Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Welcome To Ask or Share your Answers For Others

Categories

0 votes
517 views
in Technique[技术] by (71.8m points)

optimization - Efficient 4x4 matrix multiplication (C vs assembly)

I'm looking for a faster and trickier way to multiply two 4x4 matrices in C. My current research is focused on x86-64 assembly with SIMD extensions. So far, I've created a function witch is about 6x faster than a naive C implementation, which has exceeded my expectations for the performance improvement. Unfortunately, this stays true only when no optimization flags are used for compilation (GCC 4.7). With -O2, C becomes faster and my effort becomes meaningless.

I know that modern compilers make use of complex optimization techniques to achieve an almost perfect code, usually faster than an ingenious piece of hand-crafed assembly. But in a minority of performance-critical cases, a human may try to fight for clock cycles with the compiler. Especially, when some mathematics backed with a modern ISA can be explored (as it is in my case).

My function looks as follows (AT&T syntax, GNU Assembler):

    .text
    .globl matrixMultiplyASM
    .type matrixMultiplyASM, @function
matrixMultiplyASM:
    movaps   (%rdi), %xmm0    # fetch the first matrix (use four registers)
    movaps 16(%rdi), %xmm1
    movaps 32(%rdi), %xmm2
    movaps 48(%rdi), %xmm3
    xorq %rcx, %rcx           # reset (forward) loop iterator
.ROW:
    movss (%rsi), %xmm4       # Compute four values (one row) in parallel:
    shufps $0x0, %xmm4, %xmm4 # 4x 4FP mul's, 3x 4FP add's 6x mov's per row,
    mulps %xmm0, %xmm4        # expressed in four sequences of 5 instructions,
    movaps %xmm4, %xmm5       # executed 4 times for 1 matrix multiplication.
    addq $0x4, %rsi

    movss (%rsi), %xmm4       # movss + shufps comprise _mm_set1_ps intrinsic
    shufps $0x0, %xmm4, %xmm4 #
    mulps %xmm1, %xmm4
    addps %xmm4, %xmm5
    addq $0x4, %rsi           # manual pointer arithmetic simplifies addressing

    movss (%rsi), %xmm4
    shufps $0x0, %xmm4, %xmm4
    mulps %xmm2, %xmm4        # actual computation happens here
    addps %xmm4, %xmm5        #
    addq $0x4, %rsi

    movss (%rsi), %xmm4       # one mulps operand fetched per sequence
    shufps $0x0, %xmm4, %xmm4 #  |
    mulps %xmm3, %xmm4        # the other is already waiting in %xmm[0-3]
    addps %xmm4, %xmm5
    addq $0x4, %rsi           # 5 preceding comments stride among the 4 blocks

    movaps %xmm5, (%rdx,%rcx) # store the resulting row, actually, a column
    addq $0x10, %rcx          # (matrices are stored in column-major order)
    cmpq $0x40, %rcx
    jne .ROW
    ret
.size matrixMultiplyASM, .-matrixMultiplyASM

It calculates a whole column of the resultant matrix per iteration, by processing four floats packed in 128-bit SSE registers. The full vectorisation is possible with a bit of math (operation reordering and aggregation) and mullps/addps instructions for parallel multiplication/addition of 4xfloat packages. The code reuses registers meant for passing parameters (%rdi, %rsi, %rdx : GNU/Linux ABI), benefits from (inner) loop unrolling and holds one matrix entirely in XMM registers to reduce memory reads. A you can see, I have researched the topic and took my time to implement it the best I can.

The naive C calculation conquering my code looks like this:

void matrixMultiplyNormal(mat4_t *mat_a, mat4_t *mat_b, mat4_t *mat_r) {
    for (unsigned int i = 0; i < 16; i += 4)
        for (unsigned int j = 0; j < 4; ++j)
            mat_r->m[i + j] = (mat_b->m[i + 0] * mat_a->m[j +  0])
                            + (mat_b->m[i + 1] * mat_a->m[j +  4])
                            + (mat_b->m[i + 2] * mat_a->m[j +  8])
                            + (mat_b->m[i + 3] * mat_a->m[j + 12]);
}

I have investigated the optimised assembly output of the above's C code which, while storing floats in XMM registers, does not involve any parallel operations – just scalar calculations, pointer arithmetic and conditional jumps. The compiler's code seems to be less deliberate, but it is still slightly more effective than my vectorised version expected to be about 4x faster. I'm sure that the general idea is correct – programmers do similar things with rewarding results. But what is wrong here? Are there any register allocation or instruction scheduling issues I am not aware of? Do you know any x86-64 assembly tools or tricks to support my battle against the machine?

See Question&Answers more detail:os

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome To Ask or Share your Answers For Others

1 Reply

0 votes
by (71.8m points)

4x4 matrix multiplication is 64 multiplications and 48 additions. Using SSE this can be reduced to 16 multiplications and 12 additions (and 16 broadcasts). The following code will do this for you. It only requires SSE (#include <xmmintrin.h>). The arrays A, B, and C need to be 16 byte aligned. Using horizontal instructions such as hadd (SSE3) and dpps (SSE4.1) will be less efficient (especially dpps). I don't know if loop unrolling will help.

void M4x4_SSE(float *A, float *B, float *C) {
    __m128 row1 = _mm_load_ps(&B[0]);
    __m128 row2 = _mm_load_ps(&B[4]);
    __m128 row3 = _mm_load_ps(&B[8]);
    __m128 row4 = _mm_load_ps(&B[12]);
    for(int i=0; i<4; i++) {
        __m128 brod1 = _mm_set1_ps(A[4*i + 0]);
        __m128 brod2 = _mm_set1_ps(A[4*i + 1]);
        __m128 brod3 = _mm_set1_ps(A[4*i + 2]);
        __m128 brod4 = _mm_set1_ps(A[4*i + 3]);
        __m128 row = _mm_add_ps(
                    _mm_add_ps(
                        _mm_mul_ps(brod1, row1),
                        _mm_mul_ps(brod2, row2)),
                    _mm_add_ps(
                        _mm_mul_ps(brod3, row3),
                        _mm_mul_ps(brod4, row4)));
        _mm_store_ps(&C[4*i], row);
    }
}

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
OGeek|极客中国-欢迎来到极客的世界,一个免费开放的程序员编程交流平台!开放,进步,分享!让技术改变生活,让极客改变未来! Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Click Here to Ask a Question

1.4m articles

1.4m replys

5 comments

57.0k users

...