("{\n"
" cp.async.cg.shared.global [%0], [%1], %2;\n"
"}\n" :: "r"(smem), "l"(glob_ptr), "n"(BYTES));
asm volatile(
"{\n"
" cp.async.cg.shared.global [%0], [%1], %2;\n"
"}\n" ::"r"(smem),
"l"(glob_ptr), "n"(BYTES));
}
// Async copy fence.
@@ -82,28 +89,30 @@ __device__ inline void cp_async_fence() {
}
// Wait until at most `n` async copy stages are still pending.
template <int n> __device__ inline void cp_async_wait() {
template <int n>
__device__ inline void cp_async_wait() {
asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
}
// m16n8k16 tensor core mma instruction with fp16 inputs and fp32
// output/accumulation.
__device__ inline void mma(const FragA &a_frag, const FragB &frag_b,
FragC &frag_c) {
const uint32_t *a = reinterpret_cast<const uint32_t *>(&a_frag);
const uint32_t *b = reinterpret_cast<const uint32_t *>(&frag_b);
float *c = reinterpret_cast<float *>(&frag_c);
asm volatile("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 "
"{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n"
: "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
: "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]),
"r"(b[1]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]));
__device__ inline void mma(const FragA& a_frag, const FragB& frag_b,
FragC& frag_c) {
const uint32_t* a = reinterpret_cast<const uint32_t*>(&a_frag);
const uint32_t* b = reinterpret_cast<const uint32_t*>(&frag_b);
float* c = reinterpret_cast<float*>(&frag_c);
asm volatile(
"mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 "
"{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n"
: "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
: "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]),
"f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]));
}
// Instruction for loading a full 16x16 matrix fragment of operand A from shared
// memory, directly in tensor core layout.
__device__ inline void ldsm4(FragA &frag_a, const void *smem_ptr) {
uint32_t *a = reinterpret_cast<uint32_t *>(&frag_a);
__device__ inline void ldsm4(FragA& frag_a, const void* smem_ptr) {
uint32_t* a = reinterpret_cast<uint32_t*>(&frag_a);
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
asm volatile("ldmatrix.sync.aligned.m8n8.x4.shared.b16 {%0,%1,%2,%3}, [%4];\n"
: "=r"(a[0]), "=r"(a[1]), "=r"(a[2]), "=r"(a[3])
@@ -113,7 +122,8 @@ __device__ inline void ldsm4(FragA &frag_a, const void *smem_ptr) {
// Lookup-table based 3-input logical operation; explicitly used for
// dequantization as the compiler does not seem to automatically recognize it in
// all cases.
template <int lut> __device__ inline int lop3(int a, int b, int c) {
template <int lut>
__device__ inline int lop3(int a, int b, int c) {
int res;
asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
: "=r"(res)
@@ -138,24 +148,24 @@ __device__ inline FragB dequant(int q) {
const int MUL = 0x2c002c00;
const int ADD = 0xd480d480;
FragB frag_b;
frag_b[0] = __hsub2(*reinterpret_cast<half2 *>(&lo),
*reinterpret_cast<const half2 *>(&SUB));
frag_b[1] = __hfma2(*reinterpret_cast<half2 *>(&hi),
*reinterpret_cast<const half2 *>(&MUL),
*reinterpret_cast<const half2 *>(&ADD));
frag_b[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
*reinterpret_cast<const half2*>(&SUB));
frag_b[1] = __hfma2(*reinterpret_cast<half2*>(&hi),
*reinterpret_cast<const half2*>(&MUL),
*reinterpret_cast<const half2*>(&ADD));
return frag_b;
}
// Multiply dequantized values by the corresponding quantization scale; used
// only for grouped quantization.
__device__ inline void scale(FragB &frag_b, FragS &frag_s, int i) {
half2 s = __half2half2(reinterpret_cast<__half *>(&frag_s)[i]);
__device__ inline void scale(FragB& frag_b, FragS& frag_s, int i) {
half2 s = __half2half2(reinterpret_cast<__half*>(&frag_s)[i]);
frag_b[0] = __hmul2(frag_b[0], s);
frag_b[1] = __hmul2(frag_b[1], s);
}
// Wait until barrier reaches `count`, then lock for current threadblock.
