whispercpp 1.3.7 → 1.3.8

data/ext/sources/ggml/src/ggml-sycl/softmax.cpp

@@ -56,7 +56,7 @@ static void soft_max_f32(const float *         x,
                                : block_size_template;
     const int nthreads = block_size;
     const int nwarps = nthreads / WARP_SIZE;
-    size_t nreduce = nwarps / WARP_SIZE;
+    const size_t nreduce = nwarps / WARP_SIZE;
 
     const int tid = item_ct1.get_local_id(2);
 
@@ -105,17 +105,15 @@ static void soft_max_f32(const float *         x,
     max_val = warp_reduce_max<WARP_SIZE>(max_val);
 
     if (block_size > WARP_SIZE) {
-        if (warp_id == 0) {
-            buf_iw[lane_id] = -INFINITY;
-        }
-        item_ct1.barrier();
-
         if (lane_id == 0) {
             buf_iw[warp_id] = max_val;
         }
         item_ct1.barrier();
 
-        max_val = buf_iw[lane_id];
+        max_val = -INFINITY;
+        for (int i = lane_id; i < nwarps; i += WARP_SIZE) {
+            max_val = sycl::max(max_val, buf_iw[i]);
+        }
         max_val = warp_reduce_max<WARP_SIZE>(max_val);
     }
     float tmp = 0.0f; // partial sum
@@ -128,7 +126,7 @@ static void soft_max_f32(const float *         x,
             break;
         }
 
-        const float val = sycl::native::exp(vals[col] - max_val);
+        const float val = sycl::native::exp(sycl::max(vals[col] - max_val, -80.0f));
         tmp += val;
         vals[col] = val;
     }
@@ -156,7 +154,7 @@ static void soft_max_f32(const float *         x,
         tmp = warp_reduce_sum<WARP_SIZE>(tmp);
     }
     if (sinks) {
-        tmp += sycl::native::exp(sinks[i02] - max_val);
+        tmp += sycl::native::exp(sycl::max(sinks[i02] - max_val, -80.0f));
     }
     const float inv_sum = 1.0f / tmp;
 
@@ -290,7 +288,8 @@ static void soft_max_f32_sycl(const float *x, const T *mask,
 
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     soft_max_f32<false, 0, 0>(
                         x, mask, sinks, dst, params,
                         dpct_local_acc_ct1

data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp

@@ -309,6 +309,41 @@ vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh,
         vl, vh, u[0], u[1], scales[0], scales[4], d, d8[0], d8[1]);
 }
 
+#define VDR_Q1_0_Q8_1_MMVQ 1
+#define VDR_Q1_0_Q8_1_MMQ  4
+
+static __dpct_inline__ float
+vec_dot_q1_0_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q1_0 * bq1_0 = (const block_q1_0 *) vbq;
+
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+    const float        d1          = bq1_0->d;
+    const int          v           = get_int_from_uint8_aligned(bq1_0->qs, iqs);
+
+    int vi_bytes[8];
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int shift = j * 4;
+        const int bits4 = (v >> shift) & 0x0F;
+        const int b0    = (bits4 & 0x01) ? 1 : -1;
+        const int b1    = (bits4 & 0x02) ? 1 : -1;
+        const int b2    = (bits4 & 0x04) ? 1 : -1;
+        const int b3    = (bits4 & 0x08) ? 1 : -1;
+        vi_bytes[j]     = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+    }
+
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int u = get_int_from_int8_aligned(bq8_1_chunk->qs, j);
+        sumi        = ggml_sycl_dp4a(vi_bytes[j], u, sumi);
+    }
+
+    return d1 * bq8_1_chunk->ds[0] * sumi;
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 

data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt

@@ -108,6 +108,9 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_CHECK_RESULTS)
         add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        # the result-checking path computes a CPU reference graph via
+        # ggml_graph_compute_with_ctx(), which is defined in ggml-cpu
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
     endif()
 
     if (GGML_VULKAN_DEBUG)
@@ -129,6 +132,8 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_RUN_TESTS)
         add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        # the test path also calls ggml_graph_compute_with_ctx() (ggml-cpu)
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
     endif()
 
     # Set up toolchain for host compilation whether cross-compiling or not