Skip to content

LLM inference in C/C++

C++ ggml

opencl: fix two issues on flash attention for Adreno a7x (#25697)

Hongqiang Wangc3d47e696b11
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index b14ea8133b..16e851bbdb 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -7319,6 +7319,14 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
return false;
}
+ // Some compilers for A7x (Adreno 740, compiler E031.41) crashes when
+ // building FA kernels with mixed or quant types (f32_f16, f32_q8_0, f32_q4_0)
+ // Here we skip all A7x for these kernels to avoid crash
+ if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+ (is_f32_f16 || is_f32_q8_0 || is_f32_q4_0)) {
+ return false;
+ }
+
if (dk == 512) {
if (backend_ctx->gpu_family == INTEL) {
return false;
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
index 1cc0cc8c34..6e43ee81e7 100644
--- a/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
@@ -30,6 +30,10 @@
#elif defined(cl_qcom_subgroup_shuffle)
#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
#define HAS_SUBGROUP_SHUFFLE 1
+// Adreno compilers that expose only cl_qcom_subgroup_shuffle do not declare the KHR
+// name, so calling it is an implicit declaration and the program fails to build.
+// Route it to the qcom builtin.
+#define sub_group_shuffle_xor(val, mask) qcom_sub_group_shuffle_xor((val), (mask), CLK_SUB_GROUP_SHUFFLE_WIDTH_WAVE_SIZE_QCOM, 0.0f)
#endif
#define ACC_TYPE float
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl
index de09a1eaae..95d215971e 100644
--- a/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl
@@ -10,6 +10,10 @@
#elif defined(cl_qcom_subgroup_shuffle)
#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
#define HAS_SUBGROUP_SHUFFLE 1
+// Adreno compilers that expose only cl_qcom_subgroup_shuffle do not declare the KHR
+// name, so calling it is an implicit declaration and the program fails to build.
+// Route it to the qcom builtin.
+#define sub_group_shuffle_xor(val, mask) qcom_sub_group_shuffle_xor((val), (mask), CLK_SUB_GROUP_SHUFFLE_WIDTH_WAVE_SIZE_QCOM, 0.0f)
#endif
// Flash attention: Q=f32, K=q4_0, V=q4_0.
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl
index 46bc4bc9d9..7e89ed0bd8 100644
--- a/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl
@@ -10,6 +10,10 @@
#elif defined(cl_qcom_subgroup_shuffle)
#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
#define HAS_SUBGROUP_SHUFFLE 1
+// Adreno compilers that expose only cl_qcom_subgroup_shuffle do not declare the KHR
+// name, so calling it is an implicit declaration and the program fails to build.
+// Route it to the qcom builtin.
+#define sub_group_shuffle_xor(val, mask) qcom_sub_group_shuffle_xor((val), (mask), CLK_SUB_GROUP_SHUFFLE_WIDTH_WAVE_SIZE_QCOM, 0.0f)
#endif
// Flash attention: Q=f32, K=q8_0, V=q8_0.
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl
index da2e14ae99..97148d370f 100644
--- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl
@@ -24,6 +24,10 @@
#elif defined(cl_qcom_subgroup_shuffle)
#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
#define HAS_SUBGROUP_SHUFFLE 1
+// Adreno compilers that expose only cl_qcom_subgroup_shuffle do not declare the KHR
+// name, so calling it is an implicit declaration and the program fails to build.
+// Route it to the qcom builtin.
+#define sub_group_shuffle_xor(val, mask) qcom_sub_group_shuffle_xor((val), (mask), CLK_SUB_GROUP_SHUFFLE_WIDTH_WAVE_SIZE_QCOM, 0.0f)
#endif
// Assumes row size (ne00) is a multiple of 4