__device__ inline void barrier_acquire(int *lock, int count) {
__device__ inline void barrier_acquire(int* lock, int count) {
if (threadIdx.x == 0) {
int state = -1;
do
@@ -170,7 +180,7 @@ __device__ inline void barrier_acquire(int *lock, int count) {
}
// Release barrier and increment visitation count.
__device__ inline void barrier_release(int *lock, bool reset = false) {
__device__ inline void barrier_release(int* lock, bool reset = false) {
__syncthreads();
if (threadIdx.x == 0) {
if (reset) {
@@ -187,26 +197,27 @@ __device__ inline void barrier_release(int *lock, bool reset = false) {
}
}
template <const int threads, // number of threads in a threadblock
const int thread_m_blocks, // number of 16x16 blocks in the m
// dimension (batchsize) of the threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks with
// a separate quantization scale
template <const int threads, // number of threads in a threadblock
const int thread_m_blocks, // number of 16x16 blocks in the m
// dimension (batchsize) of the
// threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__global__ void
Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
const int4 *__restrict__ B, // 4bit quantized weight matrix of shape kxn
int4 *__restrict__ C, // fp16 output buffer of shape mxn
const int4
*__restrict__ s, // fp16 quantization scales of shape (k/groupsize)xn
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int *locks // extra global storage for barrier synchronization
__global__ void Marlin(
const int4* __restrict__ A, // fp16 input matrix of shape mxk
const int4* __restrict__ B, // 4bit quantized weight matrix of shape kxn
int4* __restrict__ C, // fp16 output buffer of shape mxn
const int4* __restrict__ s, // fp16 quantization scales of shape
// (k/groupsize)xn
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int* locks // extra global storage for barrier synchronization
) {
// Each threadblock processes one "stripe" of the B matrix with (roughly) the
// same size, which might involve multiple column "slices" (of width 16 *
@@ -241,11 +252,11 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
int slice_row = (iters * blockIdx.x) % k_tiles;
int slice_col_par = (iters * blockIdx.x) / k_tiles;
int slice_col = slice_col_par;
int slice_iters; // number of threadblock tiles in the current slice
int slice_iters; // number of threadblock tiles in the current slice
int slice_count =
0; // total number of active threadblocks in the current slice
int slice_idx; // index of threadblock in current slice; numbered bottom to
// top
0; // total number of active threadblocks in the current slice
int slice_idx; // index of threadblock in current slice; numbered bottom to
// top
// We can easily implement parallel problem execution by just remapping
// indices and advancing global pointers
@@ -261,27 +272,22 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
auto init_slice = [&]() {
slice_iters =
iters * (blockIdx.x + 1) - (k_tiles * slice_col_par + slice_row);
if (slice_iters < 0 || slice_col_par >= n_tiles * parallel)
slice_iters = 0;
if (slice_iters == 0)
return;
if (slice_row + slice_iters > k_tiles)
slice_iters = k_tiles - slice_row;
if (slice_iters < 0 || slice_col_par >= n_tiles * parallel) slice_iters = 0;
if (slice_iters == 0) return;
if (slice_row + slice_iters > k_tiles) slice_iters = k_tiles - slice_row;
slice_count = 1;
slice_idx = 0;
int col_first = iters * ceildiv(k_tiles * slice_col_par, iters);
if (col_first <= k_tiles * (slice_col_par + 1)) {
int col_off = col_first - k_tiles * slice_col_par;
slice_count = ceildiv(k_tiles - col_off, iters);
if (col_off > 0)
slice_count++;
if (col_off > 0) slice_count++;
int delta_first = iters * blockIdx.x - col_first;
if (delta_first < 0 || (col_off == 0 && delta_first == 0))
