opencl: avoid the vec path in GEMV for unaligned row stride (#25671)

The f16 GEMV kernels take a vectorized path for ne00 >= 128 that casts the row
pointers to half4 or float4. When the row stride is not aligned, the wide load
becomes misaligned. On devices that require natural alignment for vector loads,
the kernel reads garbage. This is the case Intel GPUs and the kernels produce
incorrect results there. Adreno happpens to be byte addressable and the kernels
happen to work.

diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl
index 9393b5494158ac099a59b054c9adedcd68d1096d..b4b03eb11a32b1d5942fd5df831dfc31ecacdaee 100644 (file)
--- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl
@@ -64,7 +64,14 @@ kernel void kernel_mul_mat_f16_f16(
 
     global half * x = (global half *) (src0 + offset_src0);
 
-    if (ne00 < 128) {
+    // The vector path below casts the row pointers to half4, which must be 8-byte aligned.
+    // A row address is r0*nb01 + ..., and a permuted or strided src leaves nb01/nb11
+    // unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned. Every
+    // src1 row this work-item walks is src1_base + r1*nb11, so require both.
+    const ulong src1_base  = (ulong) (src1 + (i12)*nb12 + (i13)*nb13);
+    const bool  row_aligned = (((ulong) x) & 7) == 0 && (src1_base & 7) == 0 && (nb11 & 7) == 0;
+
+    if (ne00 < 128 || !row_aligned) {
         for (int row = 0; row < N_F16_F16; ++row) {
             int r1 = rb + row;
             if (r1 >= ne11) {

diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl
index e52d3c6d47558c9f203c0fc6c155a4833b50b576..8f3ed9c7b9cde5a377eed88e913d17148b7f30d1 100644 (file)
--- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl
@@ -64,7 +64,14 @@ kernel void kernel_mul_mat_f16_f32(
 
     global half * x = (global half *) (src0 + offset_src0);
 
-    if (ne00 < 128) {
+    // The vector path below casts the row pointers to half4/float4, which must be 8- and
+    // 16-byte aligned. A row address is r0*nb01 + ..., and a permuted or strided src leaves
+    // nb01/nb11 unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned.
+    // Every src1 row this work-item walks is src1_base + r1*nb11, so require both.
+    const ulong src1_base  = (ulong) (src1 + (i12)*nb12 + (i13)*nb13);
+    const bool  row_aligned = (((ulong) x) & 7) == 0 && (src1_base & 15) == 0 && (nb11 & 15) == 0;
+
+    if (ne00 < 128 || !row_aligned) {
         for (int row = 0; row < N_F16_F32; ++row) {
             int r1 = rb + row;
             if (r1 >= ne11) {

diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl
index 28d30212cda902b8b394886636060ee6fee80350..eca45615efd9dacc956ca7adb5a1954ee0e968bf 100644 (file)
--- a/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl
@@ -64,8 +64,15 @@ kernel void kernel_mul_mat_f16_f32_1row(
     global half  * x = (global half  *) (src0 + offset_src0);
     global float * y = (global float *) (src1 + offset_src1);
 
+    // The vector path below casts the row pointers to half4/float4, which must be 8- and
+    // 16-byte aligned. A row address is r0*nb01 + ..., and a permuted or strided src leaves
+    // nb01/nb11 unconstrained -- an odd ne00, say, gives a row that is only 2-byte aligned.
+    // Take the vector path only when the rows this work-item touches are actually aligned;
+    // the scalar loop has no such requirement.
+    const bool row_aligned = (((ulong) x) & 7) == 0 && (((ulong) y) & 15) == 0;
+
     float sumf = 0;
-    if (ne00 < 128) {
+    if (ne00 < 128 || !row_aligned) {
         for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
             sumf += (float) x[i] * (float) y[i];
         }