diff --git a/Makefile b/Makefile index ae7189280..cffc6601f 100644 --- a/Makefile +++ b/Makefile @@ -744,6 +744,9 @@ clean: # Helper function that replaces .c, .cpp, and .cu file endings with .o: GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1)))) +# Helper function that replaces .c, .cpp, and .cu file endings with .s: +GET_ASM_FILE = $(patsubst %.c,%.s,$(patsubst %.cpp,%.s,$(patsubst %.cu,%.s,$(1)))) + main: examples/main/main.cpp ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -751,6 +754,16 @@ main: examples/main/main.cpp ggml.o llama.o $(C @echo '==== Run ./main -h for help. ====' @echo +bench-phi-knc.s: bench-phi-knc.c + $(CC) $(CFLAGS) -S $< -o $(call GET_ASM_FILE, $<) + +ggml-phi-knc.s: ggml-phi-knc.c + $(CC) $(CFLAGS) -S $< -o $(call GET_ASM_FILE, $<) + +bench-phi-knc: bench-phi-knc.c ggml-phi-knc.o + $(CC) $(CFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CC) $(CFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + infill: examples/infill/infill.cpp ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) diff --git a/ggml-phi-knc.c b/ggml-phi-knc.c index ff94104a7..648f81bcf 100644 --- a/ggml-phi-knc.c +++ b/ggml-phi-knc.c @@ -4,9 +4,6 @@ #include -static inline _Bool is_aligned(const void *restrict pointer, size_t byte_count) -{ return (uintptr_t)pointer % byte_count == 0; } - // No, we have an SIMD unit. // #define GGML_SIMD @@ -15,102 +12,135 @@ static inline _Bool is_aligned(const void *restrict pointer, size_t byte_count) // We can fit 16 of these float32s in a single vector register. #define GGML_F32_EPR 16 -// because we are not defining GGML_SIMD, we have to do this ourself. -#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR) - // a single vector. 128*32=512 typedef float float32x16_t __attribute__((vector_size (128))); #define GGML_F32x16 float32x16_t -// from chatGPT. nuke this later. -#include +// A forward declaration, to keep GCC happy... +void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); inline static void GGML_F32x16_VEC_ZERO(float32x16_t *target) { - // we only need a mask16, but register sizes... - __mmask32 mask=0xFFFFFFFF; - - // FIXME: how do we tell GNU AS to perform upconverts? + // FIXME: how do we tell GNU AS to perform upconverts? Could remove two memory reads here... float zero[4] __attribute__((aligned(64))) = {0.0f,0.0f,0.0f,0.0f}; - __asm__ __volatile__ ("movl\t%[M],\t%%eax\n\t" - "kmov %%eax,\t%%k1\n\t" - "vbroadcastf32x4\t%[Z],\t%%zmm0%{%%k1%}\n\t" - "vmovaps\t\t%%zmm0,\t%[RES]%{%%k1%}\n\t" - : [RES] "+m" (*target) - : [M] "m" (mask), - [Z] "m" (zero) - : "eax", "k1", "zmm0"); -} - -// multiply each item in mvec1 with the corresponding item in mvec2, adding the result to the corresponding item in sum. -inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x16_t *mvec2, float32x16_t *sumvec, size_t iterations) -{ - // we only need a mask16, but register sizes... - __mmask32 mask=0xFFFFFFFF; __asm__ __volatile__ ( - "vmovaps\t\t(%[RES]),\t%%zmm0\n\t" // load our initial state.. - "1:\n\t" - "cmp $0,\t%[ITER]\n\t" // Compare iterations to 0 - "je\t2f\n\t" // Jump to label 2 if zero (end of loop) - "vmovaps\t\t(%[VEC1]),\t%%zmm1\n\t" // Load two vectors. - "vmovaps\t\t(%[VEC2]),\t%%zmm2\n\t" - "vfmadd231ps\t%%zmm1,\t%%zmm2,\t%%zmm0\n\t" // Perform a fused multiply add. - "add $64,\t%[VEC1]\n\t" // Move to the next float32x16_t (64 bytes ahead) - "add $64,\t%[VEC2]\n\t" - "sub $1,\t%[ITER]\n\t" // Decrement iterations - "jmp 1b\n\t" // Jump back to the start of the loop - "2: \n\t" // Label for loop end - "vmovaps\t\t%%zmm0,\t(%[RES])\n\t" // save our