slice_idx = slice_count - 1;
else {
slice_idx = slice_count - 1 - delta_first / iters;
if (col_off > 0)
slice_idx--;
if (col_off > 0) slice_idx--;
}
}
if (slice_col == n_tiles) {
@@ -293,29 +299,30 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
};
init_slice();
int a_gl_stride = prob_k / 8; // stride of the A matrix in global memory
int a_gl_stride = prob_k / 8; // stride of the A matrix in global memory
// We typically use `constexpr` to indicate that this value is a compile-time
// constant
constexpr int a_sh_stride =
16 * thread_k_blocks / 8; // stride of an A matrix tile in shared memory
16 * thread_k_blocks / 8; // stride of an A matrix tile in shared memory
constexpr int a_gl_rd_delta_o =
16 * thread_k_blocks /
8; // delta between subsequent A tiles in global memory
8; // delta between subsequent A tiles in global memory
int a_gl_rd_delta_i =
a_gl_stride *
(threads / a_gl_rd_delta_o); // between subsequent accesses within a tile
(threads / a_gl_rd_delta_o); // between subsequent accesses within a tile
constexpr int a_sh_wr_delta =
a_sh_stride * (threads / a_gl_rd_delta_o); // between shared memory writes
a_sh_stride *
(threads / a_gl_rd_delta_o); // between shared memory writes
constexpr int a_sh_rd_delta_o =
2 * ((threads / 32) /
(thread_n_blocks / 4)); // between shared memory tile reads
(thread_n_blocks / 4)); // between shared memory tile reads
constexpr int a_sh_rd_delta_i =
a_sh_stride * 16; // within a shared memory tile
a_sh_stride * 16; // within a shared memory tile
constexpr int a_sh_stage =
a_sh_stride * (16 * thread_m_blocks); // overall size of a tile
a_sh_stride * (16 * thread_m_blocks); // overall size of a tile
constexpr int a_sh_wr_iters =
ceildiv(a_sh_stage,
a_sh_wr_delta); // number of shared write iterations for a tile
a_sh_wr_delta); // number of shared write iterations for a tile
int b_gl_stride = 16 * prob_n / 32;
constexpr int b_sh_stride = 32 * thread_n_blocks / 4;
@@ -368,7 +375,7 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// needed if there are more threads than required for a certain tilesize or
// when the batchsize is not a multiple of 16.
bool a_sh_wr_pred[a_sh_wr_iters];
#pragma unroll
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++)
a_sh_wr_pred[i] = a_sh_wr_delta * i + a_sh_wr < a_sh_stride * prob_m;
bool s_sh_wr_pred = threadIdx.x < s_sh_stride;
@@ -387,13 +394,13 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// loop unrolls, all shared memory accesses are static, we simply precompute
// both transformed reads and writes.
int a_sh_wr_trans[a_sh_wr_iters];
#pragma unroll
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++)
a_sh_wr_trans[i] = transform_a(a_sh_wr_delta * i + a_sh_wr);
int a_sh_rd_trans[b_sh_wr_iters][thread_m_blocks];
#pragma unroll
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++) {
#pragma unroll
#pragma unroll
for (int j = 0; j < thread_m_blocks; j++)
a_sh_rd_trans[i][j] =
transform_a(a_sh_rd_delta_o * i + a_sh_rd_delta_i * j + a_sh_rd);
@@ -403,16 +410,16 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// runtime; we break dependencies between subsequent accesses with a tile by
// maintining multiple pointers (we have enough registers), a tiny
// optimization.
const int4 *B_ptr[b_sh_wr_iters];
#pragma unroll
const int4* B_ptr[b_sh_wr_iters];
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++)
B_ptr[i] = B + b_gl_rd_delta_i * i + b_gl_rd;
extern __shared__ int4 sh[];
// Shared memory storage for global fetch pipelines.
int4 *sh_a = sh;
int4 *sh_b = sh_a + (stages * a_sh_stage);
int4 *sh_s = sh_b + (stages * b_sh_stage);
int4* sh_a = sh;
int4* sh_b = sh_a + (stages * a_sh_stage);
int4* sh_s = sh_b + (stages * b_sh_stage);
// Register storage for double buffer of shared memory reads.
