diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h index 968ce2888..88d9bbac1 100644 --- a/ggml/src/ggml-cpu/simd-mappings.h +++ b/ggml/src/ggml-cpu/simd-mappings.h @@ -248,43 +248,6 @@ #define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE #endif -#if !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) -#ifdef GGML_FP16_TO_FP32 -#undef GGML_FP16_TO_FP32 -#endif - -#ifdef GGML_FP32_TO_FP16 -#undef GGML_FP32_TO_FP16 -#endif - -#ifdef GGML_COMPUTE_FP16_TO_FP32 -#undef GGML_COMPUTE_FP16_TO_FP32 -#endif - -#ifdef GGML_COMPUTE_FP32_TO_FP16 -#undef GGML_COMPUTE_FP32_TO_FP16 -#endif - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - -#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - __fp16 tmp; - memcpy(&tmp, &h, sizeof(ggml_fp16_t)); - return (float)tmp; -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - ggml_fp16_t res; - __fp16 tmp = f; - memcpy(&res, &tmp, sizeof(ggml_fp16_t)); - return res; -} - -#endif // !__CUDACC__ && __CUDACC_VER_MAJOR__ <= 11 && !__MUSACC__ - #elif defined(__AVX512F__) #define GGML_SIMD @@ -447,23 +410,6 @@ do { \ // the _mm256_cvt intrinsics require F16C #define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x))) #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0)) - -#ifdef GGML_COMPUTE_FP16_TO_FP32 -#undef GGML_COMPUTE_FP16_TO_FP32 -#endif - -#ifdef GGML_COMPUTE_FP32_TO_FP16 -#undef GGML_COMPUTE_FP32_TO_FP16 -#endif - -#ifdef _MSC_VER - #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) - #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) -#else - #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) -#endif - #else static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) { float tmp[8]; @@ -573,53 +519,6 @@ static inline unsigned char ggml_endian_byte(int i) { r[i - GGML_ENDIAN_BYTE(0)]), \ 0, p - GGML_F16_EPR) -#ifdef GGML_FP16_TO_FP32 -#undef GGML_FP16_TO_FP32 -#endif - -#ifdef GGML_FP32_TO_FP16 -#undef GGML_FP32_TO_FP16 -#endif - -#ifdef GGML_COMPUTE_FP16_TO_FP32 -#undef GGML_COMPUTE_FP16_TO_FP32 -#endif - -#ifdef GGML_COMPUTE_FP32_TO_FP16 -#undef GGML_COMPUTE_FP32_TO_FP16 -#endif - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -/* the inline asm below is about 12% faster than the lookup method */ -#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - float f; - double d; - __asm__( - "mtfprd %0,%2\n" - "xscvhpdp %0,%0\n" - "frsp %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=f"(f): - /* in */ "r"(h)); - return f; -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - double d; - ggml_fp16_t r; - __asm__( /* xscvdphp can work on double or single precision */ - "xscvdphp %0,%2\n" - "mffprd %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=r"(r): - /* in */ "f"(f)); - return r; -} - #elif defined(__wasm_simd128__) #define GGML_SIMD @@ -1153,35 +1052,6 @@ static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) { #endif // __NNPA__ -#elif defined(__riscv) && defined(__riscv_zfhmin) - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - float f; - __asm__( - "fmv.h.x %[f], %[h]\n\t" - "fcvt.s.h %[f], %[f]" - : [f] "=&f" (f) - : [h] "r" (h) - ); - return f; -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - ggml_fp16_t res; - __asm__( - "fcvt.h.s %[f], %[f]\n\t" - "fmv.x.h %[h], %[f]" - : [h] "=&r" (res) - : [f] "f" (f) - ); - return res; -} - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - #endif // GGML_F32_ARR / GGML_F16_ARR diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index 5c276922f..6dc5ce0d9 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -317,78 +317,182 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1); GGML_API void * ggml_aligned_malloc(size_t size); GGML_API void ggml_aligned_free(void * ptr, size_t size); -// FP16 <-> FP32 -// ref: https://github.com/Maratyszcza/FP16 +// FP16 to FP32 conversion -static inline float fp32_from_bits(uint32_t w) { - union { - uint32_t as_bits; - float as_value; - } fp32; - fp32.as_bits = w; - return fp32.as_value; -} +// 16-bit float +// on Arm, we use __fp16 +// on x86, we use uint16_t +// +// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616 +// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843 +// +#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) + #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -static inline uint32_t fp32_to_bits(float f) { - union { - float as_value; - uint32_t as_bits; - } fp32; - fp32.as_value = f; - return fp32.as_bits; -} + #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - const uint32_t w = (uint32_t) h << 16; - const uint32_t sign = w & UINT32_C(0x80000000); - const uint32_t two_w = w + w; - - const uint32_t exp_offset = UINT32_C(0xE0) << 23; -#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) - const float exp_scale = 0x1.0p-112f; -#else - const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); -#endif - const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; - - const uint32_t magic_mask = UINT32_C(126) << 23; - const float magic_bias = 0.5f; - const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; - - const uint32_t denormalized_cutoff = UINT32_C(1) << 27; - const uint32_t result = sign | - (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); - return fp32_from_bits(result); -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { -#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) - const float scale_to_inf = 0x1.0p+112f; - const float scale_to_zero = 0x1.0p-110f; -#else - const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); - const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); -#endif - float base = (fabsf(f) * scale_to_inf) * scale_to_zero; - - const uint32_t w = fp32_to_bits(f); - const uint32_t shl1_w = w + w; - const uint32_t sign = w & UINT32_C(0x80000000); - uint32_t bias = shl1_w & UINT32_C(0xFF000000); - if (bias < UINT32_C(0x71000000)) { - bias = UINT32_C(0x71000000); + static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + __fp16 tmp; + memcpy(&tmp, &h, sizeof(ggml_fp16_t)); + return (float)tmp; } - base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; - const uint32_t bits = fp32_to_bits(base); - const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); - const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); - const uint32_t nonsign = exp_bits + mantissa_bits; - return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); -} + static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + __fp16 tmp = f; + memcpy(&res, &tmp, sizeof(ggml_fp16_t)); + return res; + } -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) +#elif defined(__F16C__) + + #ifdef _MSC_VER + #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) + #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) + #else + #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) + #endif + +#elif defined(__POWER9_VECTOR__) + + #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + /* the inline asm below is about 12% faster than the lookup method */ + #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) + #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + + static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + float f; + double d; + __asm__( + "mtfprd %0,%2\n" + "xscvhpdp %0,%0\n" + "frsp %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=f"(f): + /* in */ "r"(h)); + return f; + } + + static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + double d; + ggml_fp16_t r; + __asm__( /* xscvdphp can work on double or single precision */ + "xscvdphp %0,%2\n" + "mffprd %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=r"(r): + /* in */ "f"(f)); + return r; + } + +#elif defined(__riscv) && defined(__riscv_zfhmin) + + static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + float f; + __asm__( + "fmv.h.x %[f], %[h]\n\t" + "fcvt.s.h %[f], %[f]" + : [f] "=&f" (f) + : [h] "r" (h) + ); + return f; + } + + static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + __asm__( + "fcvt.h.s %[f], %[f]\n\t" + "fmv.x.h %[h], %[f]" + : [h] "=&r" (res) + : [f] "f" (f) + ); + return res; + } + + #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) + #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + +#else + + // FP16 <-> FP32 + // ref: https://github.com/Maratyszcza/FP16 + + static inline float fp32_from_bits(uint32_t w) { + union { + uint32_t as_bits; + float as_value; + } fp32; + fp32.as_bits = w; + return fp32.as_value; + } + + static inline uint32_t fp32_to_bits(float f) { + union { + float as_value; + uint32_t as_bits; + } fp32; + fp32.as_value = f; + return fp32.as_bits; + } + + static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + const uint32_t w = (uint32_t) h << 16; + const uint32_t sign = w & UINT32_C(0x80000000); + const uint32_t two_w = w + w; + + const uint32_t exp_offset = UINT32_C(0xE0) << 23; + #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) + const float exp_scale = 0x1.0p-112f; + #else + const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); + #endif + const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; + + const uint32_t magic_mask = UINT32_C(126) << 23; + const float magic_bias = 0.5f; + const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; + + const uint32_t denormalized_cutoff = UINT32_C(1) << 27; + const uint32_t result = sign | + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); + return fp32_from_bits(result); + } + + static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) + const float scale_to_inf = 0x1.0p+112f; + const float scale_to_zero = 0x1.0p-110f; + #else + const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); + const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); + #endif + float base = (fabsf(f) * scale_to_inf) * scale_to_zero; + + const uint32_t w = fp32_to_bits(f); + const uint32_t shl1_w = w + w; + const uint32_t sign = w & UINT32_C(0x80000000); + uint32_t bias = shl1_w & UINT32_C(0xFF000000); + if (bias < UINT32_C(0x71000000)) { + bias = UINT32_C(0x71000000); + } + + base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; + const uint32_t bits = fp32_to_bits(base); + const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); + const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); + const uint32_t nonsign = exp_bits + mantissa_bits; + return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); + } + + #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + +#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) // precomputed f32 table for f16 (256 KB) // defined in ggml.c, initialized in ggml_init()