whisper.rn 0.4.1 → 0.4.3

package/cpp/ggml-cpu/ops.h

@@ -20,6 +20,9 @@
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
+// Work buffer size for im2col operations in CONV2D
+#define WSP_GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -53,6 +56,7 @@ void wsp_ggml_compute_forward_permute(const struct wsp_ggml_compute_params * par
 void wsp_ggml_compute_forward_transpose(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_get_rows(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_get_rows_back(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_set_rows(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_diag(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_diag_mask_inf(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_diag_mask_zero(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
@@ -64,6 +68,7 @@ void wsp_ggml_compute_forward_clamp(const struct wsp_ggml_compute_params * param
 void wsp_ggml_compute_forward_conv_transpose_1d(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_im2col(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_im2col_back_f32(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_conv_2d(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_conv_transpose_2d(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_conv_2d_dw(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_pool_1d(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
@@ -93,6 +98,7 @@ void wsp_ggml_compute_forward_ssm_scan(const struct wsp_ggml_compute_params * pa
 void wsp_ggml_compute_forward_win_part(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_win_unpart(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_unary(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_glu(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_get_rel_pos(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_add_rel_pos(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_rwkv_wkv6(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
@@ -105,6 +111,7 @@ void wsp_ggml_compute_forward_custom(const struct wsp_ggml_compute_params * para
 void wsp_ggml_compute_forward_cross_entropy_loss(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_cross_entropy_loss_back(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_opt_step_adamw(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_mul_mat(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 
 #ifdef __cplusplus
 }

package/cpp/ggml-cpu/quants.c

@@ -2,6 +2,7 @@
 #include "ggml-common.h"
 
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "ggml-quants.h"
 #include "quants.h"
 
@@ -137,7 +138,7 @@ void wsp_ggml_vec_dot_q4_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*WSP_GGML_FP16_TO_FP32(x[ib].d)*WSP_GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*WSP_GGML_CPU_FP16_TO_FP32(x[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -174,7 +175,7 @@ void wsp_ggml_vec_dot_q4_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (WSP_GGML_FP16_TO_FP32(x[ib].d)*WSP_GGML_FP16_TO_FP32(y[ib].d))*sumi + WSP_GGML_FP16_TO_FP32(x[ib].m)*WSP_GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (WSP_GGML_CPU_FP16_TO_FP32(x[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + WSP_GGML_CPU_FP16_TO_FP32(x[ib].m)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -217,7 +218,7 @@ void wsp_ggml_vec_dot_q5_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (WSP_GGML_FP16_TO_FP32(x[ib].d)*WSP_GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (WSP_GGML_CPU_FP16_TO_FP32(x[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -260,7 +261,7 @@ void wsp_ggml_vec_dot_q5_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (WSP_GGML_FP16_TO_FP32(x[ib].d)*WSP_GGML_FP16_TO_FP32(y[ib].d))*sumi + WSP_GGML_FP16_TO_FP32(x[ib].m)*WSP_GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (WSP_GGML_CPU_FP16_TO_FP32(x[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + WSP_GGML_CPU_FP16_TO_FP32(x[ib].m)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -290,7 +291,7 @@ void wsp_ggml_vec_dot_q8_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(WSP_GGML_FP16_TO_FP32(x[ib].d)*WSP_GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(WSP_GGML_CPU_FP16_TO_FP32(x[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -342,7 +343,7 @@ void wsp_ggml_vec_dot_tq1_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
             }
         }
 
-        sumf += (float) sum * (WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -372,7 +373,7 @@ void wsp_ggml_vec_dot_tq2_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
             }
         }
 
-        const float d = y[i].d * WSP_GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * WSP_GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -405,8 +406,8 @@ void wsp_ggml_vec_dot_q2_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * WSP_GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * WSP_GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * WSP_GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * WSP_GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -504,7 +505,7 @@ void wsp_ggml_vec_dot_q3_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -577,9 +578,9 @@ void wsp_ggml_vec_dot_q4_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = WSP_GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = WSP_GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -657,9 +658,9 @@ void wsp_ggml_vec_dot_q5_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = WSP_GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = WSP_GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -714,7 +715,7 @@ void wsp_ggml_vec_dot_q6_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -739,7 +740,7 @@ void wsp_ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, s
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * WSP_GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * WSP_GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -778,7 +779,7 @@ void wsp_ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, si
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * WSP_GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * WSP_GGML_RESTRICT sc = x[i].scales;
         const int8_t   * WSP_GGML_RESTRICT q8 = y[i].qs;
@@ -829,7 +830,7 @@ void wsp_ggml_vec_dot_iq2_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -882,7 +883,7 @@ void wsp_ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, s
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * WSP_GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * WSP_GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * WSP_GGML_RESTRICT q8 = y[i].qs;
@@ -924,7 +925,7 @@ void wsp_ggml_vec_dot_iq3_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * WSP_GGML_RESTRICT qs = x[i].qs;
         const uint8_t * WSP_GGML_RESTRICT qh = x[i].qh;
         const uint8_t * WSP_GGML_RESTRICT signs = x[i].signs;
@@ -1002,7 +1003,7 @@ void wsp_ggml_vec_dot_iq1_s_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
             qs += 4;
         }
 
-        sumf += WSP_GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += WSP_GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -1063,7 +1064,7 @@ void wsp_ggml_vec_dot_iq1_m_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
             qh += 2;
         }
 
-        sumf += WSP_GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += WSP_GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -1087,7 +1088,7 @@ void wsp_ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, si
     float sumf = 0;
 
     for (; ib < nb; ++ib) {
-        const float d = WSP_GGML_FP16_TO_FP32(y[ib].d)*WSP_GGML_FP16_TO_FP32(x[ib].d);
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(y[ib].d)*WSP_GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1113,7 +1114,7 @@ void wsp_ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, si
 
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = WSP_GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = WSP_GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

package/cpp/ggml-cpu/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include "arch-fallback.h"
@@ -72,7 +73,7 @@ void wsp_ggml_wsp_quantize_mat_q8_0_4x4_generic(const float * WSP_GGML_RESTRICT
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = WSP_GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = WSP_GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -110,7 +111,7 @@ void wsp_ggml_wsp_quantize_mat_q8_0_4x8_generic(const float * WSP_GGML_RESTRICT
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = WSP_GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = WSP_GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -236,7 +237,7 @@ void wsp_ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -280,7 +281,7 @@ void wsp_ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -325,7 +326,7 @@ void wsp_ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -396,13 +397,13 @@ void wsp_ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
                     }
-                    sumf[j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                 }
             }
             for (int sb = 0; sb < 8; sb++) {
                 uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * WSP_GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -449,7 +450,7 @@ void wsp_ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, s
                             const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                         }
-                        sumf[j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -500,7 +501,7 @@ void wsp_ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                                 }
-                                sumf[m][j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }
@@ -555,7 +556,7 @@ void wsp_ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -609,7 +610,7 @@ void wsp_ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, siz
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -688,7 +689,7 @@ void wsp_ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
                                 sumi2 = sumi2 * scales_1[j];
                                 sumi += sumi1 + sumi2;
                             }
-                            sumf[m][j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -697,7 +698,7 @@ void wsp_ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, siz
                     for(int m = 0; m < 4; m++) {
                         const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                         for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * WSP_GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -753,7 +754,7 @@ void wsp_ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, s
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                                 }
-                                sumf[m][j] += sumi * WSP_GGML_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * WSP_GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * WSP_GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }