CUDA implementation for IQ1_S_R4 (#492)

* iq1_s_r4: CUDA dequantize

* iq1_s_r4: CUDA GEMV

* iq1_s_r4: MMQ on CUDA

Requires Turing or better (will fall back to dequantize+cuBLAS on older cards).

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

9 files changed, 198 insertions, 0 deletions

diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index 7f4b01d4..de9816fe 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -3476,6 +3476,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                     case GGML_TYPE_IQ4_KS_R4:
                     case GGML_TYPE_IQ5_K_R4:
                     case GGML_TYPE_IQ5_KS_R4:
+                    case GGML_TYPE_IQ1_S_R4:
                         return true;
                     default:
                         return false;
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 896ba0df..bc7fadb0 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -516,6 +516,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
 };
 
 template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S_R4> {
+    static constexpr int qk = 32;
+    static constexpr int qr = 2;
+    static constexpr int qi = 4;
+};
+
+template<>
 struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR1_M;
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index 644ee316..b2f77e09 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -527,6 +527,41 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
 }
 
 template<typename dst_t>
+static __global__ void dequantize_block_iq1_s_r4(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t n_per_row, int64_t row_size) {
+
+    int64_t ii = blockIdx.x;
+
+    int64_t nblock = n_per_row/32;
+    int64_t row  = (8*ii)/nblock;
+    int64_t row4 = row/4;
+    int64_t ir   = row%4;
+    int64_t ibl  = (8*ii)%nblock;
+
+    const int tid = threadIdx.x;
+    const int  il = tid/8; // 0...3
+    const int  ib = tid%8; // 0...7
+
+    const half * dptr = (const half *)((const char *)vx + 4*row4*row_size);
+    const float d = (float)dptr[ir];
+    const block_iq1_s_r4 * x = (const block_iq1_s_r4 *)(dptr + 4) + ibl;
+    dst_t * y = yy + 256*ii + 32*ib + 8*il;
+
+    float dl = d*(2*((x[ib].qh[ir] >> 12) & 7) + 1);
+    float delta = dl * (x[ib].qh[ir] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA);
+
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[ib].qs[4*il+ir] | (((x[ib].qh[ir] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+
+    if constexpr (std::is_same_v<dst_t, nv_bfloat16>) {
+        for (int j = 0; j < 8; ++j) y[j] = __float2bfloat16(dl*q[j] + delta);
+    } else {
+        for (int j = 0; j < 8; ++j) y[j] = dl*q[j] + delta;
+    }
+}
+
+template<typename dst_t>
 static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
     const int64_t i   = blockIdx.x;
@@ -1399,6 +1434,14 @@ static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t
 }
 
 template<typename dst_t>
+static void dequantize_row_iq1_s_r4_cuda(const void * vx, dst_t * y, const int64_t nrows, const int64_t n_per_row, cudaStream_t stream) {
+    const int64_t k = nrows * n_per_row;
+    const int64_t row_size = ggml_row_size(GGML_TYPE_IQ1_S_R4, n_per_row);
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq1_s_r4<<<nb, 32, 0, stream>>>(vx, y, n_per_row, row_size);
+}
+
+template<typename dst_t>
 static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t nrows, const int64_t n_per_row, cudaStream_t stream) {
     const int64_t k = nrows * n_per_row;
     const int nb = (k + QK_K - 1) / QK_K;
@@ -1651,6 +1694,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
             return dequantize_row_iq5_k_r4_cuda<nv_bfloat16>;
         case GGML_TYPE_IQ5_KS_R4:
             return dequantize_row_iq5_ks_r4_cuda<nv_bfloat16>;
+        case GGML_TYPE_IQ1_S_R4:
+            return dequantize_row_iq1_s_r4_cuda<nv_bfloat16>;
         default:
             return nullptr;
     }
@@ -1699,6 +1744,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_S_R4:
+            return dequantize_row_iq1_s_r4_cuda;
         case GGML_TYPE_IQ1_M:
             return dequantize_row_iq1_m_cuda;
         case GGML_TYPE_IQ1_BN:
@@ -1790,6 +1837,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_S_R4:
+            return dequantize_row_iq1_s_r4_cuda;
         case GGML_TYPE_IQ1_M:
             return dequantize_row_iq1_m_cuda;
         case GGML_TYPE_IQ1_BN:
diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cu b/ggml/src/ggml-cuda/iqk_mmvq.cu
index ae11ae14..2340b54a 100644
--- a/ggml/src/ggml-cuda/iqk_mmvq.cu
+++ b/ggml/src/ggml-cuda/iqk_mmvq.cu
@@ -353,6 +353,38 @@ __device__ __forceinline__ void vec_dot_iq4_ks_r4_q8_1(
     }
 }
 
