diff --git a/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c b/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c index 05ba015d2..d604235c9 100644 --- a/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c +++ b/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c @@ -31,8 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define VMOVLDUP(addr, zmm) asm("vmovsldup (%1), %0": "=v"(zmm): "r"(addr)) #define VMOVHDUP(addr, zmm) asm("vmovshdup (%1), %0": "=v"(zmm): "r"(addr)) #define BROADCAST64(base, step, n, offset, zmm) \ - if (n == 0) asm("vbroadcastsd %2(%1), %0": "=v"(zmm): "r"(base), "n"(offset*2)); \ - else asm("vbroadcastsd %4(%1, %2, %3), %0": "=v"(zmm): "r"(base), "r"(step), "n"(n*2), "n"(offset*2)) + if (n == 0) asm("vbroadcastsd %c2(%1), %0": "=v"(zmm): "r"(base), "n"(offset*2)); \ + else asm("vbroadcastsd %c4(%1, %2, %c3), %0": "=v"(zmm): "r"(base), "r"(step), "n"(n*2), "n"(offset*2)) #define DECLARE_A_PAIR(A) \ __m512i A_lo_##A; __m512i A_hi_##A; @@ -41,8 +41,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. VMOVLDUP(ptr_a##A, A_lo_##A); \ VMOVHDUP(ptr_a##A, A_hi_##A); +#define MASK_LOAD_A_PAIR(A) { \ + __m512 tmp = _mm512_maskz_loadu_ps(mmask, ptr_a##A); \ + A_lo_##A = (__m512i) _mm512_moveldup_ps(tmp); \ + A_hi_##A = (__m512i) _mm512_movehdup_ps(tmp); \ +} + #define LOAD_A_PAIR_TAIL(A) { \ - __m256i ymm = _mm256_loadu_si256(ptr_a##A); \ + __m256i ymm = _mm256_loadu_si256((void *)ptr_a##A); \ + __m512 zmm = (__m512) _mm512_cvtepu16_epi32(ymm); \ + A_lo_##A = (__m512i) _mm512_moveldup_ps(zmm); \ + A_hi_##A = (__m512i) _mm512_movehdup_ps(zmm); \ +} + +#define MASK_LOAD_A_PAIR_TAIL(A) { \ + __m256i ymm = _mm256_maskz_loadu_epi16(mmask, ptr_a##A); \ __m512 zmm = (__m512) _mm512_cvtepu16_epi32(ymm); \ A_lo_##A = (__m512i) _mm512_moveldup_ps(zmm); \ A_hi_##A = (__m512i) _mm512_movehdup_ps(zmm); \ @@ -53,13 +66,26 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define BROADCAST_B_PAIR(Bx, By) \ BROADCAST64(ptr_b##Bx, n_blksize, By, 0, B_lo); \ - BROADCAST64(ptr_b##Bx, n_blksize, By, 2, B_hi); + BROADCAST64(ptr_b##Bx, n_blksize, By, 4, B_hi); + +#define MASK_BROADCAST_B_PAIR(Bx, x) {\ + __m128 xmm = _mm_maskz_loadu_ps(nmask, ptr_b##Bx); \ + B_lo = (__m512i) _mm512_broadcastsd_pd((__m128d) xmm); \ + B_hi = (__m512i) _mm512_broadcastsd_pd(_mm_permute_pd((__m128d) xmm, 0x1)); \ +} #define BROADCAST_B_PAIR_TAIL(Bx, By) {\ - __m128i xmm = (__m128i) _mm_load_sd(ptr_b##Bx + n_blksize * By); \ + __m128i xmm = (__m128i) _mm_load_sd((double *)(ptr_b##Bx + n_blksize * By)); \ xmm = _mm_cvtepu16_epi32(xmm); \ - B_lo = _mm512_broadcastd_epi32(xmm); \ - B_hi = _mm512_broadcastd_epi32((__m128i) _mm_permute_pd((__m128d) xmm, 0x1)); \ + B_lo = _mm512_broadcast_i32x2(xmm); \ + B_hi = _mm512_broadcast_i32x2((__m128i) _mm_permute_pd((__m128d) xmm, 0x1)); \ +} + +#define MASK_BROADCAST_B_PAIR_TAIL(Bx, By) {\ + __m128i xmm = _mm_maskz_loadu_epi16(nmask, ptr_b##Bx + n_blksize * By); \ + xmm = _mm_cvtepu16_epi32(xmm); \ + B_lo = _mm512_broadcast_i32x2(xmm); \ + B_hi = _mm512_broadcast_i32x2((__m128i) _mm_permute_pd((__m128d) xmm, 0x1)); \ } #define DECLARE_RESULT_4X(A, Bx, By) \ @@ -76,25 +102,103 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. FMA(A_lo_##A, B_hi, result_10_##A##Bx##By); \ FMA(A_hi_##A, B_hi, result_11_##A##Bx##By); -#define STORE_4X(A, Bx, By) +#define _STORE_C_2nx16(addr, val0, val1) \ + asm("vfmadd213ps (%1), %2, %0": "+v"(val0) : "r"(addr), "v"(alpha_512)); \ + asm("vfmadd213ps (%1, %3, 4), %2, %0": "+v"(val1) : "r"(addr), "v"(alpha_512), "r"(ldc)); \ + asm("vmovups %0, (%1)": : "v"(val0), "r"(addr)); \ + asm("vmovups %0, (%1, %2, 4)": : "v"(val1), "r"(addr), "r"(ldc)) +#define _MASK_STORE_C_2nx16(addr, val0, val1) \ + asm("vfmadd213ps (%1), %2, %0 %{%3%} ": "+v"(val0) : "r"(addr), "v"(alpha_512), "k"(mmask)); \ + asm("vfmadd213ps (%1, %3, 4), %2, %0 %{%4%}": "+v"(val1) : "r"(addr), "v"(alpha_512), "r"(ldc), "k"(mmask)); \ + asm("vmovups %0, (%1) %{%2%}": : "v"(val0), "r"(addr), "k"(mmask)); \ + asm("vmovups %0, (%1, %2, 4) %{%3%}": : "v"(val1), "r"(addr), "r"(ldc), "k"(mmask)) + +#define _REORDER_C_2X(result_0, result_1) { \ + __m512 tmp0, tmp1; \ + tmp0 = _mm512_unpacklo_ps(result_0, result_1); \ + tmp1 = _mm512_unpackhi_ps(result_0, result_1); \ + result_0 = (__m512) _mm512_unpacklo_pd((__m512d) tmp0, (__m512d) tmp1); \ + result_1 = (__m512) _mm512_unpackhi_pd((__m512d) tmp0, (__m512d) tmp1); \ +} + +#define _STORE_2X(ptr_c, result_0, result_1) {\ + _REORDER_C_2X(result_0, result_1) \ + _STORE_C_2nx16(ptr_c, result_0, result_1); \ + ptr_c += ldc * 2; \ +} + +#define _MASK_STORE_2X(ptr_c, result_0, result_1) {\ + _REORDER_C_2X(result_0, result_1) \ + _MASK_STORE_C_2nx16(ptr_c, result_0, result_1); \ + ptr_c += ldc * 2; \ +} + +#define STORE_4X(A, Bx, By) { \ + _STORE_2X(ptr_c##A, result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _STORE_2X(ptr_c##A, result_10_##A##Bx##By, result_11_##A##Bx##By); \ +} + +#define MASK_STORE_4X(A, Bx, By) { \ + _MASK_STORE_2X(ptr_c##A, result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _MASK_STORE_2X(ptr_c##A, result_10_##A##Bx##By, result_11_##A##Bx##By); \ +} + +#define _STORE_C_16(addr, val0) \ + asm("vfmadd213ps (%1), %2, %0": "+v"(val0) : "r"(addr), "v"(alpha_512)); \ + asm("vmovups %0, (%1)": : "v"(val0), "r"(addr)); + +#define _MASK_STORE_C_16(addr, val0) \ + asm("vfmadd213ps (%1), %2, %0 %{%3%} ": "+v"(val0) : "r"(addr), "v"(alpha_512), "k"(mmask)); \ + asm("vmovups %0, (%1) %{%2%}": : "v"(val0), "r"(addr), "k"(mmask)); + +#define N_STORE_4X(A, Bx, By) { \ + _REORDER_C_2X(result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _REORDER_C_2X(result_10_##A##Bx##By, result_11_##A##Bx##By); \ + switch(n_count) { \ + case 3: _STORE_C_16(ptr_c + ldc * 2, result_10_##A##Bx##By); \ + case 2: _STORE_C_16(ptr_c + ldc * 1, result_01_##A##Bx##By); \ + case 1: _STORE_C_16(ptr_c + ldc * 0, result_00_##A##Bx##By); \ + } \ + ptr_c##A += ldc * n_count; \ +} + +#define N_MASK_STORE_4X(A, Bx, By) { \ + _REORDER_C_2X(result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _REORDER_C_2X(result_10_##A##Bx##By, result_11_##A##Bx##By); \ + switch(n_count) { \ + case 3: _MASK_STORE_C_16(ptr_c + ldc * 2, result_10_##A##Bx##By); \ + case 2: _MASK_STORE_C_16(ptr_c + ldc * 1, result_01_##A##Bx##By); \ + case 1: _MASK_STORE_C_16(ptr_c + ldc * 0, result_00_##A##Bx##By); \ + } \ + ptr_c##A += ldc * n_count; \ +} int CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * B, FLOAT * C, BLASLONG ldc) { - IFLOAT *ptr_a = A, *ptr_b = B, *ptr_c = C; + IFLOAT *ptr_a = A, *ptr_b = B; IFLOAT *ptr_b0, *ptr_b1; IFLOAT *ptr_a0, *ptr_a1; + FLOAT *ptr_c = C; + FLOAT *ptr_c0, *ptr_c1; BLASLONG n_count = n; BLASLONG m_count, k_count; BLASLONG n_blksize = 4 * k; + BLASLONG cn_offset = 0; + __m512 alpha_512 = _mm512_broadcastss_ps(_mm_load_ss(&alpha)); for (; n_count > 23; n_count -= 24) { + IFLOAT *ptr_b00 = ptr_b; + IFLOAT *ptr_b10 = ptr_b + n_blksize * 3; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; m_count = m; - ptr_b0 = ptr_b; - ptr_b1 = ptr_b0 + n_blksize * 3; for (; m_count > 15; m_count -= 16) { - DECLARE_A_PAIR(0); DECLARE_B_PAIR(); + ptr_b0 = ptr_b00; + ptr_b1 = ptr_b10; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); DECLARE_RESULT_4X(0, 1, 0); DECLARE_RESULT_4X(0, 1, 1); DECLARE_RESULT_4X(0, 1, 2); for (k_count = k; k_count > 1; k_count -=2) { @@ -105,8 +209,8 @@ int CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * BROADCAST_B_PAIR(1, 0); MATMUL_4X(0, 1, 0); BROADCAST_B_PAIR(1, 1); MATMUL_4X(0, 1, 1); BROADCAST_B_PAIR(1, 2); MATMUL_4X(0, 1, 2); - ptr_b0 += 24 * 2; - ptr_b1 += 24 * 2; + ptr_b0 += 4 * 2; + ptr_b1 += 4 * 2; ptr_a0 += 16 * 2; } if (k_count > 0) { @@ -117,10 +221,249 @@ int CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * BROADCAST_B_PAIR_TAIL(1, 0); MATMUL_4X(0, 1, 0); BROADCAST_B_PAIR_TAIL(1, 1); MATMUL_4X(0, 1, 1); BROADCAST_B_PAIR_TAIL(1, 2); MATMUL_4X(0, 1, 2); - ptr_b0 += 24; - ptr_b1 += 24; + ptr_b0 += 4; + ptr_b1 += 4; ptr_a0 += 16; } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); STORE_4X(0, 0, 1); STORE_4X(0, 0, 2); + STORE_4X(0, 1, 0); STORE_4X(0, 1, 1); STORE_4X(0, 1, 2); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + ptr_b1 = ptr_b10; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + DECLARE_RESULT_4X(0, 1, 0); DECLARE_RESULT_4X(0, 1, 1); DECLARE_RESULT_4X(0, 1, 2); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + BROADCAST_B_PAIR(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR(1, 2); MATMUL_4X(0, 1, 2); + ptr_b0 += 4 * 2; + ptr_b1 += 4 * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + BROADCAST_B_PAIR_TAIL(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR_TAIL(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR_TAIL(1, 2); MATMUL_4X(0, 1, 2); + ptr_b0 += 4; + ptr_b1 += 4; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); MASK_STORE_4X(0, 0, 1); MASK_STORE_4X(0, 0, 2); + MASK_STORE_4X(0, 1, 0); MASK_STORE_4X(0, 1, 1); MASK_STORE_4X(0, 1, 2); + ptr_c += m_count; + } + ptr_b += 24 * k; + cn_offset += 24; + } + for (; n_count > 11; n_count -= 12) { + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_a1 = ptr_a + 16 * k; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 31; m_count -= 32) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); DECLARE_A_PAIR(1); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + DECLARE_RESULT_4X(1, 0, 0); DECLARE_RESULT_4X(1, 0, 1); DECLARE_RESULT_4X(1, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); LOAD_A_PAIR(1); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); MATMUL_4X(1, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); MATMUL_4X(1, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); MATMUL_4X(1, 0, 2); + ptr_b0 += 4 * 2; + ptr_a0 += 16 * 2; + ptr_a1 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); LOAD_A_PAIR_TAIL(1); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); MATMUL_4X(1, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); MATMUL_4X(1, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); MATMUL_4X(1, 0, 2); + ptr_b0 += 4; + ptr_a0 += 16; + ptr_a1 += 16; + } + ptr_c0 = ptr_c; + ptr_c1 = ptr_c + 16; + STORE_4X(0, 0, 0); STORE_4X(1, 0, 0); + STORE_4X(0, 0, 1); STORE_4X(1, 0, 1); + STORE_4X(0, 0, 2); STORE_4X(1, 0, 2); + ptr_c += 16 * 2; + } + if (m > 31) { + ptr_a0 = ptr_a1; + } + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + ptr_a0 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + ptr_a0 += 16; + } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); STORE_4X(0, 0, 1); STORE_4X(0, 0, 2); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); MASK_STORE_4X(0, 0, 1); MASK_STORE_4X(0, 0, 2); + ptr_c += m_count; + } + ptr_b += 12 * k; + cn_offset += 12; + } + for (; n_count > 3; n_count -= 4) { + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4 * 2; + ptr_a0 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4; + ptr_a0 += 16; + } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4 * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); + ptr_c += m_count; + } + ptr_b += 4 * k; + cn_offset += 4; + } + if (n_count > 0) { + __mmask8 nmask = (1UL << n_count) - 1; + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + MASK_BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count * 2; + ptr_a0 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + MASK_BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count; + ptr_a0 += 16; + } + ptr_c0 = ptr_c; + N_STORE_4X(0, 0, 0); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + MASK_BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + MASK_BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + N_MASK_STORE_4X(0, 0, 0); + ptr_c += m_count; } } + return 0; } diff --git a/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c b/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c index 523e3b48f..eefbd7355 100644 --- a/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c +++ b/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c @@ -79,8 +79,8 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ aoffset = a; boffset = b; - BLASLONG m32 = n & ~31; - BLASLONG m8 = n & ~7; + BLASLONG m32 = m & ~31; + BLASLONG m8 = m & ~7; BLASLONG n4 = n & ~3; int permute_table[] = { @@ -115,15 +115,15 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ boffset += 32 * 4; } for (; i < m8; i += 8) { - __m128i r0 = _mm_loadu_si128(aoffset0 + i); - __m128i r1 = _mm_loadu_si128(aoffset1 + i); - __m128i r2 = _mm_loadu_si128(aoffset2 + i); - __m128i r3 = _mm_loadu_si128(aoffset3 + i); + __m128i r0 = _mm_loadu_si128((void *)(aoffset0 + i)); + __m128i r1 = _mm_loadu_si128((void *)(aoffset1 + i)); + __m128i r2 = _mm_loadu_si128((void *)(aoffset2 + i)); + __m128i r3 = _mm_loadu_si128((void *)(aoffset3 + i)); REORDER_4x8(r0, r1, r2, r3); - _mm_storeu_si128(boffset + 8*0, r0); - _mm_storeu_si128(boffset + 8*1, r1); - _mm_storeu_si128(boffset + 8*2, r2); - _mm_storeu_si128(boffset + 8*3, r3); + _mm_storeu_si128((void *)(boffset + 8*0), r0); + _mm_storeu_si128((void *)(boffset + 8*1), r1); + _mm_storeu_si128((void *)(boffset + 8*2), r2); + _mm_storeu_si128((void *)(boffset + 8*3), r3); boffset += 8 * 4; } if (i < m) { @@ -138,9 +138,9 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ // store should skip the tail odd line int num_store = remain_m/2; switch(num_store) { - case 3: _mm_storeu_si128(boffset + 8*2, r0); - case 2: _mm_storeu_si128(boffset + 8*1, r0); - case 1: _mm_storeu_si128(boffset + 8*0, r0); + case 3: _mm_storeu_si128((void *)(boffset + 