FragA frag_a[2][thread_m_blocks];
I4 frag_b_quant[2];
@@ -421,34 +428,33 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// Zero accumulators.
auto zero_accums = [&]() {
#pragma unroll
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4 * 2 * 4; i++)
reinterpret_cast<float *>(frag_c)[i] = 0;
reinterpret_cast<float*>(frag_c)[i] = 0;
};
// Asynchronously fetch the next A, B and s tile from global to the next
// shared memory pipeline location.
auto fetch_to_shared = [&](int pipe, int a_off, bool pred = true) {
if (pred) {
int4 *sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++) {
cp_async4_pred(
&sh_a_stage[a_sh_wr_trans[i]],
&A[a_gl_rd_delta_i * i + a_gl_rd + a_gl_rd_delta_o * a_off],
a_sh_wr_pred[i]);
}
int4 *sh_b_stage = sh_b + b_sh_stage * pipe;
#pragma unroll
int4* sh_b_stage = sh_b + b_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++) {
cp_async4(&sh_b_stage[b_sh_wr_delta * i + b_sh_wr], B_ptr[i]);
B_ptr[i] += b_gl_rd_delta_o;
}
// Only fetch scales if this tile starts a new group
if (group_blocks != -1 && pipe % (group_blocks / thread_k_blocks) == 0) {
int4 *sh_s_stage = sh_s + s_sh_stage * pipe;
if (s_sh_wr_pred)
cp_async4(&sh_s_stage[s_sh_wr], &s[s_gl_rd]);
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
if (s_sh_wr_pred) cp_async4(&sh_s_stage[s_sh_wr], &s[s_gl_rd]);
s_gl_rd += s_gl_rd_delta;
}
}
@@ -475,37 +481,35 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// theoretically better attempts have lead to bad instruction ordering by
// the compiler and correspondingly a noticeable drop in performance.
if (group_blocks != -1) {
int4 *sh_s_stage =
int4* sh_s_stage =
sh_s + s_sh_stage * ((group_blocks / thread_k_blocks) *
(pipe / (group_blocks / thread_k_blocks)));
reinterpret_cast<int4 *>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
}
int4 *sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++)
ldsm4(frag_a[k % 2][i], &sh_a_stage[a_sh_rd_trans[k % b_sh_wr_iters][i]]);
int4 *sh_b_stage = sh_b + b_sh_stage * pipe;
frag_b_quant[k % 2] = *reinterpret_cast<I4 *>(
int4* sh_b_stage = sh_b + b_sh_stage * pipe;
frag_b_quant[k % 2] = *reinterpret_cast<I4*>(
&sh_b_stage[b_sh_rd_delta * (k % b_sh_wr_iters) + b_sh_rd]);
};
// Execute the actual tensor core matmul of a sub-tile.
auto matmul = [&](int k) {
// We have the m dimension as the inner loop in order to encourage overlapping
// dequantization and matmul operations.
#pragma unroll
// We have the m dimension as the inner loop in order to encourage overlapping
// dequantization and matmul operations.
#pragma unroll
for (int j = 0; j < 4; j++) {
int b_quant = frag_b_quant[k % 2][j];
int b_quant_shift = b_quant >> 8;
FragB frag_b0 = dequant(b_quant);
// If there are no groups, we can just scale the final output once and can
// avoid doing so for each weight.
if (group_blocks != -1)
scale(frag_b0, frag_s[k % 2][j], 0);
if (group_blocks != -1) scale(frag_b0, frag_s[k % 2][j], 0);
FragB frag_b1 = dequant(b_quant_shift);
if (group_blocks != -1)
scale(frag_b1, frag_s[k % 2][j], 1);
#pragma unroll
if (group_blocks != -1) scale(frag_b1, frag_s[k % 2][j], 1);
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++) {
mma(frag_a[k % 2][i], frag_b0, frag_c[i][j][0]);
mma(frag_a[k % 2][i], frag_b1, frag_c[i][j][1]);
@@ -530,38 +534,38 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// unnecessary read or write iterations, e.g., for two warps we write only
// once by warp 1 and read only once by warp 0.