results. - : [RES] "+r" (sumvec), - [ITER] "+r" (iterations) - : [M] "r" (mask), - [VEC1] "r" (mvec1), - [VEC2] "r" (mvec2) - : "zmm0", "zmm1", "zmm2", "cc", "memory"); + "vbroadcastf32x4\t%[Z],\t%%zmm8\n\t" // use an upscaling operator to clear our value. + "vmovaps\t\t%%zmm8,\t%[RES]\n\t" + : [RES] "+m" (*target) + : [Z] "m" (zero) + : "zmm8"); } - -// NOTE: all inputs must be __attribute__((aligned(64))); -float DotProduct_F32(const float * restrict inVec1, const float * restrict inVec2, uint32_t count) +// Multiply each item in mvec1 with the corresponding item in mvec2, adding the result to the corresponding item in sum. optionally clear the sum before starting. +inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x16_t *mvec2, float32x16_t *sumvec, size_t iterations, int clear) { - // our single result, in the end. - float sumf = 0.0f; + // FIXME: how do we tell GNU AS to perform upconverts? Could remove two memory reads here... + float zero[4] __attribute__((aligned(64))) = {0.0f,0.0f,0.0f,0.0f}; + __asm__ __volatile__ ( + "mov\t%[ITER],%%r8\n\t" // how many register sized chunks are we responsible for + "mov\t%[VEC1],%%r10\n\t" // where do we start work in mvec1? + "mov\t%[VEC2],%%r12\n\t" // where do we start work in mvec2? + "cmp\t$1,%[CLR]\n\t" // should we clear the sum before we start? + "jne\t4f\n\t" + "vbroadcastf32x4\t%[Z],\t%%zmm0\n\t" // if so, use an upscaling operator to do it. + "vprefetchnta\t(%%r10)\n\t" + "vprefetchnta\t(%%r12)\n\t" + "vprefetch1\t128(%%r10)\n\t" + "vprefetch1\t128(%%r12)\n\t" + "vprefetch1\t256(%%r10)\n\t" + "vprefetch1\t256(%%r12)\n\t" + "vprefetch1\t384(%%r10)\n\t" + "vprefetch1\t384(%%r12)\n\t" + "vprefetch1\t512(%%r10)\n\t" + "vprefetch1\t512(%%r12)\n\t" + "jmp\t1f\n\t" + "4:\n\t" + "vprefetch0\t(%[RES])\n\t" + "vmovaps\t\t(%[RES]),\t%%zmm0\n\t" // otherwise, load our inital state from sum.. + "vprefetchnta\t(%%r10)\n\t" + "vprefetchnta\t(%%r12)\n\t" + "1:\n\t" + "cmp\t$3,\t%%r8\n\t" // Compare iterations to three. + "jnae\t6f\n\t" // If there are not three iterations left, jump to label 6. + "vmovaps\t\t(%%r10),\t%%zmm1\n\t" // Load two vectors. + "vmovaps\t\t(%%r12),\t%%zmm2\n\t" + "sub\t$3,\t%%r8\n\t" // Decrement iterations + "vprefetchnta\t192(%%r10)\n\t" // prefetch the next float32x16_t block (192 bytes ahead) + "vprefetchnta\t192(%%r12)\n\t" + "vmovaps\t\t64(%%r10),\t%%zmm3\n\t" // Load two vectors. + "vmovaps\t\t64(%%r12),\t%%zmm4\n\t" + "vprefetch1\t320(%%r10)\n\t" // prefetch the block after the block after the next float32x16_t block (320 bytes ahead) + "vprefetch1\t320(%%r12)\n\t" + "vmovaps\t\t128(%%r10),\t%%zmm5\n\t" // Load two vectors. + "vmovaps\t\t128(%%r12),\t%%zmm6\n\t" + "vprefetch1\t576(%%r10)\n\t" + "vprefetch1\t576(%%r12)\n\t" + "vprefetch1\t704(%%r10)\n\t" + "vprefetch1\t704(%%r12)\n\t" + "add\t$192,\t%%r10\n\t" // Move to the next float32x16_t block (192 bytes ahead) + "add\t$192,\t%%r12\n\t" + "vfmadd231ps\t%%zmm1,\t%%zmm2,\t%%zmm0\n\t" // Perform a fused multiply add + "vfmadd231ps\t%%zmm3,\t%%zmm4,\t%%zmm0\n\t" // Perform a fused multiply add + "vfmadd231ps\t%%zmm5,\t%%zmm6,\t%%zmm0\n\t" // Perform a fused multiply add + "jmp\t1b\n\t" // Jump back to the start of the loop + "6:\n\t" // we know we are near the tail. handle 2, 1, and 0 cases. + "cmp\t$0,\t%%r8\n\t" // Compare iterations to zero + "je\t2f\n\t" // Jump to label 2 if zero (end of loop) + "cmp\t$1,\t%%r8\n\t" // Compare iterations to one + "vmovaps\t\t(%%r10),\t%%zmm1\n\t" // Load two vectors. + "vmovaps\t\t(%%r12),\t%%zmm2\n\t" + "vfmadd231ps\t%%zmm1,\t%%zmm2,\t%%zmm0\n\t" // Perform a fused multiply add + "je\t2f\n\t" // Jump to label 3 if one (end