8*2), r2); + case 2: _mm_storeu_si128((void *)(boffset + 8*1), r1); + case 1: _mm_storeu_si128((void *)(boffset + 8*0), r0); } boffset += 8 * num_store; @@ -152,7 +152,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ * need to extract lo words of data and store */ tail = _mm_cvtepi32_epi16(tail); - _mm_store_sd(boffset, (__m128d) tail); // only lower 4 bfloat valid + _mm_store_sd((double *)boffset, (__m128d) tail); // only lower 4 bfloat valid boffset += 4; } } @@ -167,16 +167,16 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ __m128i r0, r1, r2, r3; for (i = 0; i < m8; i += 8) { switch (remain_n) { - case 3: r2 = _mm_loadu_si128(aoffset2 + i); - case 2: r1 = _mm_loadu_si128(aoffset1 + i); - case 1: r0 = _mm_loadu_si128(aoffset0 + i); + case 3: r2 = _mm_loadu_si128((void *)(aoffset2 + i)); + case 2: r1 = _mm_loadu_si128((void *)(aoffset1 + i)); + case 1: r0 = _mm_loadu_si128((void *)(aoffset0 + i)); } REORDER_4x8(r0, r1, r2, r3); - _mm_mask_storeu_epi16(boffset + remain_n * 0, nmask, r0); - _mm_mask_storeu_epi16(boffset + remain_n * 1, nmask, r1); - _mm_mask_storeu_epi16(boffset + remain_n * 2, nmask, r2); - _mm_mask_storeu_epi16(boffset + remain_n * 3, nmask, r3); - boffset += 4 * remain_n; + _mm_mask_storeu_epi32(boffset + remain_n * 0, nmask, r0); + _mm_mask_storeu_epi32(boffset + remain_n * 2, nmask, r1); + _mm_mask_storeu_epi32(boffset + remain_n * 4, nmask, r2); + _mm_mask_storeu_epi32(boffset + remain_n * 6, nmask, r3); + boffset += 8 * remain_n; } if (i < m) { int remain_m = m - i; @@ -190,9 +190,9 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ int num_store = remain_m/2; switch (num_store) { - case 3: _mm_mask_storeu_epi16(boffset + remain_n * 2, nmask, r2); - case 2: _mm_mask_storeu_epi16(boffset + remain_n * 1, nmask, r1); - case 1: _mm_mask_storeu_epi16(boffset + remain_n * 0, nmask, r0); + case 3: _mm_mask_storeu_epi32(boffset + remain_n * 4, nmask, r2); + case 2: _mm_mask_storeu_epi32(boffset + remain_n * 2, nmask, r1); + case 1: _mm_mask_storeu_epi32(boffset + remain_n * 0, nmask, r0); } boffset += 2 * num_store * remain_n; @@ -204,4 +204,5 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ } } } + return 0; } diff --git a/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c b/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c index 16bf48f0b..ce4458d2c 100644 --- a/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c +++ b/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c @@ -29,134 +29,135 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include #include "common.h" + int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ - BLASLONG i, j; + BLASLONG i, j; - IFLOAT *boffset0, *boffset1; + IFLOAT *boffset0, *boffset1; - boffset0 = b; + boffset0 = b; - BLASLONG n32 = n & ~31; - BLASLONG m4 = m & ~3; - BLASLONG m2 = m & ~1; + BLASLONG n32 = n & ~31; + BLASLONG m4 = m & ~3; + BLASLONG m2 = m & ~1; - uint32_t permute_table = { - 0x00, 0x10, 0x01, 0x11, 0x02, 0x12, 0x03, 0x13, 0x04, 0x14, 0x05, 0x15, 0x06, 0x16, 0x07, 0x17, - 0x08, 0x18, 0x09, 0x19, 0x0a, 0x1a, 0x0b, 0x1b, 0x0c, 0x1c, 0x0d, 0x1d, 0x0e, 0x1e, 0x0f, 0x1f, - }; + uint32_t permute_table[] = { + 0x00, 0x01, 0x02, 0x03, 0x10, 0x11, 0x12, 0x13, 0x04, 0x05, 0x06, 0x07, 0x14, 0x15, 0x16, 0x17, + 0x08, 0x09, 0x0a, 0x0b, 0x18, 0x19, 0x1a, 0x1b, 0x0c, 0x0d, 0x0e, 0x0f, 0x1c, 0x1d, 0x1e, 0x1f, + }; - __m512i idx_lo = _mm512_loadu_si512(permute_table); - __m512i idx_hi = _mm512_loadu_si512(permute_table + 16); + __m512i idx_lo = _mm512_loadu_si512(permute_table); + __m512i idx_hi = _mm512_loadu_si512(permute_table + 16); - for (j = 0; j < n32; j += 32) { - /* process 2x16 n at the same time */ - boffset1 = boffset0 + m * 16; - for (i = 0; i < m4; i += 4) { - /* bf16 fma need special memory layout: - * for memory layout like below: - * a00, a01, a02, a03, a04, a05 .... - * a10, a11, a12, a13, a14, a15 .... - * need to copy as: - * a00, a10, a01, a11, a02, a12, a03, a13, ... - */ - __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); - __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); - __m512i a2 = _mm512_loadu_si512(&a[(i + 2)*lda + j]); - __m512i a3 = _mm512_loadu_si512(&a[(i + 3)*lda + j]); + for (j = 0; j < n32; j += 32) { + /* process 2x16 n at the same time */ + boffset1 = boffset0 + m * 16; + for (i = 0; i < m4; i += 4) { + /* bf16 fma need special memory layout: + * for memory layout like below: + * a00, a01, a02, a03, a04, a05 .... + * a10, a11, a12, a13, a14, a15 .... + * need to copy as: + * a00, a10, a01, a11, a02, a12, a03, a13, ... + */ + __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); + __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); + __m512i a2 = _mm512_loadu_si512(&a[(i + 2)*lda + j]); + __m512i a3 = _mm512_loadu_si512(&a[(i + 3)*lda + j]); - __m512i a00 = _mm512_unpacklo_epi16(a0, a1); - __m512i a01 = _mm512_unpackhi_epi16(a0, a1); - __m512i a10 = _mm512_unpacklo_epi16(a2, a3); - __m512i a11 = _mm512_unpackhi_epi16(a2, a3); + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + __m512i a10 = _mm512_unpacklo_epi16(a2, a3); + __m512i a11 = _mm512_unpackhi_epi16(a2, a3); - a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); - a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); - a2 = _mm512_permutex2var_epi32(a10, idx_lo, a11); - a3 = _mm512_permutex2var_epi32(a10, idx_hi, a11); + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + a2 = _mm512_permutex2var_epi32(a10, idx_lo, a11); + a3 = _mm512_permutex2var_epi32(a10, idx_hi, a11); - _mm512_storeu_si512(boffset0, a0); - _mm512_storeu_si512(boffset1, a1); - _mm512_storeu_si512(boffset0 + 32, a2); - _mm512_storeu_si512(boffset1 + 32, a3); - boffset0 += 64; - boffset1 += 64; - } - for (; i < m2; i += 2) { - __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); - __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); + _mm512_storeu_si512(boffset0, a0); + _mm512_storeu_si512(boffset1, a1); + _mm512_storeu_si512(boffset0 + 32, a2); + _mm512_storeu_si512(boffset1 + 32, a3); + boffset0 += 64; + boffset1 += 64; + } + for (; i < m2; i += 2) { + __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); + __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); - __m512i a00 = _mm512_unpacklo_epi16(a0, a1); - __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); - a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); - a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); - _mm512_storeu_si512(boffset0, a0); - _mm512_storeu_si512(boffset1, a1); - boffset0 += 32; - boffset1 += 32; - } - for (; i < m; i++) { - /* just copy the only remains row */ - __m256i a0 = _mm256_loadu_si256(&a[(i + 0)*lda + j]); - __m256i a1 = _mm256_loadu_si256(&a[(i + 0)*lda + j + 16]); - _mm256_storeu_si256(boffset0, a0); - _mm256_storeu_si256(boffset1, a1); - boffset0 += 16; - boffset1 += 16; - } - boffset0 = boffset1; - } - if (j < n) { - uint32_t remains = n - j; - __mmask32 r_mask = (1UL << remains) - 1; - if (remains > 16) { - boffset1 = boffset0 + m * 16; - uint32_t tail1 = remains - 16; - __mmask16 w_mask1 = (1UL << tail1) - 1; - for (i = 0; i < m2; i += 2) { - __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); - __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + _mm512_storeu_si512(boffset0, a0); + _mm512_storeu_si512(boffset1, a1); + boffset0 += 32; + boffset1 += 32; + } + for (; i < m; i++) { + /* just copy the only remains row */ + __m256i a0 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j]); + __m256i a1 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j + 16]); + _mm256_storeu_si256((void *)boffset0, a0); + _mm256_storeu_si256((void *)boffset1, a1); + boffset0 += 16; + boffset1 += 16; + } + boffset0 = boffset1; + } + if (j < n) { + uint32_t remains = n - j; + __mmask32 r_mask = (1UL << remains) - 1; + if (remains > 16) { + boffset1 = boffset0 + m * 16; + uint32_t tail1 = remains - 16; + __mmask16 w_mask1 = (1UL << tail1) - 1; + for (i = 0; i < m2; i += 2) { + __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); - __m512i a00 = _mm512_unpacklo_epi16(a0, a1); - __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); - a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); - a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); - _mm512_storeu_si512(boffset0, a0); - _mm512_mask_storeu_epi32(boffset1, w_mask1, a1); + _mm512_storeu_si512(boffset0, a0); + _mm512_mask_storeu_epi32(boffset1, w_mask1, a1); - boffset0 += 32; - boffset1 += 2 * tail1; - } - for (; i < m; i++) { - __m256i a0 = _mm256_loadu_si256(&a[(i + 0)*lda + j]); - __m256i a1 = _mm256_maskz_loadu_epi16(w_mask1, &a[(i + 0)*lda + j + 16]); - _mm256_storeu_si256(boffset0, a0); - _mm256_mask_storeu_epi16(boffset1, w_mask1, a1); - boffset0 += 16; - boffset1 += tail1; - } - } else { - __mmask16 w_mask = (1UL << remains ) - 1; - for (i = 0; i < m2; i += 2) { - __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); - __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + boffset0 += 32; + boffset1 += 2 * tail1; + } + for (; i < m; i++) { + __m256i a0 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j]); + __m256i a1 = _mm256_maskz_loadu_epi16(w_mask1, (void *)&a[(i + 0)*lda + j + 16]); + _mm256_storeu_si256((void *)boffset0, a0); + _mm256_mask_storeu_epi16((void *)boffset1, w_mask1, a1); + boffset0 += 16; + boffset1 += tail1; + } + } else { + __mmask16 w_mask = (1UL << remains ) - 1; + for (i = 0; i < m2; i += 2) { + __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); - __m512i a00 = _mm512_unpacklo_epi16(a0, a1); - __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); - a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); - _mm512_mask_storeu_epi32(boffset0, w_mask, a0); - boffset0 += 2 * remains; - } - for (; i < m; i++) { - __m256i a0 = _mm256_maskz_loadu_epi16(w_mask, &a[(i + 0)*lda + j]); - _mm256_mask_storeu_epi16(boffset0, w_mask, a0); - boffset0 += remains; - } - } - } + _mm512_mask_storeu_epi32(boffset0, w_mask, a0); + boffset0 += 2 * remains; + } + for (; i < m; i++) { + __m256i a0 = _mm256_maskz_loadu_epi16(w_mask, &a[(i + 0)*lda + j]); + _mm256_mask_storeu_epi16(boffset0, w_mask, a0); + boffset0 += remains; + } + } + } }