ggml-cuda: ds_read_b128 for q4_0 and q4_1 mmq kernels (#21168)

* ds_read_b128 for q4_0 and q4_1 mmq kernels

     Current for loop generates ds_read_b32 instructions with hip compiler, the new solution generates ds_read_b128 instructions for the same operation, saving some LDS bandwidth. Tested on MI50 and RX6800XT, its faster on both.

* Vectorized lds load update: used ggml_cuda_get_max_cpy_bytes and ggml_cuda_memcpy_1 functions for generic implementation

* Explicit for loop in mmq, renamed vec into tmp

* Fixed max_cpy usage in the loading loop

* Fixed typo in q4_1 kernel

* Update ggml/src/ggml-cuda/mmq.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Update ggml/src/ggml-cuda/mmq.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Update ggml/src/ggml-cuda/mmq.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Renoved trailing white line 500

* Update mmq.cuh removed other whitelines

* Remove trailing whitespaces

---------

Co-authored-by: iacopPBK <iacopPBK@users.noreply.github.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: iacopPBK <iacop@deneb.com>

Author: 3 people

ggml/src/ggml-cuda/mmq.cuh

@@ -386,17 +386,25 @@ static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
 #pragma unroll
             for (int i0 = 0; i0 < mmq_y; i0 += warp_size) {
                 const int i = i0 + threadIdx.x;
-
                 const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
 
                 int u[2*VDR_Q4_0_Q8_1_MMQ];
 
-#pragma unroll
-                for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
-                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
-                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_0)];
+                constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes();
+                constexpr int mcpy_int = max_cpy / sizeof(int);
+                static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ");
+
+                int tmp0[4], tmp1[4];
+
+                #pragma unroll
+                for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) {
+                    ggml_cuda_memcpy_1<max_cpy>(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int]  );
+                    ggml_cuda_memcpy_1<max_cpy>(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_0 + l0 * mcpy_int]);
                 }
 
+                u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3];
+                u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3];
+
                 sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
                     (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_0], u,
                      x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
@@ -489,17 +497,25 @@ static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
 #pragma unroll
             for (int i0 = 0; i0 < mmq_y; i0 += warp_size) {
                 const int i = i0 + threadIdx.x;
-
                 const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
 
                 int u[2*VDR_Q4_1_Q8_1_MMQ];
 
-#pragma unroll
-                for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
-                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
-                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_1)];
+                constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes();
+                constexpr int mcpy_int = max_cpy / sizeof(int);
+                static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ");
+
+                int tmp0[4], tmp1[4];
+
+                #pragma unroll
+                for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) {
+                    ggml_cuda_memcpy_1<max_cpy>(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int]  );
+                    ggml_cuda_memcpy_1<max_cpy>(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_1 + l0 * mcpy_int]);
                 }
 
+                u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3];
+                u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3];
+
                 sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
                     (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_1], u,
                      x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
@@ -4170,3 +4186,4 @@ void ggml_cuda_op_mul_mat_q(
     const int64_t src1_padded_row_size, cudaStream_t stream);
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts);
+