of loop) + // No compare. we must be two. + "vmovaps\t\t64(%%r10),\t%%zmm3\n\t" // Load two vectors. + "vmovaps\t\t64(%%r12),\t%%zmm4\n\t" + "vfmadd231ps\t%%zmm3,\t%%zmm4,\t%%zmm0\n\t" // Perform a fused multiply add + "2:\n\t" // Label for loop end + "vmovaps\t\t%%zmm0,\t(%[RES])\n\t" // save our results. + : [RES] "+r" (sumvec) + : [ITER] "r" (iterations), + [VEC1] "r" (mvec1), + [VEC2] "r" (mvec2), + [CLR] "r" (clear), + [Z] "m" (zero) + : "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6", "cc", "memory", "r8", "r10", "r12"); +} + +// NOTE: x and y inputs must be __attribute__((aligned(64))); +void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) +{ // our sum. float32x16_t sum __attribute__((aligned(64))); // the number of vector-sized steps we will need to do. - const uint32_t np = (count & ~(GGML_F32_EPR - 1)); + const uint32_t np = (n & ~(GGML_F32_EPR - 1)); - GGML_F32x16_VEC_ZERO(&sum); + GGML_F32x16_VEC_FMA((const float32x16_t *)x, (const float32x16_t *)y, &sum, np/GGML_F32_EPR, 1); - // 0 indexed cycle count - // for (uint32_t cycle = 0; cycle < (np/GGML_F32_EPR); ++cycle) - GGML_F32x16_VEC_FMA((float32x16_t *)inVec1, (float32x16_t *)inVec2, &sum, np/GGML_F32_EPR); - - if (count != np) + // FIXME: replace this with a final round using masked vectors. + if ( n - np != 0 ) { - printf("handling remainder %u\n",count-np); // add the leftovers, that could not be handled by the vector loop. - // our extended last part of inVec1. + // our extended last part of x. float32x16_t v1 __attribute__((aligned(64))); GGML_F32x16_VEC_ZERO(&v1); - // our extended last part of inVec2. + // our extended last part of y. float32x16_t v2 __attribute__((aligned(64))); GGML_F32x16_VEC_ZERO(&v2); - memcpy(&v1, &inVec1[np], (count - np)*sizeof(float)); - memcpy(&v2, &inVec2[np], (count - np)*sizeof(float)); + memcpy(&v1, &x[np], (n - np)*sizeof(float)); + memcpy(&v2, &y[np], (n - np)*sizeof(float)); GGML_F32x16_VEC_FMA(&v1, - &v2, - &sum, 1); + &v2, + &sum, 1, 0); + } - // reduce sum0..sumX to sumf + // reduce sum, and store it in s. for (uint32_t i=0; i - -GGML_CALL float DotProduct_F32(const float * restrict vec1, const float * restrict vec2, uint32_t count); +void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); #ifdef __cplusplus } #endif - diff --git a/ggml.c b/ggml.c index 66ed13b5e..872858d0c 100644 --- a/ggml.c +++ b/ggml.c @@ -42,6 +42,7 @@ #pragma warning(disable: 4996) #endif +// hand assembled replacement functions are cool. #if defined(__k1om__) #include #endif @@ -500,7 +501,11 @@ FILE * ggml_fopen(const char * fname, const char * mode) { static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); +#if defined(__k1om__) +// We get this function from elsewhere. +#else static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); +#endif static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc); static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { @@ -1502,6 +1507,9 @@ inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; } inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; } +#if defined(__k1om__) +// we get this function from elsewhere. +#else static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); @@ -1534,8 +1542,6 @@ static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * for (int i = np; i < n; ++i) { sumf += x[i]*y[i]; } -#elif defined(__k1om__) - float sumf = DotProduct_F32(x, y, n); #else // scalar ggml_float sumf = 0.0; @@ -1546,6 +1552,7 @@ static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * *s = sumf; } +#endif static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) { assert(nrc == 1);