#pragma unroll
#pragma unroll
for (int m_block = 0; m_block < thread_m_blocks; m_block++) {
#pragma unroll
#pragma unroll
for (int i = red_off; i > 0; i /= 2) {
if (i <= red_idx && red_idx < 2 * i) {
#pragma unroll
#pragma unroll
for (int j = 0; j < 4 * 2; j++) {
int red_sh_wr =
red_sh_delta * j + (red_sh_rd - red_sh_stride * i);
if (i < red_off) {
float *c_rd = reinterpret_cast<float *>(
&sh[red_sh_delta * j + red_sh_rd]);
float *c_wr = reinterpret_cast<float *>(&sh[red_sh_wr]);
#pragma unroll
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * j + red_sh_rd]);
float* c_wr = reinterpret_cast<float*>(&sh[red_sh_wr]);
#pragma unroll
for (int k = 0; k < 4; k++)
reinterpret_cast<FragC *>(frag_c)[4 * 2 * m_block + j][k] +=
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + j][k] +=
c_rd[k] + c_wr[k];
}
sh[red_sh_wr] =
reinterpret_cast<int4 *>(&frag_c)[4 * 2 * m_block + j];
reinterpret_cast<int4*>(&frag_c)[4 * 2 * m_block + j];
}
}
__syncthreads();
}
if (red_idx == 0) {
#pragma unroll
#pragma unroll
for (int i = 0; i < 4 * 2; i++) {
float *c_rd =
reinterpret_cast<float *>(&sh[red_sh_delta * i + red_sh_rd]);
#pragma unroll
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * i + red_sh_rd]);
#pragma unroll
for (int j = 0; j < 4; j++)
reinterpret_cast<FragC *>(frag_c)[4 * 2 * m_block + i][j] +=
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + i][j] +=
c_rd[j];
}
}
@@ -571,9 +575,9 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
};
// Since multiple threadblocks may process parts of the same column slice, we
// finally have to globally reduce over the results. As the striped partitioning
// minimizes the number of such reductions and our outputs are usually rather
// small, we perform this reduction serially in L2 cache.
// finally have to globally reduce over the results. As the striped
// partitioning minimizes the number of such reductions and our outputs are
// usually rather small, we perform this reduction serially in L2 cache.
auto global_reduce = [&](bool first = false, bool last = false) {
// We are very careful here to reduce directly in the output buffer to
// maximize L2 cache utilization in this step. To do this, we write out
@@ -592,39 +596,39 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
int row = (threadIdx.x % 32) / 4;
if (!first) {
// Interestingly, doing direct global accesses here really seems to mess up the
// compiler and lead to slowdowns, hence we also use async-copies even though
// these fetches are not actually asynchronous.
#pragma unroll
// Interestingly, doing direct global accesses here really seems to mess up
// the compiler and lead to slowdowns, hence we also use async-copies even
// though these fetches are not actually asynchronous.
