PowerPC:Fix for issue #12076

This patch adds cpu backend support and fix build failures on POWER
systems

1. Added `ggml-cpu/arch/powerpc` directory to host Power-specific code.
2. Fixed compilation errors in `sgemm.cpp` and simd-mappings.h where `vector` was not
   defined by explicitly mapping it to `__vector`.
3. Resolved build errors from `-mcpu=power11` by replacing it with
   the supported `-mcpu=power10`.

Signed-off-by: Shalini Salomi Bodapati <Shalini.Salomi.Bodapati@ibm.com>

ml/backend/ggml/ggml/src/CMakeLists.txt

@@ -359,7 +359,7 @@ if (GGML_CPU_ALL_VARIANTS)
             ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
             ggml_add_cpu_backend_variant(power9         POWER9  VSX)
             ggml_add_cpu_backend_variant(power10        POWER10 VSX)
-            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
+            ggml_add_cpu_backend_variant(power11        POWER10 VSX)
         else()
             message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
         endif()

ml/backend/ggml/ggml/src/ggml-cpu/arch-fallback.h

@@ -69,7 +69,6 @@
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

ml/backend/ggml/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp

@@ -0,0 +1,82 @@
+# include "ggml-backend-impl.h"
+
+#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#endif
+
+#include <string>
+
+struct powerpc_features {
+    std::string platform = "";
+    int power_version    = -1;
+
+    bool has_vsx         = false;
+
+    powerpc_features() {
+#if defined(__linux__)
+        unsigned long auxval = getauxval(AT_PLATFORM);
+        if (auxval) {
+            platform = std::string(reinterpret_cast<const char*>(auxval));
+            // TBD: Do systems exist that return this in uppercase?
+            if (platform.substr(0, 5) == "power") {
+                // Extractt a numeric suffix, if one exists
+                int vpos = -1;
+                for (int i = platform.length() - 1; i >= 0; i--) {
+                    if (std::isdigit(platform[i])) {
+                        vpos = i;
+                    } else {
+                        break;
+                    }
+                }
+                if (vpos > -1) {
+                    power_version = std::stoi(platform.substr(vpos));
+                }
+            }
+        }
+#endif
+        if (power_version >= 9) {
+            has_vsx = true;
+        }
+    }
+};
+
+static int ggml_backend_cpu_powerpc_score() {
+    int score = 1;
+    powerpc_features pf;
+
+// Platform scores
+#if defined(GGML_USE_POWER7)
+    if (pf.power_version < 7) { return 0; }
+    score += 1<<1;
+#endif
+#if defined(GGML_USE_POWER8)
+    if (pf.power_version < 8) { return 0; }
+    score += 1<<2;
+#endif
+#if defined(GGML_USE_POWER9)
+    if (pf.power_version < 9) { return 0; }
+    score += 1<<3;
+#endif
+#if defined(GGML_USE_POWER10)
+    if (pf.power_version < 10) { return 0; }
+    score += 1<<4;
+#endif
+#if defined(GGML_USE_POWER11)
+    if (pf.power_version < 11) { return 0; }
+    score += 1<<5;
+#endif
+
+// Feature scores
+#if defined(GGML_USE_VSX)
+    if (!pf.has_vsx) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
+
+#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)

ml/backend/ggml/ggml/src/ggml-cpu/arch/powerpc/powerpc.go

@@ -0,0 +1,5 @@
+package powerpc
+
+// #cgo CXXFLAGS: -std=c++17
+// #cgo CPPFLAGS: -I${SRCDIR}/../.. -I${SRCDIR}/../../.. -I${SRCDIR}/../../../../include
+import "C"

ml/backend/ggml/ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -117,6 +117,9 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
 #endif
 
 #if defined(__MMA__)
+#ifndef vector
+#define vector __vector
+#endif
 typedef vector unsigned char vec_t;
 typedef __vector_quad acc_t;
 #endif

ml/backend/ggml/ggml/src/ggml-cpu/simd-mappings.h

@@ -605,7 +605,6 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F16_VEC_REDUCE         GGML_F32Cx8_REDUCE
 
 #elif defined(__POWER9_VECTOR__)
-
 #define GGML_SIMD
 
 // F32 POWER9
@@ -613,7 +612,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_STEP 32
 #define GGML_F32_EPR  4
 
-#define GGML_F32x4              vector float
+#define GGML_F32x4              __vector float
 #define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)