+// TODO
+__device__ __forceinline__ void vec_dot_iq1_s_r4_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs, float * result) {
+
+    const half * dptr = (const half *)vbq;
+    const block_iq1_s_r4 * bq1 = (const block_iq1_s_r4 *)(dptr + 4) + kbx;
+
+    // iqs is 0 or 2
+    const float d8 = __low2float(bq8_1->ds);
+    const int32_t  * q8 = (const int *)bq8_1->qs;
+
+    int32_t grid32[2];
+    const int * igrid = (const int *)grid32;
+
+    int minus = 0;
+    for (int k = 0; k < 4; ++k) minus = ggml_cuda_dp4a(0x01010101, q8[4*(iqs/2)+k], minus);
+
+    for (int i = 0; i < 4; ++i) {
+        float dl = (float)dptr[i]*(2*((bq1->qh[i] >> 12) & 7) + 1) * d8;
+        float ml = dl * (bq1->qh[i] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA);
+        grid32[0] = iq1s_grid_gpu[bq1->qs[4*iqs+i] | (((bq1->qh[i] >> 3*iqs) & 7) << 8)];
+        grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+        grid32[0] &= 0x0f0f0f0f;
+        int sumi = ggml_cuda_dp4a(igrid[0], q8[4*(iqs/2)+0], ggml_cuda_dp4a(igrid[1], q8[4*(iqs/2)+1], 0));
+        grid32[0] = iq1s_grid_gpu[bq1->qs[4*iqs+i+4] | (((bq1->qh[i] >> (3*iqs+3)) & 7) << 8)];
+        grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+        grid32[0] &= 0x0f0f0f0f;
+        sumi = ggml_cuda_dp4a(igrid[0], q8[4*(iqs/2)+2], ggml_cuda_dp4a(igrid[1], q8[4*(iqs/2)+3], sumi));
+        result[i] += dl * sumi + ml * minus;
+    }
+}
+
 #define VDR_IQ4_KS_Q8_1_MMVQ 4
 #define VDR_IQ4_KS_Q8_1_MMQ  4
 
@@ -1106,6 +1138,14 @@ void mul_mat_vec_iq4_ks_r4_q8_1_cuda(
     iqk_mul_mat_vec_q_cuda<GGML_TYPE_IQ4_KS_R4, 2, vec_dot_iq4_ks_r4_q8_1, 4>(vx, vy, dst, ids_data, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream);
 }
 
+void mul_mat_vec_iq1_s_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, int64_t ids_nb0, cudaStream_t stream) {
+
+    iqk_mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S_R4, 2, vec_dot_iq1_s_r4_q8_1, 4>(vx, vy, dst, ids_data, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream);
+}
+
 void mul_mat_vec_iq5_k_r4_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const char * ids_data,
     const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cuh b/ggml/src/ggml-cuda/iqk_mmvq.cuh
index 5278ef1e..1e4257e8 100644
--- a/ggml/src/ggml-cuda/iqk_mmvq.cuh
+++ b/ggml/src/ggml-cuda/iqk_mmvq.cuh
@@ -90,3 +90,8 @@ void mul_mat_vec_iq5_ks_r4_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const char * ids_data,
     const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
     const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq1_s_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 2f7a9bfd..60c2037f 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -85,6 +85,9 @@ void ggml_cuda_op_mul_mat_q(
         case GGML_TYPE_IQ1_S:
             mul_mat_q_case<GGML_TYPE_IQ1_S>(ctx, args, stream);
             break;
+        case GGML_TYPE_IQ1_S_R4:
+            mul_mat_q_case<GGML_TYPE_IQ1_S_R4>(ctx, args, stream);
+            break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_q_case<GGML_TYPE_IQ4_XS>(ctx, args, stream);
             break;
@@ -150,6 +153,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_KS:
@@ -174,6 +178,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     if (int8_mma_available(cc)) {
         return true;
     }
+    if (type == GGML_TYPE_IQ1_S_R4) {
+        return false;
+    }
 