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
cp_async4_pred(&sh[c_sh_wr + c_sh_wr_delta * i],
&C[c_gl_wr + c_gl_wr_delta_o * (i / 2) +
c_gl_wr_delta_i * (i % 2)],
i < (thread_m_blocks - 1) * 4 ||
8 * (i / 2) + row < prob_m);
cp_async4_pred(
&sh[c_sh_wr + c_sh_wr_delta * i],
&C[c_gl_wr + c_gl_wr_delta_o * (i / 2) +
c_gl_wr_delta_i * (i % 2)],
i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m);
}
cp_async_fence();
cp_async_wait<0>();
}
#pragma unroll
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
if (i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m) {
if (!first) {
int4 c_red = sh[c_sh_wr + i * c_sh_wr_delta];
#pragma unroll
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<float *>(
reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)] +=
__half2float(reinterpret_cast<__half *>(&c_red)[j]);
__half2float(reinterpret_cast<__half*>(&c_red)[j]);
}
}
if (!last) {
int4 c;
#pragma unroll
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<__half *>(&c)[j] =
__float2half(reinterpret_cast<float *>(
reinterpret_cast<__half*>(&c)[j] =
__float2half(reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)]);
}
C[c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2)] =
@@ -658,17 +662,17 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// We first reorder in shared memory to guarantee the most efficient final
// global write patterns
auto write = [&](int idx, float c0, float c1, FragS &s) {
auto write = [&](int idx, float c0, float c1, FragS& s) {
half2 res = __halves2half2(__float2half(c0), __float2half(c1));
if (group_blocks ==
-1) // for per-column quantization we finally apply the scale here
-1) // for per-column quantization we finally apply the scale here
res = __hmul2(res, s[0]);
((half2 *)sh)[idx] = res;
((half2*)sh)[idx] = res;
};
if (threadIdx.x / 32 < thread_n_blocks / 4) {
#pragma unroll
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++) {
#pragma unroll
#pragma unroll
for (int j = 0; j < 4; j++) {
int wr = c_sh_wr + 8 * j;
write(wr + (4 * c_sh_stride) * 0 + 0, frag_c[i][j][0][0],
@@ -685,7 +689,7 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
}
__syncthreads();
#pragma unroll
#pragma unroll
for (int i = 0;
i < ceildiv(16 * thread_m_blocks, threads / (2 * thread_n_blocks));
i++) {
@@ -699,9 +703,8 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// Start global fetch and register load pipelines.
auto start_pipes = [&]() {
#pragma unroll
for (int i = 0; i < stages - 1; i++)
fetch_to_shared(i, i, i < slice_iters);
#pragma unroll
for (int i = 0; i < stages - 1; i++) fetch_to_shared(i, i, i < slice_iters);
zero_accums();
wait_for_stage();
fetch_to_registers(0, 0);
@@ -711,12 +714,12 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// Main loop.
while (slice_iters) {
// We unroll over both the global fetch and the register load pipeline to ensure
// all shared memory accesses are static. Note that both pipelines have even
// length meaning that the next iteration will always start at index 0.
#pragma unroll
// We unroll over both the global fetch and the register load pipeline to
// ensure all shared memory accesses are static. Note that both pipelines have
// even length meaning that the next iteration will always start at index 0.
#pragma unroll
for (int pipe = 0; pipe < stages;) {
#pragma unroll
#pragma unroll
for (int k = 0; k < b_sh_wr_iters; k++) {
fetch_to_registers(k + 1, pipe % stages);
if (k == b_sh_wr_iters - 2) {
@@ -728,8 +731,7 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
matmul(k);
}
slice_iters--;
if (slice_iters == 0)
break;
if (slice_iters == 0) break;
}
a_gl_rd += a_gl_rd_delta_o * stages;
@@ -742,8 +744,7 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// For per-column scales, we only fetch them here in the final step before
// write-out
if (group_blocks == -1 && last) {
if (s_sh_wr_pred)
cp_async4(&sh_s[s_sh_wr], &s[s_gl_rd]);
if (s_sh_wr_pred) cp_async4(&sh_s[s_sh_wr], &s[s_gl_rd]);
cp_async_fence();
}
thread_block_reduce();
@@ -751,17 +752,17 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
cp_async_wait<0>();
__syncthreads();
if (threadIdx.x / 32 < thread_n_blocks / 4) {
reinterpret_cast<int4 *>(&frag_s)[0] = sh_s[s_sh_rd + 0];
