Add tests cases for q @ k attention variant

torchao/experimental/kernels/cpu/aarch64/matmul/channelwise_8bit_a_channelwise_8bit_b_4x8x8_f32_neondot-impl.h

@@ -288,7 +288,7 @@ struct KernelImpl<true, true, false, true> {
     constexpr int nr = 8;
     constexpr int kr = 8;
     assert(m % mr == 0);
-    assert(k % kr == 0);
+    assert(k % 16 == 0);
     assert(n >= nr);
     std::vector<int8_t> rhs_packed(n * k);
     // Since we are casting int8_t to float32_t in order to tranpose matrix in a

torchao/experimental/kernels/cpu/aarch64/tests/test_qmatmul.cpp

@@ -53,7 +53,7 @@ struct test_channelwise_8bit_channelwise_8bit_b<
         const int,
         const int);
     kernel_fn_type kernel_fn = nullptr;
-    if (use_gemm && (m % 4 == 0) && (n % 8 == 0) && (k % 8 == 0)) {
+    if (use_gemm && (m % 4 == 0) && (n % 8 == 0) && (k % 16 == 0)) {
       using namespace torchao::kernels::cpu::aarch64::quantized_matmul::
           channelwise_8bit_a_channelwise_8bit_b_4x8x8_f32_neondot;
       kernel_fn = kernel<a_has_zeros, b_has_zeros, false, true>;
@@ -531,9 +531,6 @@ static void test_8bit_per_token_q_at_k_matmul_attention(
       channelwise_8bit_a_channelwise_8bit_b_q_at_k_attention_test_case::
           generate(b, s_q, s_k, h, d, transpose);
 
-  using namespace torchao::kernels::cpu::aarch64::quantized_matmul::
-      channelwise_8bit_a_channelwise_8bit_b_1x8x16_f32_neondot;
-
   size_t q_b_stride = test_case.b_q_stride;
   size_t q_h_stride = test_case.h_q_stride;
   size_t q_s_q_stride = test_case.s_q_stride;
@@ -553,9 +550,36 @@ static void test_8bit_per_token_q_at_k_matmul_attention(
   size_t output_h_stride = s_q * s_k;
   size_t output_s_q_stride = s_k;
 
+  using kernel_fn_type = void (*)(
+      int,
+      int,
+      int,
+      const void*,
+      int,
+      const void*,
+      int,
+      float*,
+      int,
+      const int8_t*,
+      const int8_t*,
+      const float*,
+      const float*,
+      const int,
+      const int);
+  kernel_fn_type kernel_fn = nullptr;
+  if ((s_q % 4 == 0) && (s_k % 8 == 0) && (d % 16 == 0)) {
+    using namespace torchao::kernels::cpu::aarch64::quantized_matmul::
+        channelwise_8bit_a_channelwise_8bit_b_4x8x8_f32_neondot;
+    kernel_fn = kernel<true, true, false, true>;
+  } else {
+    using namespace torchao::kernels::cpu::aarch64::quantized_matmul::
+        channelwise_8bit_a_channelwise_8bit_b_1x8x16_f32_neondot;
+    kernel_fn = kernel<true, true, false, true>;
+  }
+
   for (int b_idx = 0; b_idx < b; b_idx++) {
     for (int h_idx = 0; h_idx < h; h_idx++) {
-      kernel<true, true, false, true>(
+      kernel_fn(
           s_q,
           s_k,
           d,
@@ -587,6 +611,14 @@ TEST(test_8bit_per_token_q_at_k_matmul_attention, Basic) {
   test_8bit_per_token_q_at_k_matmul_attention(1, 16, 16, 8, 16);
 }
 
+TEST(test_8bit_per_token_q_at_k_matmul_attention, BasicGemmKernel) {
+  test_8bit_per_token_q_at_k_matmul_attention(1, 4, 16, 4, 16);
+}
+
+TEST(test_8bit_per_token_q_at_k_matmul_attention, BasicGemmKernelNoTranspose) {
+  test_8bit_per_token_q_at_k_matmul_attention(1, 4, 16, 4, 16, false);
+}
+
 TEST(test_8bit_per_token_q_at_k_matmul_attention, PrimeHeadsAndHeadDim) {
   test_8bit_per_token_q_at_k_matmul_attention(1, 8, 8, 7, 33);
 }