     if (cc < MIN_CC_DP4A) {
         return false;
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 7a51c514..c98fa561 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -79,6 +79,7 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
         case GGML_TYPE_IQ3_S:
             return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_S_R4:
             return MMQ_Q8_1_DS_LAYOUT_DS4;
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
@@ -186,6 +187,7 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
         case GGML_TYPE_IQ3_XXS : return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_IQ3_S   : return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_IQ1_S   : return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ1_S_R4: return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_IQ4_XS  : return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_IQ4_NL  : return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_IQ4_KS  : return MMQ_DP4A_TXS_Q8_0;
@@ -231,6 +233,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
         case GGML_TYPE_IQ3_XXS : return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_IQ3_S   : return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_IQ1_S   : return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ1_S_R4: return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_IQ4_XS  : return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_IQ4_NL  : return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_IQ4_KS  : return MMQ_MMA_TILE_X_K_Q8_0;
@@ -318,6 +321,74 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     }
 }
 
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s_r4(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / 4;
+    const int kqsx = threadIdx.x % 4;
+
+    int32_t grid32[2];
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+        const int i4 = i/4;
+        const int ir = i%4;
+
+        const block_iq1_s_r4 * bxi = (const block_iq1_s_r4 *)(x + 4*i4*stride + 4*sizeof(half)) + kbx0 + kbx;
+
+        grid32[0] = iq1s_grid_gpu[bxi->qs[4*kqsx+ir] | (((bxi->qh[ir] >> 3*kqsx) & 7) << 8)];
+        grid32[1] = ((grid32[0] >> 4) & 0x0f0f0f0f) << 3;
+        grid32[0] = (grid32[0] & 0x0f0f0f0f) << 3;
+        const int shift = bxi->qh[ir] & 0x8000 ? 0x09090909 : 0x07070707;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kbx + 2*kqsx + 0] = __vsubss4(grid32[0], shift);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kbx + 2*kqsx + 1] = __vsubss4(grid32[1], shift);
+#else
+        // TODO
+        //x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / 4;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + threadIdx.y * 4 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+        const int i4 = i/4;
+        const int ir = i%4;
+
+        const half * dptr = (const half *)(x + 4*i4*stride);
+        const block_iq1_s_r4 * bxi = (const block_iq1_s_r4 *)(dptr + 4) + kbx0 + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0       + kbxd] = 0.125f * __half2float(dptr[ir]) * (((bxi->qh[ir] >> 11) & 14) + 1);
+#else
+        // TODO
+        //x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
@@ -3133,6 +3204,14 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ1_S> {
 };
 
 template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ1_S_R4> {
+    static constexpr int              vdr          = VDR_Q4_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq1_s_r4<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_DS4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_NL> {
     static constexpr int              vdr          = VDR_IQ4_NL_Q8_1_MMQ;
     static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_nl<mmq_y, nwarps, need_check>;
@@ -3656,6 +3735,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_K);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ5_K);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ5_KS);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ6_K);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S_R4);
 
 // -------------------------------------------------------------------------------------------------------------------------
 
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 6b5d37fa..cc00d278 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -560,6 +560,9 @@ static void ggml_cuda_op_mul_mat_vec_q_impl(ggml_backend_cuda_context & ctx, ggm
         case GGML_TYPE_IQ5_KS_R4:
             mul_mat_vec_iq5_ks_r4_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ids_data, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst,   ne2, nb02, nb12, nb2, ids_nb0, stream);
             break;
+        case GGML_TYPE_IQ1_S_R4:
+            mul_mat_vec_iq1_s_r4_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ids_data, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst,   ne2, nb02, nb12, nb2, ids_nb0, stream);
+            break;
         default:
             GGML_ABORT("fatal error");
             break;
@@ -679,6 +682,7 @@ bool ggml_cuda_mmvq_type_supported(ggml_type src0_type) {
         case GGML_TYPE_IQ4_KS_R4:
         case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ5_KS_R4:
+        case GGML_TYPE_IQ1_S_R4:
             return true;
         default:
             return false;
diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s_r4.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s_r4.cu
new file mode 100644
index 00000000..7bc5cf1a
--- /dev/null
+++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s_r4.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ1_S_R4);