reinterpret_cast<int4 *>(&frag_s)[1] = sh_s[s_sh_rd + 4];
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd + 0];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
}
if (slice_count > 1) { // only globally reduce if there is more than one
// block in a slice
if (slice_count > 1) { // only globally reduce if there is more than one
// block in a slice
barrier_acquire(&locks[slice_col], slice_idx);
global_reduce(slice_idx == 0, last);
barrier_release(&locks[slice_col], last);
}
if (last) // only the last block in a slice actually writes the result
if (last) // only the last block in a slice actually writes the result
write_result();
slice_row = 0;
slice_col_par++;
@@ -770,13 +771,12 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
if (slice_iters) {
a_gl_rd = a_gl_stride * (threadIdx.x / a_gl_rd_delta_o) +
(threadIdx.x % a_gl_rd_delta_o);
#pragma unroll
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++)
B_ptr[i] += b_sh_stride - b_gl_rd_delta_o * k_tiles;
if (slice_col == 0) {
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++)
B_ptr[i] -= b_gl_stride;
#pragma unroll
for (int i = 0; i < b_sh_wr_iters; i++) B_ptr[i] -= b_gl_stride;
}
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
start_pipes();
@@ -787,26 +787,27 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
#else
template <const int threads, // number of threads in a threadblock
const int thread_m_blocks, // number of 16x16 blocks in the m
// dimension (batchsize) of the threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks with
// a separate quantization scale
template <const int threads, // number of threads in a threadblock
const int thread_m_blocks, // number of 16x16 blocks in the m
// dimension (batchsize) of the
// threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__global__ void
Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
const int4 *__restrict__ B, // 4bit quantized weight matrix of shape kxn
int4 *__restrict__ C, // fp16 output buffer of shape mxn
const int4
*__restrict__ s, // fp16 quantization scales of shape (k/groupsize)xn
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int *locks // extra global storage for barrier synchronization
__global__ void Marlin(
const int4* __restrict__ A, // fp16 input matrix of shape mxk
const int4* __restrict__ B, // 4bit quantized weight matrix of shape kxn
int4* __restrict__ C, // fp16 output buffer of shape mxn
const int4* __restrict__ s, // fp16 quantization scales of shape
// (k/groupsize)xn
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int* locks // extra global storage for barrier synchronization
) {
// Marlin is not implemented yet for SM < 8.0
assert(false);
@@ -819,10 +820,10 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
// than 1 warp per schedule allows some more latency hiding. At the same time,
// we want relatively few warps to have many registers per warp and small tiles.
const int USER_THREADS =
256; // Note: This is only used with user-provided thread_k/n
const int STAGES = 4; // 4 pipeline stages fit into shared memory
256; // Note: This is only used with user-provided thread_k/n
const int STAGES = 4; // 4 pipeline stages fit into shared memory
const int SHARED_MEM =
96 * 1024; // max shared memory on compute capability 8.6 (< 8.0)
96 * 1024; // max shared memory on compute capability 8.6 (< 8.0)
static constexpr int min_thread_n = 64;
static constexpr int min_thread_k = 64;
@@ -831,7 +832,7 @@ static constexpr int tile_size = 16;
static constexpr int max_par = 16;
static constexpr int pack_factor_4bit =
8; // We have 8 4-bit vals inside a 32 bit
8; // We have 8 4-bit vals inside a 32 bit
#define __CALL_IF(THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
GROUP_BLOCKS, NUM_THREADS) \
@@ -858,23 +859,23 @@ thread_config_t small_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{128, 128, 256}, // Default
{128, 64, 128}, // Reduce N 2X, same K
{64, 256, 256}, // Reduce K 2X, increase N 2X
{64, 128, 128}, // Reduce K 2X, same N
{128, 128, 256}, // Default
{128, 64, 128}, // Reduce N 2X, same K
{64, 256, 256}, // Reduce K 2X, increase N 2X
{64, 128, 128}, // Reduce K 2X, same N
};
thread_config_t large_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{64, 256, 256}, // Default
{128, 128, 256}, // Reduce N 2X, increase K 2X
{64, 128, 128}, // Reduce N 2X, same K
{128, 64, 128}, // Reduce N 4X, increase K 2X
{64, 256, 256}, // Default
{128, 128, 256}, // Reduce N 2X, increase K 2X
{64, 128, 128}, // Reduce N 2X, same K
{128, 64, 128}, // Reduce N 4X, increase K 2X
};
bool is_valid_config(thread_config_t const &th_config, int prob_m, int prob_n,
bool is_valid_config(thread_config_t const& th_config, int prob_m, int prob_n,
int prob_k) {
// Sanity
if (th_config.thread_k == -1 || th_config.thread_n == -1 ||
@@ -907,7 +908,6 @@ bool is_valid_config(thread_config_t const &th_config, int prob_m, int prob_n,
}
thread_config_t determine_thread_config(int prob_m, int prob_n, int prob_k) {
if (prob_m <= 16) {
for (auto th_config : small_batch_thread_configs) {
if (is_valid_config(th_config, prob_m, prob_n, prob_k)) {
@@ -926,20 +926,20 @@ thread_config_t determine_thread_config(int prob_m, int prob_n, int prob_k) {
return thread_config_t{-1, -1, -1};
}
#define CALL_IF(N_BLOCKS, K_BLOCKS, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(2, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(2, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(3, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(3, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(4, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
#define CALL_IF(N_BLOCKS, K_BLOCKS, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(1, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(2, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(2, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(3, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(3, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS) \
__CALL_IF(4, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(4, N_BLOCKS, K_BLOCKS, 8, NUM_THREADS)
void marlin_cuda(const void *A, const void *B, void *C, void *s, int prob_m,
int prob_n, int prob_k, void *workspace, int groupsize = -1,
void marlin_cuda(const void* A, const void* B, void* C, void* s, int prob_m,
int prob_n, int prob_k, void* workspace, int groupsize = -1,
int dev = 0, cudaStream_t stream = 0, int thread_k = -1,
int thread_n = -1, int sms = -1, int max_par = 16) {
int tot_m = prob_m;
@@ -996,12 +996,12 @@ void marlin_cuda(const void *A, const void *B, void *C, void *s, int prob_m,
" is not divisible by group_blocks = ", group_blocks);
}
const int4 *A_ptr = (const int4 *)A;
const int4 *B_ptr = (const int4 *)B;
int4 *C_ptr = (int4 *)C;
const int4 *s_ptr = (const int4 *)s;
const int4* A_ptr = (const int4*)A;
const int4* B_ptr = (const int4*)B;
int4* C_ptr = (int4*)C;
const int4* s_ptr = (const int4*)s;
int *locks = (int *)workspace;
int* locks = (int*)workspace;
for (int i = 0; i < tot_m_blocks; i += 4) {
int thread_m_blocks = tot_m_blocks - i;
@@ -1011,8 +1011,7 @@ void marlin_cuda(const void *A, const void *B, void *C, void *s, int prob_m,
// Note that parallel > 1 currently only works for inputs without any
// padding
par = (16 * thread_m_blocks - pad) / 64;
if (par > max_par)
par = max_par;
if (par > max_par) par = max_par;
prob_m = 64 * par;
i += 4 * (par - 1);
thread_m_blocks = 4;
@@ -1041,12 +1040,11 @@ void marlin_cuda(const void *A, const void *B, void *C, void *s, int prob_m,
}
}
} // namespace marlin
} // namespace marlin
torch::Tensor marlin_gemm(torch::Tensor &a, torch::Tensor &b_q_weight,
torch::Tensor &b_scales, torch::Tensor &workspace,
torch::Tensor marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
torch::Tensor& b_scales, torch::Tensor& workspace,
int64_t size_m, int64_t size_n, int64_t size_k) {
// Verify M
TORCH_CHECK(size_m == a.size(0),
"Shape mismatch: a.size(0) = " + str(a.size(0)) +
@@ -1074,9 +1072,9 @@ torch::Tensor marlin_gemm(torch::Tensor &a, torch::Tensor &b_q_weight,
int actual_size_n =
(b_q_weight.size(1) / marlin::tile_size) * marlin::pack_factor_4bit;
TORCH_CHECK(size_n == actual_size_n,
"size_n = " + str(size_n) +
", actual_size_n = " + str(actual_size_n));
TORCH_CHECK(
size_n == actual_size_n,
"size_n = " + str(size_n) + ", actual_size_n = " + str(actual_size_n));
// Verify A device and strides
TORCH_CHECK(a.device().is_cuda(), "A is not on GPU");