File size: 3,916 Bytes
fb6457c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | #include "blocking.h"
#include "../scalar/primitives.h"
#include <stdlib.h>
#include <string.h>
block_config_t* blocking_config_create(size_t mc, size_t nc, size_t kc) {
block_config_t* cfg = (block_config_t*)malloc(sizeof(block_config_t));
if (cfg) {
cfg->mc = mc > 0 ? mc : MC;
cfg->nc = nc > 0 ? nc : NC;
cfg->kc = kc > 0 ? kc : KC;
}
return cfg;
}
void blocking_config_free(block_config_t* cfg) {
if (cfg) free(cfg);
}
void pack_A_f32(
float* packed,
const matrix_t* A,
size_t k_start, size_t k_end,
size_t m_start, size_t m_end
) {
if (!packed || !A) return;
size_t kc = k_end - k_start;
size_t mc = m_end - m_start;
const float* a_ptr = matrix_ptr((matrix_t*)A);
for (size_t i = 0; i < mc; ++i) {
for (size_t k = 0; k < kc; ++k) {
packed[i * kc + k] = a_ptr[(m_start + i) * A->stride + (k_start + k)];
}
}
}
void pack_B_f32(
float* packed,
const matrix_t* B,
size_t k_start, size_t k_end,
size_t n_start, size_t n_end
) {
if (!packed || !B) return;
size_t kc = k_end - k_start;
size_t nc = n_end - n_start;
const float* b_ptr = matrix_ptr((matrix_t*)B);
for (size_t k = 0; k < kc; ++k) {
for (size_t j = 0; j < nc; ++j) {
packed[k * nc + j] = b_ptr[(k_start + k) * B->stride + (n_start + j)];
}
}
}
void microkernel_f32(
float* C,
const float* packed_A,
const float* packed_B,
size_t mc, size_t nc, size_t kc,
size_t ldc
) {
if (!C || !packed_A || !packed_B) return;
for (size_t i = 0; i < mc; ++i) {
for (size_t j = 0; j < nc; ++j) {
float sum = 0.0f;
for (size_t k = 0; k < kc; ++k) {
sum = scalar_fma_f32(packed_A[i * kc + k], packed_B[k * nc + j], sum);
}
C[i * ldc + j] += sum;
}
}
}
void gemm_blocked_f32(
const matrix_t* A,
const matrix_t* B,
matrix_t* C,
float alpha,
float beta,
const block_config_t* cfg
) {
if (!A || !B || !C || !cfg) return;
if (A->cols != B->rows || A->rows != C->rows || B->cols != C->cols) return;
size_t M = A->rows;
size_t N = B->cols;
size_t K = A->cols;
size_t mc = cfg->mc;
size_t nc = cfg->nc;
size_t kc = cfg->kc;
float* packed_A = (float*)malloc(mc * kc * sizeof(float));
float* packed_B = (float*)malloc(kc * nc * sizeof(float));
if (!packed_A || !packed_B) {
free(packed_A);
free(packed_B);
return;
}
float* c_ptr = matrix_ptr(C);
for (size_t i_block = 0; i_block < M; i_block += mc) {
size_t i_end = (i_block + mc < M) ? i_block + mc : M;
size_t i_size = i_end - i_block;
for (size_t j_block = 0; j_block < N; j_block += nc) {
size_t j_end = (j_block + nc < N) ? j_block + nc : N;
size_t j_size = j_end - j_block;
for (size_t k_block = 0; k_block < K; k_block += kc) {
size_t k_end = (k_block + kc < K) ? k_block + kc : K;
size_t k_size = k_end - k_block;
pack_A_f32(packed_A, A, k_block, k_end, i_block, i_end);
pack_B_f32(packed_B, B, k_block, k_end, j_block, j_end);
microkernel_f32(
c_ptr + i_block * C->stride + j_block,
packed_A, packed_B,
i_size, j_size, k_size,
C->stride
);
}
}
}
free(packed_A);
free(packed_B);
if (alpha != 1.0f || beta != 0.0f) {
for (size_t i = 0; i < M; ++i) {
for (size_t j = 0; j < N; ++j) {
c_ptr[i * C->stride + j] = scalar_fma_f32(alpha, c_ptr[i * C->stride + j],
scalar_mul_f32(beta, c_ptr[i * C->stride + j]));
}
}
}
}
|