ggml : retire custom ggml_2d impl in favor of general (#352)

diff --git a/src/ggml.c b/src/ggml.c
index e45f8ef892f304a6ccc23dacc93d7b30b573aee5..f5821f1f1aa042d3c04eabc36b9aca9678ea4e87 100644 (file)
--- a/src/ggml.c
+++ b/src/ggml.c
@@ -12992,12 +12992,13 @@ static void ggml_compute_forward_conv_1d(
     };
 }
 
-// ggml_compute_forward_conv_2d_sk_p0
+// ggml_compute_forward_conv_2d
 
-static void ggml_compute_forward_conv_2d_sk_p0_f16_f32(
+static void ggml_compute_forward_conv_2d_f16_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
+        const struct ggml_tensor * opt0,
               struct ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -13017,120 +13018,12 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32(
     // size of the convolution row - the kernel size unrolled across all channels
     const int ew0 = nk0*nk1*ne02;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
-        memset(params->wdata, 0, params->wsize);
-
-        // prepare source data (src1)
-        {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-
-            for (int i13 = 0; i13 < ne13; i13++) {
-                for (int i12 = 0; i12 < ne12; i12++) {
-                    const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12);
-                    ggml_fp16_t * dst_data = wdata + i13*(ne1*ne0*ew0);
-
-                    for (int i1 = 0; i1 < ne1; i1++) {
-                        for (int i0 = 0; i0 < ne0; i0++) {
-                            for (int ik1 = 0; ik1 < nk1; ik1++) {
-                                for (int ik0 = 0; ik0 < nk0; ik0++) {
-                                    dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
-                                        GGML_FP32_TO_FP16(src[(i1*nk1 + ik1)*ne10 + (i0*nk0 + ik0)]);
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-        }
-
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // total patches in dst
-    const int np = ne2;
-
-    // patches per thread
-    const int dp = (np + nth - 1)/nth;
-
-    // patch range for this thread
-    const int ip0 = dp*ith;
-    const int ip1 = MIN(ip0 + dp, np);
-
-    ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ip0; i2 < ip1; i2++) {
-            float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2);
-
-            for (int i1 = 0; i1 < ne1; ++i1) {
-                for (int i0 = 0; i0 < ne0; ++i0) {
-                    ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0,
-                            (ggml_fp16_t *) ((char *) src0->data + i2*nb03),
-                            (ggml_fp16_t *)                wdata + i3*nb3 + (i1*ne0 + i0)*ew0);
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_conv_2d_sk_p0(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_conv_2d_sk_p0_f16_f32(params, src0, src1, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                //ggml_compute_forward_conv_2d_sk_p0_f32(params, src0, src1, dst);
-                GGML_ASSERT(false);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-// ggml_compute_forward_conv_2d_any
-
-static void ggml_compute_forward_conv_2d_f16_f32(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
-              struct ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
     const int32_t s0 = ((const int32_t*)(opt0->data))[0];
     const int32_t s1 = ((const int32_t*)(opt0->data))[1];
     const int32_t p0 = ((const int32_t*)(opt0->data))[2];
     const int32_t p1 = ((const int32_t*)(opt0->data))[3];
     const int32_t d0 = ((const int32_t*)(opt0->data))[4];
     const int32_t d1 = ((const int32_t*)(opt0->data))[5];
-    GGML_TENSOR_BINARY_OP_LOCALS;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nk0 = ne00;
-    const int nk1 = ne01;
-
-    const int ew0 = nk0*nk1*ne02;
 
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
@@ -13150,12 +13043,12 @@ static void ggml_compute_forward_conv_2d_f16_f32(
                     for (int i0 = 0; i0 < ne0; i0++) {
                         for (int ik1 = 0; ik1 < nk1; ik1++) {
                             for (int ik0 = 0; ik0 < nk0; ik0++) {
-                                const int idx1 = i1*s1 + ik1*d1 - p1;
                                 const int idx0 = i0*s0 + ik0*d0 - p0;
-                                
+                                const int idx1 = i1*s1 + ik1*d1 - p1;
+
                                 if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) {
                                     dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
-                                        GGML_FP32_TO_FP16(src[idx1 * ne10 + idx0]);
+                                        GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]);
                                 }
                             }
                         }
@@ -13183,27 +13076,28 @@ static void ggml_compute_forward_conv_2d_f16_f32(
 
     ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-    for (int i2 = ip0; i2 < ip1; i2++) {
-        float * dst_data = (float *)((char *) dst->data + i2*nb2);
+    for (int i3 = 0; i3 < ne3; i3++) {
+        for (int i2 = ip0; i2 < ip1; i2++) {
+            float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2);
 
-        for (int i1 = 0; i1 < ne1; ++i1) {
-            for (int i0 = 0; i0 < ne0; ++i0) {
-                ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0,
-                        (ggml_fp16_t *) ((char *) src0->data + i2*nb03),
-                        (ggml_fp16_t *)                wdata + (i1*ne0 + i0)*ew0);
+            for (int i1 = 0; i1 < ne1; ++i1) {
+                for (int i0 = 0; i0 < ne0; ++i0) {
+                    ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0,
+                            (ggml_fp16_t *) ((char *) src0->data + i2*nb03),
+                            (ggml_fp16_t *)                wdata + i3*nb3 + (i1*ne0 + i0)*ew0);
+                }
             }
-
         }
     }
 }
 
-static void ggml_compute_forward_conv_2d_any(
+static void ggml_compute_forward_conv_2d(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         const struct ggml_tensor * opt0,
         struct ggml_tensor * dst
-    ) {
+        ) {
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
@@ -13219,30 +13113,6 @@ static void ggml_compute_forward_conv_2d_any(
                 GGML_ASSERT(false);
             } break;
     }
-
-}
-
-// ggml_compute_forward_conv_2d
-
-static void ggml_compute_forward_conv_2d(
-    const struct ggml_compute_params* params,
-    const struct ggml_tensor* src0,
-    const struct ggml_tensor* src1,
-    const struct ggml_tensor* opt0,
-    struct ggml_tensor* dst) {
-    const int32_t s0 = ((const int32_t*)(opt0->data))[0];
-    const int32_t s1 = ((const int32_t*)(opt0->data))[1];
-    const int32_t p0 = ((const int32_t*)(opt0->data))[2];
-    const int32_t p1 = ((const int32_t*)(opt0->data))[3];
-    const int32_t d0 = ((const int32_t*)(opt0->data))[4];
-    const int32_t d1 = ((const int32_t*)(opt0->data))[5];
-
-    if (s0 == src0->ne[0] && s1 == src0->ne[1] && p0 == 0 && p1 == 0) {
-        ggml_compute_forward_conv_2d_sk_p0(params, src0, src1, dst);
-    }
-    else {
-        ggml_compute_forward_conv_2d_any(params, src0, src1, opt0, dst);
-    };
 }
 
 // ggml_compute_forward_pool_1d_sk_p0