cuda : do warp-based block reduce

ggml-cuda.cu

@@ -502,6 +502,31 @@ static size_t g_scratch_offset = 0;
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+static __device__ __forceinline__ float warp_reduce_max(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+}
+
 static __global__ void add_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -578,15 +603,6 @@ static __global__ void sqr_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] * x[i];
 }
 
-static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
-    }
-    return a;
-}
-
 template <int block_size>
 static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
@@ -625,14 +641,6 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     }
 }
 
-static __device__ __forceinline__ float warp_reduce_sum(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
-    }
-    return x;
-}
-
 template <int block_size>
 static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
@@ -4718,7 +4726,6 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
 }
 
-// TODO: maybe can be improved with some warp-based primitives
 static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale) {
     const int tid  = threadIdx.x;
     const int rowx = blockIdx.x;
@@ -4726,24 +4733,26 @@ static __global__ void soft_max_f32(const float * x, const float * y, float * ds
 
     const int block_size = blockDim.x;
 
-    __shared__ float buf[CUDA_SOFT_MAX_BLOCK_SIZE];
+    float max_val = -INFINITY;
-
-    buf[tid] = -INFINITY;
 
     for (int col = tid; col < ncols; col += block_size) {
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
-        buf[tid] = max(buf[tid], x[ix]*scale + (y ? y[iy] : 0.0f));
+        max_val = max(max_val, x[ix]*scale + (y ? y[iy] : 0.0f));
     }
 
-    __syncthreads();
-
     // find the max value in the block
-    for (int i = block_size/2; i > 0; i >>= 1) {
+    max_val = warp_reduce_max(max_val);
-        if (tid < i) {
+    if (block_size > WARP_SIZE) {
-            buf[tid] = max(buf[tid], buf[tid + i]);
+        __shared__ float buf[CUDA_SOFT_MAX_BLOCK_SIZE/WARP_SIZE];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            buf[warp_id] = max_val;
         }
         __syncthreads();
+        max_val = buf[lane_id];
+        max_val = warp_reduce_max(max_val);
     }
 
     float tmp = 0.f;
@@ -4751,26 +4760,26 @@ static __global__ void soft_max_f32(const float * x, const float * y, float * ds
     for (int col = tid; col < ncols; col += block_size) {
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
-        const float val = expf((x[ix]*scale + (y ? y[iy] : 0.0f)) - buf[0]);
+        const float val = expf((x[ix]*scale + (y ? y[iy] : 0.0f)) - max_val);
         tmp += val;
         dst[ix] = val;
     }
 
-    __syncthreads();
+    // find the sum of exps in the block
-
+    tmp = warp_reduce_sum(tmp);
-    buf[tid] = tmp;
+    if (block_size > WARP_SIZE) {
-
+        __shared__ float buf[CUDA_SOFT_MAX_BLOCK_SIZE/WARP_SIZE];
-    __syncthreads();
+        int warp_id = threadIdx.x / WARP_SIZE;
-
+        int lane_id = threadIdx.x % WARP_SIZE;
-    // sum up partial sums
+        if (lane_id == 0) {
-    for (int i = block_size/2; i > 0; i >>= 1) {
+            buf[warp_id] = tmp;
-        if (tid < i) {
-            buf[tid] += buf[tid + i];
         }
         __syncthreads();
+        tmp = buf[lane_id];
+        tmp = warp_reduce_sum(tmp);
     }
 
-    const float inv_tmp = 1.f / buf[0];
+    const float inv_tmp = 1.f / tmp;
 
     for (int col = tid; col < ncols; col += block_size) {
         const int i = rowx*ncols + col;